• Chip codename: "Tahiti"
• 4.3 billion transistors (more than 60% more than Cayman and exactly twice as much as Cypress)
• 384-bit memory bus: six 64-bit wide controllers with GDDR5 memory support
• Core clock: up to 925 MHz (for Radeon HD 7970)
• 32 GCN Compute Units with 128 SIMD cores for a total of 2048 floating point ALUs (Integer and Float formats, IEEE 754 FP32 and FP64 precision support)
• 128 texture units, with support for trilinear and anisotropic filtering for all texture formats
• 32 ROPs with support for anti-aliasing modes with the possibility of programmable sampling of more than 16 samples per pixel, including with FP16 or FP32 framebuffer format. Peak performance up to 32 samples per clock, and in colorless mode (Z only) - 128 samples per clock
• Integrated support for six monitors including HDMI 1.4a and DisplayPort 1.2

Radeon HD 7970 Graphics Specifications

• Core clock: 925 MHz
• Number of universal processors: 2048
• Number of texture units: 128, blending units: 32
• Effective memory frequency: 5500 MHz (4×1375 MHz)
• Memory type: GDDR5
• Memory capacity: 3 gigabytes
• Memory bandwidth: 264 gigabytes per second
• theoretical maximum speed fills: 29.6 gigapixels per second
• Theoretical texture sampling rate: 118.4 gigatexels per second
• Two CrossFire connectors
• PCI Express 3.0 bus
• Power consumption: 3 to 250 W
• One 8-pin and one 6-pin power connector
• Dual slot design
• US MSRP: $549

Radeon HD 7950 Graphics Specifications

• Core frequency: 800 MHz
• Number of universal processors: 1792
• Number of texture units: 112, blending units: 32
• Effective memory frequency: 5000 MHz (4×1250 MHz)
• Memory type: GDDR5
• Memory capacity: 3 gigabytes
• Memory bandwidth: 240 gigabytes per second
• Theoretical maximum fill rate: 25.6 gigapixels per second.
• Theoretical texture sampling rate: 89.6 gigatexels per second
• Two CrossFire connectors
• PCI Express 3.0 bus
• Connectors: DVI Dual Link, HDMI 1.4, two Mini-DisplayPort 1.2
• Power consumption: 3 to 200 W
• Dual slot design
• US MSRP: $449

The high complexity of the new chip draws attention - 4.3 billion transistors, which is more than half the number of transistors in the previous top-end graphics processor. The ability to make such a complex crystal was made possible by the use of a modern 28-nanometer process technology, and the new chip turned out to be even slightly smaller in area than the Cayman size. And its practical characteristics that affect performance have been noticeably improved: the number of ALUs, TMUs, memory bus. Only the number of ROPs did not increase, and the frequency of GDDR5 video memory remained at the same level.

The principle of naming video cards of the company remained the same. The Radeon HD 7970 is the most productive single-chip solution from the company, after some time the younger model HD 7950 was released, which was announced a little later. Initially, the HD 7970 had no competitors in the market and did not replace any specific video card from the AMD line, but rather moved it down. As for the comparison with the competitor, NVIDIA released its 28nm solution much later.

The same GDDR5 memory is installed on the new AMD video card, but its volume, instead of 2 gigabytes in the previous generation, has grown to 3 gigabytes. This happened due to the expansion of the memory bus from 256-bit to 384-bit. And now you can put either 1.5 GB or 3 GB on the new board. Naturally, from a marketing point of view, installing a smaller volume would be a clear disadvantage, so the decision was made to put 3 GB, although today this is a bit overkill. Only in ultra-high resolutions and with MSAA 16x 1.5-2 GB will not be enough. However, AMD also has Eyefinity, and for games on three or more monitors, the screen buffer will just take up a very large amount.

So, let's take a look at the Radeon HD 7970. The new video card of the upper price range has a two-slot cooling system, which is covered with a plastic casing, which is familiar to all modern AMD motherboards, along the entire length of the card. Only the design of this casing has changed a little, although rear end still goes beyond the printed circuit board. But the design of the bar with pins was changed - to improve the cooling of the video card, one of the two slots (half of the bar) was occupied exclusively by a ventilation hole for heat dissipation.

But users shouldn't suffer from the reduction in the number of DVI connectors soldered right on the board. For their convenience, a special HDMI-DVI adapter will be included in the package, which will allow you to connect two monitors with DVI connectors. By the way, the power consumption of the new card is not lower than that of the Radeon HD 6970, so it had to install a set of one 8-pin and one 6-pin power connectors.

But in the new Radeon HD 7970, the cooling system has changed for the better. A new generation of evaporator chamber and a new larger cooler are used, with a modified blade shape and increased performance (greater airflow is provided). The result is an increase in cooler efficiency while reducing noise.

The Dual BIOS firmware switch, which we wrote about in the Radeon HD 6900 description, has not disappeared from the board. Both users and AMD liked this convenient solution so much that AMD decided to continue to complete top-end solutions with it.

We can only say hello this decision, which really helps in various cases related to both unexpected problems during flashing (turning off the power in the process, for example), and allows you to fearlessly conduct various experiments with BIOS images. Unsurprisingly, AMD hints again and again at the excellent overclocking capabilities of the new graphics card:

As you can see, overclocking to a frequency of 1 GHz and higher is practically promised, if you do not take into account the small inscription (which was not included in the screenshot) that the guarantee ceases to apply even if the video card fails as a result of an experiment with raising the frequency from the settings video driver.

Architectural features Radeon HD 7970

To assess the relevance of architectural modifications in the Southern Islands, first consider the development of the GPU over the past few years in the AMD view. Prior to 2002, graphics chips were specific hardware capable of purely graphics computing. Video chips of that time had limited functionality, they could only apply and filter textures, process geometry, engage in primitive rasterization, and therefore were not at all suitable for universal computing tasks.

Over the next few years, basic programmability was added to the GPU, but also focused exclusively on graphics tasks. This was the time of support for DirectX 8 and 9, limited functionality shader programs with floating point capability. Video chips of that time had specialized ALUs for vertex and pixel processing, as well as dedicated caches for pixels, textures, and other data. Universality was still not even close.

And only in 2007, AMD got a unified shader architecture DirectX 10, as well as the ability to program the GPU using special tools: CAL, Brook, ATI Stream. GPUs of the time already had advanced caching and support for local and global shared data. Architecturally, the chips were based on VLIW5 and VLIW4 blocks, flexible enough for some basic graphic computing, but still focused on graphical algorithms.

And now it's time for a new architecture, even better suited for universal computing - Graphics Core Next (GCN). For AMD, this is a new architectural era, which is why the name was chosen. The new GPUs offer excellent graphics processing capabilities and performance, but the architectural changes made are intended primarily to improve the position in non-graphics computing - increasing performance and efficiency in complex general-purpose tasks. New design The GPU is designed for so-called heterogeneous computing - a mixture of graphical and universal computing in a multitasking environment. The GCN architecture has become more flexible and should be even better suited for energy efficient execution of various tasks.

The basic block in the new architecture is the GCN block. It is on these "bricks" that all new GPUs of the Southern Islands series are based. The architecture for the first time for AMD graphics chips uses a non-VLIW design, it uses vector and scalar blocks, and one of the most important changes is that each of the GCN computing blocks has its own scheduler and can execute instructions from various programs (kernel).

The new computing architecture is designed for high efficiency loading of computing units in a multitasking environment. The GCN computational unit is divided into four subsections, each of which works on its own instruction stream every clock cycle. Threads can also use the scalar block found in GCN for flow control or pointer operations. The combination of vector and scalar blocks offers a very simple programming model. For example, function pointers and stack pointers are much easier to program, and the task of the compiler is now greatly simplified, since the execution units are scalar.

Each GCN block has a dedicated local storage data for 64 KB for data exchange or local stack extension for registers. Also, the block includes both a first-level cache memory with the ability to read and write, and a full-fledged texture pipeline (sampling and filtering blocks). Therefore, the new computing unit is able to work independently, without a central scheduler, which in previous architectures was responsible for distributing work among blocks. Now each of the GCN blocks is capable of scheduling and distributing commands on its own, one computing unit can execute up to 32 different command streams, which can be from different virtual address spaces in memory and are completely protected and independent of each other.

AMD's previous GPU architectures used the VLIW4 and VLIW5 architecture models, and although they are good enough for graphics tasks, but are not efficient enough for universal computing, since it is very difficult to load all the execution units with work in such conditions. The new GCN architecture offers a similarly large number of execution units, but in a scalar execution that removes the limitations and dependencies of registers and instructions. The transition from the VLIW architecture to the scalar execution provides a noticeable simplification of code optimization tasks.

When executing instructions on the previous VLIW4 architecture, the compiler has to deal with register conflict resolution, complex distribution of instructions to execution units at the code compilation stage, etc. At the same time, to achieve high performance, non-trivial optimization is often required, which is suitable for most graphics tasks and much less flexible for other calculations. The new architecture offers a significant simplification of development and support, simplified creation, analysis and catching errors in low-level code, stable and predictable performance.

Memory caching subsystem

Bandwidth and memory and caches are never enough, and there is always a need and methods to increase them. AMD's new GPUs use a full two-level read/write cache. Each computing unit has 16 kilobytes of the first level cache, and the total volume of the second level cache is 768 kilobytes (in total, the chip gets 512 KB L1 and 768 KB L2), which is 50% more than in the previous chip, which does not have the ability to write at all to L2 cache.

In terms of performance, each GCN computational unit can receive or write 64 bytes of data from / to the L1 cache or global memory in one cycle, which serves to exchange data between command streams. The same amount of data is capable of transmitting and receiving each section of the L2 cache memory. As a result, the company's top-end GPU achieves 2 terabytes / s for L1 and 700 GB / s for L2, which is 50% more than the previous top AMD solution.

Tahiti GPU

Now that we've looked at the low-level architectural changes to the new Southern Islands series, it's time to move on to the details of the line's most powerful solution, the Radeon HD 7900, which includes two models. First of all, let's just note the huge complexity of the new GPU, because it includes more than 4.3 billion transistors, which is twice as much as was in the chip on which the Radeon HD 5870 is based! Naturally, such a powerful chip became possible only thanks to the use of a new 28 nm process technology. So what does he have inside?

The number of geometric blocks has not changed, compared to the Cayman, there are still two of them, but the efficiency of their work has been significantly increased - we will dwell on this in more detail a little later. On the graphics processor diagram, we see 32 GCN architecture compute units available on the Radeon HD 7970, and in the case of the junior solution, some of them will be disabled. If we consider the peak computing performance of the solution, then it is almost 3.8 teraflops (the number of floating point operations per second), which is an absolute record for a GPU today.

Each GCN block has 16 texture units, which gives a total figure of 128 TMUs per chip, or more than 118 gigatexels / sec - and this is another record at the time of release, and it will not be the last. But the number of ROP blocks has not changed, there are still 32 of them in 8 enlarged RBE blocks. Another interesting architectural change is that now ROP blocks are “attached” not to memory channels, as it was before, but to GCN blocks.

Although the theoretical framebuffer write speed has not changed much, and the maximum possible are the same 32 color values ​​and 128 depth values ​​per clock, the practical fill rate (fill rate) in real applications has increased significantly due to the increased memory bandwidth. AMD measured the Cayman at only 23 pixels per clock, while the new Tahiti approached the theoretical 32 pixels per clock.

This is understandable, because the new AMD video chip has a 384-bit memory bus - six 64-bit channels, just like the current top-end competitor's solution. It is this 1.5-fold increase in memory bandwidth that makes it possible to increase the actual speed of texture fetches and writing to the framebuffer. The bandwidth of 264 GB / s should help to squeeze out close to theoretical figures of 118 gigatexels / sec and 30 gigapixels / sec, and in the practical part we will check this.

In the case of the “stripped down” Radeon HD 7950 graphics processor, Tahiti includes 28 active computing units of the GCN architecture out of 32 physically available on the chip. In the case of the junior solution of the Radeon HD 7970 series, it was decided to disable four of them. Since each GCN has 16 texture units, the total TMU figure for the new model is 112 TMUs, which gives a throughput of almost 90 gigatexels/sec.

But the number of ROPs and memory controllers in the HD 7950 has not changed, they decided not to cut them down and keep the same 32 and 6 pieces, respectively. Therefore, the Tahiti Pro video chip has the same 384-bit memory bus, assembled from six 64-bit channels, as the top AMD solution. Apparently, it is computing functional devices that suffer the most from marriage during production, and they decided not to cut everything else.

Tessellation and Geometry Processing

From an architectural point of view, nothing special has changed in the geometric blocks of the Tahiti since the Cayman. It still uses two blocks for processing (setting vertices and tessellation) of geometric data and rasterization, and the scheme is very similar to the one we saw earlier, except that the tessellators are already called the 9th generation:

Despite the schematic similarities, the latest generation of these blocks is capable of significantly better tessellation and geometry processing performance, since the blocks have undergone significant modifications. Although peak performance has grown only to almost two billion vertices and primitives per second (925 MHz and two vertices per clock), real performance grew more. This was achieved by increasing the size of caches, improving buffering of geometry data, and reusing vertex data.

As a result, tessellation performance is improved at all triangle split ratios by up to four times over the previous generation Radeon HD 6970. But four times is not achieved in all cases, even on the diagram from AMD itself:

The chart shows a comparison of the tessellation performance of the Radeon HD 7970 compared to the HD 6970 at split ratios from 1 to 32. And as you can see, the performance difference is from 1.7 to 4 times. But this is naked synthetics. And in order to get closer to reality, we will give more data on the tessellation speed already in gaming applications:

As you can see, AMD's synthetic numbers are well supported by gaming ones - performance in real applications with "heavy" tessellation has grown significantly. This is a very good result, which we will definitely check in the practical part, using the example of synthetics and gaming applications.

Non-graphic calculations

From the point of view of heterogeneous and non-graphical computing tasks, two asynchronous computing engines (Asynchronous Compute Engines - ACE) have appeared. They are designed to schedule and distribute work between execution units for efficient multitasking and work in conjunction with a graphic command processor (Command Processor).

The Radeon HD 7900 has two independent compute engines and one graphics engine. In total, this gives three programmable blocks and three instruction streams, completely separated from each other. And in addition to asynchronously issuing commands for fast context switching, the new GPU also features two bi-directional direct memory access (DMA) controllers introduced in the Cayman. These two controllers are required to take full advantage of the new PCI Express 3.0 bus.

As we know, from the point of view of serious calculations, not only the speed of performing floating point operations with single precision, but also double precision (double precision floating point) is important. And AMD's new architecture does a pretty good job of that. On the this moment It is assumed that there are two versions of GCN computing units with different execution rates for FP64 instructions. For the older GPU, the execution rate is 1/4 of the FP32 speed, and for the younger chips, the execution rate is 1/16, which is quite enough to maintain compatibility, but does not complicate inexpensive solutions too much. As a result, the Radeon HD 7970 is capable of 947 billion double-precision operations per second (oh, they just barely reached the teraflop!) - there is another highest achievement of the new AMD chip.

Moreover, these are not the same gigaflops as in the case of previous architectures, but more “fat” ones. After all, the efficiency of the new GPU in complex computing tasks should seriously increase. First, the memory and caching subsystem has been improved. Secondly, each GCN compute unit has its own scheduler, which should improve branching code execution and overall efficiency. And thirdly, we note the scalar execution, which does not require complex optimizations from the compiler, as a result of which the computing units will be idle much less often. And as a result, in any tasks, it will be easier for the new chip to show high performance and ALU loading.

Among other innovations related to computing capabilities, we note the full support for ECC for DRAM and SRAM. On the software side, it is important that Tahiti is the first GPU with full support for new API versions: OpenCL 1.2, DirectCompute 11.1 and C++ AMP and their capabilities. For example, OpenCL 1.2 allows you to combine the capabilities of several computing devices into one, and AMD has already released support in the form of AMD APP SDK 2.6 and the Catalyst 11.12 driver.

Architectural Performance and Efficiency

After reviewing all the architectural innovations on the example of the top chip of the Southern Island series, it's time to talk about the effectiveness of all these changes. It is clear that the performance of the new chips is much higher than that of the previous ones, the opposite would be quite surprising. The question is how much faster. In various tasks, figures are obtained from 40-50% (minimum!) To a fivefold difference. Improvements in the architecture make it possible to exceed the theoretical 1.4-fold difference in dumb gigaflops. Let's look at this with examples:

The diagram compares the new top solution and the previous single-chip solution: Radeon HD 7970 and HD 6970, which is quite fair. Various performance tests are chosen: SmallptGPU and LuxMark are ray tracing on OpenCL, SHA256 is a secure hashing algorithm, and AES256 is a symmetric encryption algorithm. Well, Mandelbrot is a well-known problem calculated with double precision calculations.

The vertical broken line in the graph indicates the theoretical difference in performance, but the speed data shows that in three out of five tasks the speed of the new GPU was significantly higher. This is due to all the changes aimed at increasing efficiency: moving away from VLIW, the presence of a scheduler in each computing unit, improved caching, etc.

Changes in rendering quality

Actually, this part could well have been skipped, since there are no special claims to image quality lately and cannot be - for various reasons. For example, the quality of full-screen anti-aliasing for video cards from different manufacturers is very close, especially considering the widespread use of software anti-aliasing methods using post-processing filters, which are performed on all GPUs in exactly the same way.

The same applies to texture filtering - now its quality is such that it's very difficult to distinguish between AMD and NVIDIA solutions even if you do a pixel-by-pixel comparison. The Radeon HD 6900 - the previous generation of the company - has improved anisotropic filtering a little more, and now even a "microscope" will not help to find any significant flaws there. The only note is that in motion the Radeon video cards were slightly inferior to the GeForce due to more noticeable specific artifacts, such as "noise" or "sand".

With the release of the new generation of video chips, the texel weights in the texture filter were revised once again, modified in such a way as to reduce such artifacts, sometimes visible on the Radeon HD 6900 in the presence of textures of a certain type (“high-frequency”, with sharp transitions from dark to light, for example). The change in quality is so hard to show with examples that AMD doesn't provide HD 7900 versus HD 6900 comparison pictures, but simply compares the quality of the "hardware" algorithm with a purely software algorithm running on GPU stream processors, and therefore ideal:

On such a small screenshot, the difference in quality is not visible, but AMD assures that all the changes made did not bring any performance drop and did not worsen the image quality in any of the aspects - it still does not depend on the angle and the filtering quality is close to ideal. In one of the future practical materials, we will definitely check this.

Partially Resident Textures

The idea of ​​Partially Resident Textures (PRT) is to use the hardware capability of the presented GPU - virtual memory. Surely many users have already seen id Software's RAGE game, which uses virtual texturing technology, the so-called megatexturing ("MegaTexture"), which provides the ability to use huge amounts of texture data and swap (streaming) them into video memory.

Using virtual video memory, it is very easy to get effective hardware support for such algorithms, which allow using up to 32 terabytes of textures in an application, which makes it possible to make unique locations in games, without repeating pieces of textures, while total absence problems loading texture data. True, AMD's illustrative example is too strange, from which nothing is particularly clear:

PRT allows you to achieve high picture quality and helps to increase the efficiency of video memory usage. Similar algorithms are already used in the id Software engine, and are expected to appear in many next generation engines. Games of the future need to work with huge amounts of data, and the advantage of the new GPU is that local graphics memory in PRT algorithms works like hardware cache memory, and textures are loaded into it when necessary. GPUs from the Southern Islands family support "mega-textures" up to 32 terabytes (resolution up to 16384×16384) and, most importantly, hardware texture filtering for them, which is not available on earlier video chips.

Virtual textures are divided into pieces of 64 kilobytes (kilobytes, not texels) and this piece size is fixed. And only those that are needed when rendering the current frame are loaded into the local memory of the video card. The technology works regardless of the texture format, just the sizes of the pieces in texels will be different. For example, for a regular uncompressed texture with 32 bits per color, the chunk size will be 128x128 texels, and for a DXT3-compressed texture it will be 256x256 texels.

The technology also involves the use of mip-levels of textures (smaller copies used in texture filtering). When rendering and filtering, they need to be accessed multiple times. Consider the operation of the algorithm on an example.

This figure highlights four different pieces from different mip levels required for rendering. When the shader program requests data from them, some of the pieces are already in local memory and this data is immediately sent to the shader for further calculations. But some pieces are missing from the table, and the application must decide what to do next on such a miss. For example, you can request data from a lower resolution mip-level, then the image will be fuzzy, but at least it will look like the truth and will be drawn without delay. And by the time the next frame is rendered, it can already be loaded into the cache - local video memory. Those who played RAGE will understand us.

This is a powerful algorithm that allows you to use huge textures that are unique to each of the objects. Similar algorithms have long been used in offline rendering, except for the need for real-time calculations. AMD even made a demo using the Per-Face Texture Mapping technique developed by Walt Disney Animation Studios for their animated films. Unfortunately, the demo isn't ready yet, and we've only seen low-res screenshots.

The essence of this texture mapping technique is to assign a certain piece of texture to each polygon, without the need to use UV-transformation (finding a correspondence between the coordinates of the surface of a three-dimensional object and the coordinates on a two-dimensional texture). This approach solves some of the problems with creating tessellated content by making the displacement mapping algorithm very simple. And PRT in this method is used for efficient storage and access to texture data.

Media Handling Instructions

An interesting innovation in Southern Islands seems to be support for specialized instructions used in image processing, both static and dynamic. For example, a widely used instruction called "sum of absolute differences", better known as SAD (Sum of Absolute Differences), has been improved. The speed of its execution is a very performance-critical bottleneck in many image and video data processing algorithms, such as motion detection, gesture recognition, image search, computer vision, and many others.

But in our review of the ancient video card Radeon HD 5870, we already wrote about SAD support. Now, in addition to the usual SAD (4 × 1), Southern Islands has new instruction- QSAD (quadruple SAD), which combines SAD with shift operators to increase performance and power efficiency, as well as the "mask" instruction MQSAD, which ignores background pixels and is used to isolate objects moving in the frame from the background.

New GPUs can process up to 256 pixels per GCN compute unit per clock, which in the case of the model AMD Radeon HD 7970 means the ability to process up to 7.6 trillion pixels per second in the case of 8-bit integer color values. Although this is a theoretical figure, the visual processing capabilities of the new GPUs are quite impressive - many video processing tasks can be performed in real time.

PCI Express 3.0

We couldn't pass by the support of the third version of PCI Express by the whole line of new graphic solutions from Southern Islands. This support was quite expected, since the specifications of the third version of PCI Express were finally approved in the fall of 2010, but there were still no hardware solutions with its support, although motherboards are already appearing, video cards were released at the end of 2011, and there are corresponding central processors.

The updated interface has a transfer rate of 8 gigatransactions per second instead of 5 GT / s for version 2.0, and it throughput once again doubled (up to 32 Gb / s), compared with the PCI Express 2.0 standard. The new bus uses a different coding scheme for data sent over the bus, but compatibility with previous versions of PCI Express has been preserved.

First motherboards with PCI support Express 3.0 was introduced in the summer of 2011, in the main base Intel chipset Z68, and they appeared on wide sale only in the autumn of the same year. So video cards arrived in time, and AMD again became ahead of the rest in terms of the speed of release of new graphics processors with support for the most advanced technologies. But it's too early to judge whether PCI-E 3.0 will be of any practical use.

AMD PowerTune technology

One of the most interesting innovations in the Cayman was PowerTune advanced power management technology. Flexible GPU power management has been used for a long time, but before the Radeon HD 6900, all these technologies were rather primitive and mostly software methods and changed the frequency and voltage in steps, not being able to turn off large parts of the video chips.

Even in the Radeon HD 5000 family, a performance limiter appeared when a certain consumption level was exceeded, and in the Radeon HD 6900 the system moved to a qualitatively different level. To do this, special sensors were included in the chip in all blocks that monitor boot parameters. The GPU constantly measures the load and power consumption and does not allow the latter to go beyond a certain threshold, automatically adjusting the frequency and voltage so that the parameters remain within the specified thermal package.

Unlike early power management technologies, PowerTune provides direct control over GPU power consumption, as opposed to indirect control by changing frequencies and voltages. This technology helps to set high GPU frequencies, get high performance in games, and not be afraid that consumption may go beyond safe limits. After all, most games and regular applications that use GPU computing have significantly lower power requirements and do not approach dangerous power consumption limits, unlike stability tests like Furmark and OCCT.

Even the heaviest games do not require maximum power consumption, and if you limit consumption by frequency, testing video cards with extreme tests, then in the case of 3D games, there will be quite a lot of unused performance and power opportunities. In the event that the video card has not reached the limit of the safe consumption level, the GPU will operate at the frequency set at the factory, and in the FurMark and OCCT tests, the GPU frequency will decrease to stay within the consumption limits.

Thus, PowerTune helps to set higher factory frequencies and tune the system for the most efficient use of GPU resources at the maximum power level set. In the example shown above, the HD 5870 does not use PowerTune and due to the GPU frequency limitation of high consumption in endurance tests does not use its full potential. While the maximum TDP is set for the Radeon HD 7970, and the video chip resets the frequencies only when it is exceeded, getting the highest possible performance in any application.

This is clearly shown in the following diagram. In the case of gaming applications, TDP can be achieved by increasing the frequency of the GPU, and for peak loads, endurance tests reduce the frequency to a safe level of power consumption. Without PowerTune, you would have to choose - either to get the probability of a video card failure when FurMark and OCCT are running for a long time, or to cut down on the potential performance in games. New technology solves these issues as efficiently as possible.

AMD PowerTune features fast responsiveness to changing conditions (microseconds) as it is a hardware technology. It is also distinguished by flexible frequency tuning, and not stepped, as it was in previous chips. All measurements are independent of the driver, but can be adjusted by the user using the video card settings.

The difference between PowerTune and the previously generally accepted approach is that in other cases, thermal throttling is used, which puts the GPU in a significantly reduced consumption mode, and PowerTune simply smoothly reduces its frequency, bringing the GPU consumption to the set limiter. This results in higher clock speeds and higher performance.

AMD ZeroCore Technology

AMD did not limit itself to using power management technology already known from previous solutions. In the first chips of the Southern Islands family, it introduces AMD ZeroCore technology, which helps to achieve even greater energy efficiency in the "deep idle" (or "sleep") mode with a disabled display device, which is supported by all operating systems.

After all, almost any system, even a gaming one, spends most of the time in low-load mode on the graphics processor. And the video card should not consume much power in this mode. And even more so, not to mention the mode with the monitor turned off - in this case, it is advisable to turn off the GPU altogether. That's what AMD did. Thanks to ZeroCore, the new GPU consumes less than 5% of the power in full mode when in deep idle, disabling most of the functional blocks in this mode.

AMD provides a schematic comparison with its own Radeon HD 5870, which did not support this technology. ZeroCore is a Southern Islands-exclusive innovation in desktop solutions from mobile GPUs designed for laptops. By the way, the advantages of this technology are associated not only with reduced consumption. In addition, in the long idle mode, when the display is turned off, the video card also completely turns off the fan on the video card cooler!

This is exactly what many users have been waiting for a long time. The most interesting thing is that according to our data, laboratory tests of solutions like PowerTune and ZeroCore took place several generations of video cards ago. Some of the engineering samples of video cards from AMD's series that have long since left the market worked exactly like this, completely turning off the cooler in idle time.

But it's not just single-GPU users who benefit from noise reduction and power consumption with AMD's new ZeroCore-enabled graphics cards. Similar improvements await happy owners of CrossFire systems based on two, three, and even four GPUs. Is it logical, after all, that in the mode of rendering the two-dimensional interface of the operating system, all video cards, except for the main one, should not work at all? But that's how they work now!

In the case of CrossFire systems on video cards with ZeroCore support in 2D mode, all secondary video cards are immersed in a deep sleep with minimal power consumption and a disabled cooler. This mode works both for several single-chip video cards and for two-chip solutions. In addition, the primary CrossFire graphics card will also enter this mode in the event of a long idle time configured in Windows. Visually, the difference in work looks like this:

By the way, the technology is not as simple as it might seem. AMD engineers had to solve a lot of issues related to the operation of the operating system in idle mode. For example, they found that Windows tries to update information on the screen even when the monitor is turned off. Which, of course, does not allow you to disable the GPU at all. Therefore, the company's programmers had to take a workaround, ignoring all screen drawing commands when the monitor was turned off in sleep mode.

AMD Eyefinity 2.0 Technology

Naturally, in the new architecture there was a place for improvements to the proven technology for displaying images on multiple monitors - AMD Eyefinity, now in version 2.0. It has received new features, higher resolutions, support for more displays and increased flexibility.

This technology is quite interesting, although an extremely small number of users will find space in the room and muster up the courage in front of the family to install more than two monitors. But it is better to have the opportunity to always be able to use it than not to have it at all. Moreover, the prices for monitors of large diagonals almost do not decrease, but mid-range solutions are constantly getting cheaper.

Indeed, now it is more profitable to buy three monitors with a screen diagonal of 24″ than one 30-inch one. AMD gives just such an example, where a 30″ 2560x1600 monitor costs over $1,000, while three 24″ FullHD monitors can be bought for half that price:

But how to spend your money and space in the room is a personal matter for each user. The main thing is that there is such an opportunity. Plus, Eyefinity 2.0 now supports image output in HD3D stereo mode - something that was lacking in previous solutions, which were inferior to competing ones in this parameter. Combining AMD Eyefinity and HD3D technologies, the Radeon HD 7970 is the first single-chip solution to support three monitors in stereo mode.

High resolution stereo rendering requires a very fast data transfer interface. And with previous HDMI versions outputs, the possibilities were limited to 24 Hz per eye, which is quite enough for watching movies on Blu-ray 3D, but clearly too little for gamers.

For such tasks, they began to use the frame packing format, when the frames for the left and right eyes are combined into one, and AMD Radeon HD 7970 supports the HDMI 1.4a frame packing format for stereo output. This is the first video card to support 3 GHz HDMI with frame packing, where each eye has a FullHD picture at 60 Hz (120 Hz in total):

Another interesting novelty seems to us to be Discrete Digital Multi-Point Audio (DDMA) multi-channel audio output technology, which works together with Eyefinity. All previous GPUs are capable of outputting via HDMI and DisplayPort only one audio stream. That is, even if three monitors located in different rooms are connected to the PC via HDMI, then sound channel only one is transmitted. But the AMD Radeon HD 7900 received support for the simultaneous output of several independent audio channels at once, which may well come in handy in some multi-monitor configurations.

The same feature will be very useful for video conferencing applications with the output of several interlocutors on separate screens, as well as multitasking applications such as playing on three monitors with game audio and watching news on a separate screen with an independent audio stream. Previously, for all this it was necessary to use several separate audio systems, but now everything works as conveniently as possible.

The software support of Eyefinity is also not forgotten, almost every month the technology is updated - new opportunities appear. So, back in October, support for resolutions up to 16384 × 16384 and new multi-monitor configurations appeared: horizontal and vertical 5 × 1, as well as based on six monitors in 3 × 2 mode.

The AMD Catalyst video driver update in December brings Eyefinity and HD3D to work together, and in February, support for custom resolutions, taskbar placement tweaks, and improvements to preset management are announced.

Output to six monitors can be achieved using two DisplayPort 1.2 ports and two MST hubs (which we wrote about earlier), while three or even four monitors require only one port and the corresponding hub. These hubs offer flexible display configurations, support up to four FullHD devices per DisplayPort 1.2 connector, and should be available by summer 2012.

Speaking of permission. High resolution or even ultra-high - Ultra High Resolution. Current devices with a resolution of 4000 pixels on the larger side require connection using several cables at once: two DP 1.1 or four DVI. Monitors of this resolution of the next generation will be connected using only one cable: DP 1.2 HBR2 or HDMI 1.4a 3 GHz. And new graphics card AMD is already ready for such monitors, again it became the first in the world.

Video encoding and decoding

It is quite natural that the AMD Radeon HD 7970 includes the same UVD block for decoding video data, which appeared in the previous generation of the company's video chips. It simply doesn't need any modifications, supporting the MVC multi-stream codec, MPEG-2/MPEG-4 (DivX), VC-1 and H.264 decoding, as well as decoding two FullHD streams in all supported formats.

AMD solutions provide maximum quality video stream decoding, use dozens of special quality improvement algorithms and provide maximum results in quality tests like HQV. Among the supported features, we note: color and tone adjustment, noise reduction, sharpening, high-quality scaling, dynamic contrast, advanced deinterlacing, and inverse telecine. Here is an example of on-the-fly contrast enhancement:

But with decoding, all video chips have been more or less in order for a long time. All new GPUs provide decent quality and performance when viewing video data. But video encoding on the GPU is still in its infancy, and the main complaints from users are directed to the low quality of the resulting compressed image.

Perhaps the new Radeon HD 7000 series can help with this, because all the graphics processors in the series have a Video Codec Engine (VCE) video encoding unit. The Radeon HD 7970 was the first video card to support hardware-accelerated encoding and video compression using a specialized block (earlier, stream processors took part in encoding).

Quality and performance should be clearly better than before, with support for 1080p at 60fps encoding, even faster than real-time. It's hard to say anything about quality without tests, but we are promised different levels of encoder optimization for video data and games, as well as variable compression quality (the ability to choose between improving quality or performance).

So far, there is no place to try VCE - there are simply no applications with support for it, but AMD is working with partners such as ArcSoft to provide support for VCE in relevant software products. In the future, we plan to release a software library for accelerating video encoding, which will make it easier for developers to support next-generation AMD products.

Encoding can be done in two modes: full and hybrid (using the capabilities of GPU stream processors). Full mode is designed for applications that require maximum energy efficiency and a consistent level of performance. Encoding in full mode on VCE is faster than real time and provides low latency. But there is also a hybrid mode:

In this mode, GPU math blocks work together with VCE. All highly parallelizable stages, which are circled in yellow in the diagram, can use the power of the GCN computational units, and the dedicated VCE unit is engaged in efficient hardware entropy coding. This mode is well suited for video cards with great mathematical power, such as the Radeon HD 7970. There are still questions about the quality of these two modes, but this requires a thorough analysis in a separate article.

AMD Steady Video

In addition to encoding and decoding video data, there is another area where the power of AMD's new graphics can be used - improving poor-quality handheld videos without the use of a tripod or other similar image stabilization tools. The video stabilization technology is called AMD Steady Video, and its second version has already been released.

The algorithm of the software stabilizer is quite simple: based on the video stream, statistics are collected about the movement of the camera (shift, rotation, zoom) and this movement is compensated in the current frame, relative to the previous ones - the image is shifted, rotated and scaled so that the picture does not jump much and remains stable.

As simple as it is in words, it is just as difficult to implement. Simply because there are two million pixels on the screen, and up to 30 or even 60 frames per second. Imagine how many calculations you need to do to track all possible frame shifts. We have already written above about the QSAD function used in video processing, and it is also used in Steady Video 2.0 to speed up the motion detection algorithm. So the GPU must process random shifts with an amplitude of up to 32 pixels in any direction, and this requires a performance corresponding to more than 500 billion SAD operations per second (for 1920x1080 at 60 FPS).

By supporting the new QSAD instructions in the Radeon HD 7970, its advantage over powerful CPUs in the motion detection algorithm exceeds 10x! That is, high-quality video will now be provided to us, and not only when processing home videos in video editors, but also when watching other people's online videos, shot by no one knows what and no one knows how.

Details: Radeon HD 7800 series

• Chip codename: "Pitcairn"
• Production technology: 28 nm
• 2.8 billion transistors (slightly more than the Cayman, which is the basis of the Radeon HD 6900 series)
• Unified architecture with an array of common processors for streaming processing of multiple types of data: vertices, pixels, and more.
• Hardware support for DirectX 11.1, including shader model Shader Model 5.0
• 256-bit memory bus: four 64-bit wide controllers with GDDR5 memory support
• Core clock: up to 1000 MHz (for Radeon HD 7870)
• 20 GCN Compute Units with 80 SIMD cores for a total of 1280 floating point ALUs (Integer and Float formats, IEEE 754 FP32 and FP64 precision support)
• 80 texture units, with support for trilinear and anisotropic filtering for all texture formats
• 32 ROPs with support for anti-aliasing modes with the possibility of programmable sampling of more than 16 samples per pixel, including with FP16 or FP32 framebuffer format. Peak performance up to 32 samples per clock, and in colorless mode (Z only) - 128 samples per clock

Radeon HD 7870 Graphics Specifications

• Core frequency: 1000 MHz
• Number of universal processors: 1280
• Number of texture units: 80, blending units: 32
• Memory type: GDDR5
• Memory capacity: 2 gigabytes
• Theoretical maximum fill rate: 32.0 gigapixels per second.
• Theoretical texture sampling rate: 80.0 gigatexels per second.
• One CrossFire connector
• PCI Express 3.0 bus
• Connectors: DVI Dual Link, HDMI 1.4, two Mini-DisplayPort 1.2
• Power consumption: 3 to 175 W
• Two 6-pin power connectors
• Dual slot design
• US MSRP: $349

Radeon HD 7850 Graphics Specifications

• Core frequency: 860 MHz
• Number of universal processors: 1024
• Number of texture units: 64, blending units: 32
• Effective memory frequency: 4800 MHz (4×1200 MHz)
• Memory type: GDDR5
• Memory capacity: 2 gigabytes
• Memory bandwidth: 153.6 gigabytes per second
• Theoretical maximum fill rate: 27.5 gigapixels per second.
• Theoretical texture sampling rate: 55.0 gigatexels per second.
• One CrossFire connector
• PCI Express 3.0 bus
• Connectors: DVI Dual Link, HDMI 1.4, two Mini-DisplayPort 1.2
• Power consumption: 3 to 130 W
• Dual slot design
• US MSRP: $249

And this time, the principle of naming the company's products was not changed and the trends of the previous series were continued. The mid-budget series of video cards based on the GCN architecture differs from the top and budget lines in the second digit in the index: instead of 7 and 9, the number 8 is set, which is quite logical. Since AMD has taken the psychological threshold of 1000 MHz for the frequency of the GPU, the Radeon HD 7870 received the addition of "GHz Edition" to the name, indicating the adoption of this frequency.

From the name it is clear that the Radeon HD 7800 is more productive than the HD 7700, but has a lower speed compared to the older models - HD 7900. As for comparison with NVIDIA solutions, the older HD 7870 released at the time of release competes with the video card GeForce GTX 570, and the younger one is aimed at fighting the GTX 560 Ti, and NVIDIA has not yet released new 28 nm mid-range chips.

Both models of video cards from AMD have GDDR5 memory of the same amount of 2 gigabytes. They both use a 256-bit memory bus, and so you could put 1, 2, or 4 GB on them. 1 gigabyte is too small, and 4 GB is too expensive for this price segment. Therefore, we can say that the ideal amount of 2 GB of video memory has been chosen, which is quite sufficient for the vast majority of games even at high resolutions, and not too costly in terms of cost.

In other respects, from the point of view of the consumer, the HD 7850 and HD 7870 models are still different. The older Radeon HD 7870 has a higher power consumption, so it needs two additional 6-pin power connectors, and the HD 7850 is content with only one of them. Both boards have a two-slot cooling system design, but most manufacturers produce boards with their own design of at least a cooler, and even a PCB.

Architectural features of the Radeon HD 7800 family

Above, we have carefully described all the features of the new Graphics Core Next (GCN) architecture, so we will repeat only the most important. All of the company's new GPUs offer excellent features and performance not only in graphics processing, but also in non-graphical computing, including a mixture of different types of calculations. Also, the new GCN architecture offers a serious simplification of code optimization tasks, simplification of development and support, as well as stable and predictable performance and, in general, quite high efficiency.

The base block of the new architecture is the GCN block, and all GPUs of the Southern Islands series are assembled from them. Consider the block diagram of the Pitcairn chip:

The diagram shows the Radeon HD 7870 GPU (the "simplified" HD 7850 differs from it by several disconnected blocks), we see 20 computing units of the GCN architecture. In the case of the junior solution of the Radeon HD 7800 series, four of them were disabled, and the number of active blocks in it is 16. This corresponds to 1280 and 1024 stream processors, respectively (just like in the case of the HD 7700 family, only there are exactly twice as many blocks) . Since each GCN unit has four texture units, the total number of TMUs for the older model is 80 TMUs, and for the younger one - 64 TMUs.

But the number of ROPs and memory controllers in the HD 7870 and HD 7850 is also the same as in the solutions of the youngest line. The number of ROP blocks was left quite high - 32 pieces for both models. The memory bus for boards based on Pitcairn has been cut down to 256-bit, it is assembled from four 64-bit channels. This is not bad for a solution of this level, although it is one and a half times less than in the top line, because the memory bus is traditionally cut down first. It's good that the use of fast GDDR5 memory gave a relatively high bandwidth of 153 GB / s.

Like the rest of the GCN architecture chips, Pitcairn incorporates a 9th generation tesselator block, featuring numerous buffering and caching optimizations, which can significantly improve geometry processing performance. Here is a comparison of the new AMD board with the solution of the previous generation in a synthetic problem, according to which we can assume an increase in the tessellation speed up to four times:

A lot of AMD technologies that have been introduced and improved in the new Radeon HD 7000 video chips are also supported in the same way. Here is an incomplete list of them: PowerTune, ZeroCore, Eyefinity 2.0, HD3D, Steady Video, texture filtering quality improvements, etc. All this is described in more detail above. To add to the list, the Radeon HD 7800 fully supports both the improved MLAA 2.0 anti-aliasing algorithm and supersampling anti-aliasing (SSAA).

As far as gaming performance is concerned, the Radeon HD 7870 is significantly faster than its direct competitor GeForce GTX 570, especially given the latter's 1.25 GB of VRAM (compared to 2 GB for the solutions in question) observed in modern games at high rendering resolutions . The younger Radeon HD 7850 can be compared with the GeForce GTX 560 Ti, and here it can no longer boast of the amount of memory. However, according to AMD's measurements, their new solution is still faster than the competition in most games.

Details: Radeon HD 7700 Series

• Chip codename: "Cape Verde"
• Production technology: 28 nm
• 1.5 billion transistors (less than Barts, which is the basis of the Radeon HD 6800 series)
• Unified architecture with an array of common processors for streaming processing of multiple types of data: vertices, pixels, and more.
• Hardware support for DirectX 11.1, including shader model Shader Model 5.0
• Core clock: up to 1000 MHz (for Radeon HD 7770)
• 10 GCN Compute Units with 40 SIMD cores for a total of 640 floating point ALUs (Integer and Float formats, IEEE 754 FP32 and FP64 precision support)
• 40 texture units, with support for trilinear and anisotropic filtering for all texture formats
• Integrated support for up to six monitors including HDMI 1.4a and DisplayPort 1.2

Radeon HD 7770 Graphics Specifications

• Core frequency: 1000 MHz
• Number of universal processors: 640
• Number of texture units: 40, blending units: 16
• Memory type: GDDR5
• Memory capacity: 1 gigabyte
• Theoretical texture sampling rate: 40.0 gigatexels per second.
• One CrossFire connector
• PCI Express 3.0 bus
• Connectors: DVI Dual Link, HDMI 1.4, two Mini-DisplayPort 1.2
• Power consumption: 3 to 80 W
• One 6-pin power connector
• Dual slot design
• US MSRP: $159

Radeon HD 7750 Graphics Specifications

• Core frequency: 800 MHz
• Number of universal processors: 512
• Number of texture units: 32, blending units: 16
• Effective memory frequency: 4500 MHz (4×1125 MHz)
• Memory type: GDDR5
• Memory capacity: 1 gigabyte
• Memory bandwidth: 72 gigabytes per second
• Theoretical maximum fill rate: 12.8 gigapixels per second.
• Theoretical texture sampling rate: 25.6 gigatexels per second
• PCI Express 3.0 bus
• Connectors: DVI Dual Link, HDMI 1.4, one DisplayPort 1.2
• Power consumption: 3 to 55 W
• Does not require additional power
• Single slot design
• US MSRP: $109

The low-cost series of video cards based on the GCN architecture differs from the top and middle lines in the second digit in the index: the 9th place was taken by the number 7, as it was before. The Radeon HD 7770 is a more productive solution, but there is also a younger model - the HD 7750. The older board had no direct competitors on the market at the time of release, being located somewhere between the GeForce GTX 560 and GTX 550 Ti, and the younger one is aimed at fighting the GTX 550 Ti. For the HD 7770, a competitor was later announced in the face of the GeForce GTX 560 SE (all NVIDIA solutions are based on older GPUs).

Both considered models of AMD video cards have GDDR5 memory of the same amount of 1 gigabyte. Due to the use of a 128-bit memory bus, 2 GB of memory could be put on them, but this amount of GDDR5 memory will cost too much for their price segment. Therefore, models with such a volume have so far been released, although options with 2 GB of video memory may be released in the future. In the meantime, we decided to leave this volume for the HD 7800.

In terms of other consumer characteristics, the HD 7750 and HD 7770 models are quite different. If the older Radeon HD 7770 has a two-slot design of the cooling system and its cooler is covered with a plastic casing like in older solutions, then the younger HD 7750 looks noticeably simpler, occupying one slot and having a simple cooler. However, most manufacturers still produce boards with their own designs. The power consumption of the new models in this price range is also different, the older one has one 6-pin auxiliary power connector, and the younger one is powered by PCI Express.

Architectural features Radeon HD 7700

The base block of the new architecture is the GCN block, and all GPUs of the series are assembled from them. Each of the available GCN blocks is capable of scheduling and distributing instructions on its own, and one computing unit can execute up to 32 independent instruction streams. Let's look at the block diagram of the Cape Verde chip:

The diagram shows the Radeon HD 7770 GPU (the “stripped down” HD 7750 features a few disconnected units), we see 10 compute units of the GCN architecture. In the case of the junior solution of the Radeon HD 7700 series, it was decided to disable two of them, and the number of blocks became 8. This corresponds to 640 and 512 stream processors. And since each GCN unit has 4 texture units in its composition, the final figure for the number of TMUs for the older model is 40 TMUs, and for the younger one - 32 TMUs.

The number of ROPs and memory controllers in the HD 7770 and HD 7750 is the same, and we decided not to cut the ROPs too much, leaving them at 16 each. But the memory bus at Cape Verde is cut down to 128-bit, which is assembled from two 64-bit channels. In general, this is three times less than in the top series, and we saw another confirmation that the memory bus is traditionally cut down in inexpensive chips first of all. Although the use of fast GDDR5 memory made it possible to leave a relatively high (for such inexpensive solutions) bandwidth of 72 GB / s.

It remains for us to note a rather large amount of L2 cache - as much as 512 kilobytes (compared to 768 KB for a top-end chip - apparently, the L2 cache does not take up too much space on the chip), as well as improvements in geometric performance. Like the top-of-the-line chip, Cape Verde features a 9th generation tesselator featuring multiple buffering and caching optimizations to deliver a notable improvement in geometry processing performance over the Radeon HD 6000 series.

In general, we will not repeat all the information about AMD technologies that have been implemented and improved in the new Radeon HD 7000 video chips (here is a partial list: PowerTune, ZeroCore, Eyefinity 2.0, HD3D, Steady Video, texture filtering quality improvements, etc. .p.), all this is described in more detail above. The HD 7700 series supports all the features listed there, including AMD Eyefinity 2.0 with six monitors and stereo rendering, as well as an improved video decoding and encoding unit.

But what about the most important thing - performance in games? The first estimates of rendering speed can always be made from the manufacturer's presentations. AMD believes that the Radeon HD 7770 is somewhere in the middle between the GeForce GTX 560 and GeForce GTX 550 Ti, respectively, and compares it in its materials with the second competitor model.

But they do not compare the Radeon HD 7750 with anything, simply noting that most modern games are playable on this model at maximum settings in FullHD resolution. However, this is not surprising, since last years There are practically no PC exclusives, and multi-platform games are significantly less demanding. So the Radeon HD 7700 series boards are perfect for undemanding users.

Details: Model Radeon HD 7790

• Chip codename: "Bonaire"
• Production technology: 28 nm
• 2.08 billion transistors (more than Cape Verde in the Radeon HD 7700, but less than Pitcairn in the Radeon HD 7800)
• Unified architecture with an array of common processors for streaming processing of multiple types of data: vertices, pixels, and more.
• Hardware support for DirectX 11.1, including shader model Shader Model 5.0
• 128-bit memory bus: two 64-bit wide controllers with GDDR5 memory support
• Core frequency: 1000 MHz
• 14 GCN Compute Units of 56 SIMD Cores with a total of 896 Floating Point ALUs (Integer and Float formats, supports IEEE 754 FP32 and FP64 precision)
• 56 texture units, with support for trilinear and anisotropic filtering for all texture formats
• 16 ROPs with support for anti-aliasing modes with the possibility of programmable sampling of more than 16 samples per pixel, including with FP16 or FP32 framebuffer format. Peak performance up to 16 samples per clock, and in colorless mode (Z only) - 64 samples per clock

Radeon HD 7790 Graphics Specifications

• Core frequency: 1000 MHz
• Number of universal processors: 896
• Number of texture units: 56, blending units: 16
• Memory type: GDDR5
• Memory capacity: 1 gigabyte
• Memory bandwidth: 96 gigabytes per second
• Theoretical maximum fill rate: 16.0 gigapixels per second.
• Theoretical texture sampling rate: 56.0 gigatexels per second.
• One CrossFire connector
• PCI Express 3.0 bus
• Connectors: DVI Dual Link, HDMI 1.4, two Mini-DisplayPort 1.2
• Power consumption: 3 to 85 W
• One 6-pin power connector
• Dual slot design
• US MSRP: $149

An inexpensive video card model based on a new mid-budget chip differs from the previous top model of the HD 7700 subfamily by the third digit in the index: instead of 7, they put the number 9, which indicates an increase in performance. At the same time, the Radeon HD 7790 index clearly indicates that this is a less productive video card compared to the one step higher line - HD 7800.

However, everything is not so simple here either - it will certainly be able to argue with the younger HD 7850. But the recommended price of the Radeon HD 7790 is set at $149, that is, approximately in the middle between the prices of the HD 7770 and HD 7850. As for the competitor's solutions from the same price segment, the HD 7790 was clearly intended to have NVIDIA GeForce The GTX 650 Ti, based on the GK106 chip, sits right between the HD 7770 and HD 7850 in terms of price and speed. But NVIDIA immediately responded to the release of the new board by AMD by launching an overclocked version of the GeForce GTX 650 Ti Boost, which is characterized by greater performance.

This AMD graphics card model has GDDR5 memory with a capacity of only 1 gigabyte. The GPU has a 128-bit memory bus, and theoretically 2 GB could be supplied, but this amount of fast GDDR5 memory is still too expensive for this price segment, and AMD has released a model with a smaller memory, although it may not be enough for some modern games even at low settings and resolutions. However, video cards from partners with 2 GB of video memory are also possible.

Like the models standing next to it in the line, the Radeon HD 7790 has a dual-slot design of the cooling system, which is covered with a plastic casing. Although most manufacturers still release boards with their own cooler design, so the reference one is not that important. Interestingly, the new model's power consumption hasn't increased much compared to the HD 7770, but the improvement in power efficiency was to be expected. By the way, that's why the novelty also has only one 6-pin auxiliary power connector.

architectural features

The new Bonaire GPU, on which the released Radeon HD 7790 is based, belongs to the same Graphics Core Next (GCN) architecture that we have known for a year and a half, but AMD calls it GCN 1.1, hinting at minor changes. In fact, the chip is architecturally almost the same as the previous ones, although there are indeed some minor changes. For example, the new architecture introduced instructions that are useful for heterogeneous architecture (Heterogeneous System Architecture - HSA), support for more simultaneously executing threads, as well as a new version of AMD PowerTune technology, which we will talk about later. But all these changes cannot be called significant, because there is nothing new in the basic blocks and improving their efficiency.

Therefore, we can safely refer to, which carefully describes all the features of the new Graphics Core Next (GCN) architecture, and here we will only repeat the most important characteristics and features of a particular product. All of AMD's latest GPUs offer excellent features and performance in both graphics and non-graphics processing, including mixtures of the two. The new GCN architecture also provides a major simplification of optimization and software development tasks, while maintaining high efficiency.

As you know, the basic block of the architecture is the GCN block, from which all the GPUs of the Southern Islands series are assembled. The GCN computing unit is divided into subsections, each of which works on its own instruction stream. GCN blocks have a dedicated 64 KB local data storage for data exchange or local register stack expansion. Also, the block has a first-level cache memory with the ability to read and write and a full-fledged texture pipeline with sampling and filtering blocks. Each of the existing GCN units is capable of scheduling and distributing commands on its own, and one computing unit can execute several independent instruction streams. Let's look at the block diagram of the new chip:

The Bonaire scheme confirms the goal of the new solution to offer performance between Cape Verde, which has 10 GCN compute units, and Pitcairn, which has 20 GCN units. These two GPUs, released in 2012, differ from each other by almost exactly half, so there was a rather large performance gap in the middle between them, which Bonaire has now filled.

The diagram shows the graphics processor in the form of a Radeon HD 7790, which is a complete solution without cutting any blocks. The chip includes 14 computing units of the GCN architecture, which corresponds to 896 stream processors. Since each GCN has 4 texture units, the total number of TMUs for the new model is 56 TMUs. That is, Bonaire is exactly 1.4 times faster than the Cape Verde chip in terms of the speed of mathematical calculations and texture fetches, provided that the frequency is equal.

But the number of ROP units and memory controllers in Bonaire and Radeon HD 7790 is similar to what we saw in Cape Verde and Radeon HD 7770 - it was decided to leave 16 ROP units, and the memory bus of the new chip is 128-bit, assembled from two 64- bit channels. The small number of ROPs can be the “Achilles heel” of the solution, since the use of fast GDDR5 memory made it possible to provide a relatively high throughput of 96 GB / s, but nothing can be done about ROP performance.

But in the new GPU there are improvements in geometric performance and tessellation speed. Yes, Cape Verde also has a 9th generation tesselator, but Bonaire also doubled the number of geometry blocks, rasterizers and command processors (shown as ACE in the diagram) - now there are two of them. This improvement gives Bonaire the ability to process up to two geometric primitives per clock - just like the more powerful Pitcairn and Tahiti.

As you remember, it was in the Radeon HD 7770 that AMD first took the important psychological threshold of the GPU clock frequency, equal to 1 GHz. So, the HD 7790 also has exactly the same reference frequency of 1 GHz, so the increase in performance compared to the HD 7770 will be justified solely by architectural changes and an increase in the number of execution units.

But the frequency of the video memory of the novelty is much higher. If the HD 7770 had a relatively low memory frequency of 4.5 GHz, then the HD 7790 is equipped with a fast GDDR5 memory operating at 6 GHz, which provides a third more bandwidth. The 33% increase in video memory bandwidth compared to the Radeon HD 7700 sub-family resulted in a clear increase in gaming performance. AMD provides this chart, comparing HD 7790 frame rates with memory running at 4.5 and 6.0 GHz:

The maximum speedup from the increase in memory bandwidth was achieved in games such as StarCraft II and Crysis 2. And on average, a 33% increase in memory bandwidth gives somewhere around 10% increase in average frame rate in a set of modern games. Not a bad indicator, showing that memory bandwidth is quite important in our time, although it is not the only emphasis on performance. Although it is quite possible that with more ROP, Bonaire's speed would be even higher...

It is clear that the average power consumption has increased slightly compared to the HD 7770. If for the old model this value is 80 W, then for the HD 7790 it is 85 W - this is a very small price to pay for a theoretical performance increase of 33-40%! Architectural improvements (PowerTune), the design of a new GPU based on the experience of previous ones, as well as the continuous improvement of the technical process at TSMC - all this led to a small increase in consumption with a significant improvement in speed characteristics.

As for the area of ​​the chip and the number of transistors in Bonaire, the new chip is clearly larger than Cape Verde, but the addition of computational, texture and geometric units could not go unnoticed. According to these parameters, Bonaire is also located approximately in the middle between Cape Verde and Pitcairn. Bonaire contains 2.08 billion transistors in a 160 mm 2 chip, for Cape Verde these figures are 1.5 billion and 123 mm 2 , respectively, and for Pitcairn - 2.8 billion transistors and 212 mm 2 chip area.

Naturally, the new chip supports all AMD technologies that have been introduced and improved in the new Radeon HD 7000 family (their incomplete list: PowerTune, ZeroCore, Eyefinity, HD3D, Steady Video, texture filtering quality improvements, etc.), both all this is described in detail in the article AMD Radeon HD 7970: New single-processor leader. The HD 7790 model supports all the features listed there, including AMD Eyefinity 2.0 with six monitors and stereo rendering, and also has an improved video decoding and encoding unit.

Enhanced PowerTune Technology

Back in 2010, AMD introduced PowerTune technology in its Cayman chip (AMD Radeon HD 6900 series). This GPU was the first to feature dynamic power management called PowerTune. It allowed us to increase the maximum clock speeds for typical applications, while avoiding too much power consumption in specialized stability tests like FurMark. Then the technology was applied to the dual-chip model AMD Radeon HD 6990, which needed it even more for obvious reasons.

The technology received a major update in mid-2012, when automatic frequency increase - Boost - was added to AMD PowerTune. In the AMD Radeon HD 7970 GHz Edition, this algorithm allowed even more performance improvements compared to regular version video cards. The PowerTune operation algorithm in video cards without automatic overclocking uses three states: idle (idle), light load mode (low-3D) and full speed. In the HD 7970 GHz, the Boost overclocking mode was added to them. PowerTune serves to stay within the required consumption by switching to a lower load mode when necessary. In this case, the technology sharply reduces the clock frequency values. In practice, such jumps are rare - due to the large gap between the two active modes.

Reducing the GPU clock speed reduces power consumption, but you need to reduce the voltages for better control. This is exactly what the Radeon HD 7790 does. The new Bonaire graphics chip has eight states with different frequency and voltage settings, allowing for higher clock speeds than before, while ensuring that the GPU is always running at optimal voltage and frequency. Switching between states is based on GPU load as well as current GPU power consumption.

In the new algorithm, PowerTune does not have to abruptly drop the frequency when the consumption level is exceeded, and along with the frequency, the voltage also decreases. State transitions must be as fast as possible so as not to exceed the consumption limit even for a short time, so Bonaire switches PowerTune states every 10 ms, that is, in every second, the state of the chip changes 100 times.

With such a constant change in frequencies, third-party applications like MSI Afterburner and GPU-Z will not show instantaneous clock rates, but averages over a period of time - the so-called "effective" frequency. Another interesting innovation is that AMD is opening up new PowerTune settings for third party applications. Partners can also set their own PowerTune settings to help create factory overclocked graphics cards and provide more features beyond AMD's reference values. Truth, different settings PowerTune can lead to the fact that video cards of the same model from different manufacturers will not only have different clock frequencies, but also the algorithm for changing them over time, which makes it difficult to compare under the same conditions.

Sales of video cards of the Radeon HD 7790 model began on the market at the very beginning of April 2013. AMD, together with its partners, organized the release of both motherboards with reference frequencies and factory overclocked solutions. And now, both manufacturers are launching new video cards to the market in much the same way, with various options quickly available from their partners. In fact, partners have released almost more overclocked versions of the HD 7790 than conventional ones, and the graphics chips in them operate at frequencies of about 1075 MHz.

Details: Radeon HD 7990 Model

• Codename "Malta"
• Production technology: 28 nm
• 2 chips with 4.3 billion transistors each
• Unified architecture with an array of common processors for streaming processing of multiple types of data: vertices, pixels, and more.
• Hardware support for DirectX 11.1, including shader model Shader Model 5.0
• Dual 384-bit memory bus: twice six 64-bit wide controllers with GDDR5 memory support
• GPU frequency: 1000 MHz
• Twice 32 GCN Compute Units, including 128 SIMD cores, consisting of a total of 4096 floating point ALUs (integer and floating formats, support for FP32 and FP64 precision within the IEEE 754 standard)
• 2x128 texture units, with support for trilinear and anisotropic filtering for all texture formats
• 2x32 ROPs with support for anti-aliasing modes with the possibility of programmable sampling of more than 16 samples per pixel, including with FP16 or FP32 framebuffer format. Peak performance up to 64 samples per clock, and in colorless mode (Z only) - 256 samples per clock
• Integrated support for up to six monitors via HDMI 1.4a and DisplayPort 1.2

Radeon HD 7990 Graphics Specifications

• Core frequency: 1000 MHz
• Number of universal processors: 4096
• Number of texture units: 2x128, blending units: 2x32
• Effective memory frequency: 6000 MHz (4×1500 MHz)
• Memory type: GDDR5
• Memory capacity: 2x3 gigabytes
• Memory bandwidth: 2x288 gigabytes per second
• Theoretical maximum fill rate: 64 gigapixels per second.
• Theoretical texture sampling rate: 256 gigatexels per second.
• One CrossFire connector
• PCI Express 3.0 bus
• Connectors: DVI Dual Link, four Mini-DisplayPort 1.2
• Power consumption up to 375 W
• Two 8-pin auxiliary power connectors
• Dual slot design
• The recommended price for Russia is 32,999 rubles. (for the USA - $999).

Already in the second generation of AMD video cards, the naming principle for dual-chip models remains unchanged. The top solution based on two most powerful video chips differs from the corresponding class of the previous generation model by the first digit in the index: instead of 6, it received the number 7, indicating a new series. The announced video card differs from the single-chip solution by the third digit, indicating the maximum performance within the generation.

As for the comparison with competitors, for the Radeon HD 7990 model announced today, the main rival is the GeForce GTX 690 video card, released almost a year ago, and it is these two-chip solutions that will have to fight each other. True, NVIDIA also has one more powerful solution, but already based on a single GPU - GeForce GTX Titan, which can also be considered a competitor to the AMD board in question.

The new dual-chip Radeon video card is equipped with GDDR5 memory with a capacity of 3 gigabytes per GPU, which is due to the 384-bit memory bus of the Tahiti chips. Such a volume is quite justified for a product of such a high level, since in some modern gaming applications with maximum settings, enabled anti-aliasing and high resolutions, a smaller amount of memory (2 gigabytes per chip or less) may no longer be enough. And even more so this applies to rendering in stereo mode or on multiple monitors in Eyefinity mode.

It is clear that such a powerful dual-chip graphics card has a massive dual-slot cooling system that is different from traditional AMD card coolers. It features a massive heatsink hidden under a shroud with three large fans running at relatively low speeds. The power consumption of the dual-GPU card is quite high for obvious reasons, and it has two 8-pin power connectors, but at least it's not three, as was the case with non-reference samples based on two Tahiti chips.

Architecture

Since the video card codenamed “Malta” is based on two “Tahiti” GPUs from the Southern Islands family, you can simply refer to, which thoroughly describes all the features of the current Graphics Core Next (GCN) architecture. In the base materials, we repeat only the most important characteristics and features of specific products.

The basic block of the architecture is the GCN block, from which all the GPUs of the series are assembled. The computational unit is divided into subsections, each of which works on its own stream of commands, it has a dedicated local storage for data, a first-level cache memory with the ability to read and write, and a full-fledged texture pipeline with sampling and filtering units. Each of the GCN blocks is capable of scheduling and distributing commands on its own, and one computing block can execute several independent command streams. The Radeon HD 7990 uses two Tahiti chips already known to us:

The graphics processor diagram (there are two of them in the Radeon HD 7990) shows 32 computing units of the GCN architecture, and all of them are active. Previously, it was assumed that for a two-chip solution, some of them would have to be turned off, and even lowered the frequency in order to enter the power consumption of 375 W, but AMD engineers managed to successfully solve this difficult task. Perhaps a special new revision of the Tahiti with lower power consumption has been released, or the chips just pass a very strict selection.

Since each GCN unit has 16 texture units, the number of TMUs is 128 units per chip, which gives a total performance of 256 gigatexels per second, which is very good for a GeForce GTX 690 competitor. The number of ROP units and memory controllers in HD 7990 also did not change compared to its single-chip counterpart, they were left in the amount of 32 and 6 pieces per GPU, respectively. The Radeon HD 7990 has a dual 384-bit memory bus made up of twelve 64-bit channels for a total memory bandwidth of 576 GB/s, another record.

Otherwise, the new board supports everything modern technologies by AMD, which were introduced and improved in the new video chips of the Radeon HD 7000 line: PowerTune, ZeroCore, Eyefinity 2.0, HD3D, Steady Video, improved quality of texture filtering, etc. All this is described in detail above in the description of the Radeon HD 7970, and there is simply no point in repeating it.

Cooling system and power consumption

In the case of such serious dual-chip boards, a highly efficient cooling system becomes especially important. If, in the case of solutions from partners based on two Tahiti, three-slot solutions were used, and in the case of ASUS ARES II, even water cooling, in this case it was necessary to get by with less power, so a cooler was designed that has a very massive heatsink and three fans with improved acoustic characteristics.

The noise of the cooling system and the provided temperature for GPUs are one of the most important consumer characteristics for any video card, including the top solution designed for enthusiasts. Too loud or inefficient cooling system will be regarded by buyers as a less profitable purchase, other things (roughly) being equal. So AMD took this issue very seriously with the Radeon HD 7990 when compared to other top solutions on the market. Consider the acoustic characteristics of the new system:

The diagram shows the noise level from three different video cards: the Radeon HD 7990 and two competitors: dual-chip GeForce GTX 690 and single-chip GTX Titan from NVIDIA. Moreover, the noise was measured in different conditions - in idle mode (System Idle) and at maximum load using Furmark. If AMD's numbers are to be believed, even the single-chip Titan falls short of their novelty in terms of cooler noise, not to mention the dual-chip GTX 690, which is the loudest in this comparison.

But isn't it to the detriment of the GPU temperature that such impressive acoustic performance was achieved? The following chart shows GPU temperatures measured on AMD's Radeon HD 7990 and the same two competitors. This time, AMD used only the high load mode when testing at Furmark.

And again, a “cunning” coordinate axis is used with a non-zero origin. The real difference between 80 and 82 degrees for the Radeon HD 7990 and GTX Titan will be virtually imperceptible, although the 87 degrees for the GTX 690 clearly stand out for the worse. Again, we note that all of these tests were carried out by the interested party and are subject to independent verification.

In terms of power consumption, the dual-chip solution is nothing new, but support for the previously announced ZeroCore Power technology is also here. This technology helps to achieve significantly lower power consumption in "deep idle" (or "sleep") mode with the display device turned off. In this mode, the idle GPU is almost completely disabled, and consumes less than 5% of the power of the full mode, turning off most of the functional blocks. And in the case of a two-chip board, it is even more important that in the CrossFire system, when drawing a two-dimensional interface of the operating system, all GPUs, except for the main one, do not work at all. That is, in the case of the Radeon HD 7990, one of the chips in 2D mode will be immersed in a deep sleep with minimal power consumption, and the second one can “fall asleep” in the PC deep idle mode.

AMD Radeon HD 7850M- graphics card with DirectX 11 support, based on the GCN architecture. The card is designed for medium and large laptops. It is one of the most productive video cards of 2012. In its manufacture, a 28 nm process technology is used.

Radeon 7800M series adapters are built on the Cape Verde chip, as well as Radeon HD 7770 for desktop computers, with 640 1D shader cores and 40 texture units. But in terms of clock speed, the HD 7850M is inferior to the HD 7870M (675 MHz to 800 MHz, respectively). Overall, the 7850M is about on par with HD 7750 for desktops.

As expected, the Radeon HD 7850M has become a competitor to NVIDIA's GeForce GTX 560M, since, in addition to higher performance, it also shows lower power consumption (like the GeForce GTX 660M). Thus, modern demanding games run freely on medium and high settings.

The 7850M features a new UVD3 video decoder for decoding MPEG-4 AVC/H.264, VC-1, MPEG-2, Flash, Multi-View Codec (MVC) and MPEG-4 part 2 (DivX, xVid) HD video formats .

The 7800M series cards also support automatic switching between integrated and discrete graphics. This feature is called Enduro and is similar to NVIDIA's Optimus technology. In addition, the HD 7850M can support up to 6 monitors simultaneously using Eyefinity technology when Enduro is disabled.

The next feature HD 7850M is a ZeroCore feature that automatically reduces power consumption when in "sleep" mode with the display off. PowerTune technology allows the graphics card to overclock as long as the power consumption is within acceptable limits. For example, the frequency may decrease when running FurMark and OCCT, and increase in some games (Planet, Crysis or Resident Evil 5).

The built-in HD audio processor is capable of delivering high quality audio (TrueHD or DTS Master Audio formats) via HDMI and DisplayPort (such as Blu-ray video). In addition, with the help of DDMA technology, it is possible to output sound from several devices simultaneously.

Energy level 7850M much lower than the 7870M due to the lower core clock.

* Specified clock speeds are subject to change by the manufacturer

7800 Series with different specifications. The chip, built on the Graphic Core Next microarchitecture, occupies a space equal to 2.8 billion transistors. Like most cards from Radeon, there is Eyefinity technology here, which allows you to connect up to six monitors at the same time. They can work independently of each other, or they can form one large monitor. It all depends on what settings will be exposed.

Radeon 7850

This AMD 7800 Series graphics card has 800 megahertz of processor frequency. High performance and bandwidth (153 gigabits per second) provides a bus size of 256 bits. The computing system processes data equal to 1.76 teraflops. Computing units are present in the amount of 16 pieces, and texture units - in the amount of 64 pieces. There are two cores for computing processes.

The memory format corresponds to the GDDR5 marking, and support for DirectX version 11 will help speed up interaction with operating system applications. For better optimization operation of the card, it is necessary to follow the driver updates, since only they are able to fully reveal all the capabilities of the graphics processor and provide access to necessary settings. The base drivers, which designate the video card in the system, are included with the card, and the updated version can be viewed on the AMD website.

This AMD Radeon HD 7800 Series graphics processor is equipped with the latest in-built technologies that allow you to enjoy high-quality and smooth images at 60 frames, while the resolution can reach up to 4096 x 2160 pixels. The same applies to the audio stream, which meets all modern requirements, giving out high-quality sound.

Radeon 7870

This AMD Radeon HD 7800 Series graphics card is a powerful successor of the previous card in terms of characteristics. She has a whole gigahertz to work with the graphics processor. Performance for computing operations is much higher than in the previous version - 2.56 teraflops. There are 20 compute units and 80 texture units.

Since this is the flagship of the 7800 series, it surpasses its brother in many respects. Support for tessellation technology has been introduced into the video cards of this manufacturer for a long time, but in this version it is brought to the limit. Now you can enjoy a three-dimensional image, striking in its realism and detail. And improved anti-aliasing will help to achieve a smooth and pleasant picture.

In other parameters, this representative of the AMD Radeon HD 7800 Series is completely identical in terms of characteristics to the previous video card. Both cards are capable of supporting 3D technology in both video and games. It is also possible to connect several cards in order to increase the performance, but this parameter may also depend on the capabilities of the motherboard.

Page 2 of 5

"Southern Islands"

First, a little about AMD's labeling of its latest products. The manufacturer has divided them into three levels according to performance. The code name "Cape Verde" refers to the Radeon HD 7700. The name "Pitcairn" refers to today's Radeon HD 7870 and HD 7850 testers.

• Entry level = Cape Verde = Radeon HD 7700 series;
• Mainstream = Pitcairn = Radeon HD 7800 series;
• High performance products = Tahiti = Radeon HD 7900 series.

That is, at the moment, AMD has covered all market segments with its 28 nm graphics chips. Only the release of a dual-core video card based on Tahiti chips is expected. Tentative name Radeon HD 7990.

AMD Radeon HD 7800 Series Features

The Radeon HD 7800 graphics processor (Pitcairn) has about 2.8 billion transistors and Graphic Core Next microarchitecture. As mentioned above, the Radeon HD 7850 chip (Pitcairn Pro) has 16 computing units, and its maximum TDP is 130 watts. For the Radeon HD 7870 (Pitcairn XT), these figures are 20 and 175, respectively.

The slide below shows the main specifications of the Radeon HD 7850 and HD 7870 video cards.

2GB of GDDR5 memory is already becoming the standard for most mid-range and high-end models. Thanks to 256-bit. bus and a high clock frequency of 1200 MHz (4800 MHz effective), the bandwidth is 154 GB / s. This will have a positive effect on performance in games with high resolution and picture quality.

PCI Express 3 interface

In the second half of 2011, almost all motherboard manufacturers presented their models of motherboards with PCI Express 3rd generation interface. With the release of the Radeon HD 7000 series, video cards with this interface also appeared. PCI Express 3 has twice the bandwidth (32 Gb/s) of the previous generation PCI Express. Compared to PCIe 2, the throughput per lane has been doubled from 500 Mb/s to 1 Gb/s.

Naturally, to take advantage of the new PCIe 3, you need not only a video card and motherboard with this interface, but also support from the processor (not all models from the Ivy Bridge family will support PCIe 3).

Eyefinity 2.0

AMD has gone further with their Eyefinity technology, which is designed to display images across multiple monitors. Thanks to the high computing power HD 7000 series and Eyefinity 2.0 support, you can now display an image on multiple monitors with a total resolution of 16000 x 16000. This allows you to display the image on 5 displays with a resolution of 2560x1600 set in landscape orientation. To work with such resolutions, a record 3 GB GDDR5 (HD 7970 and HD 7950) is installed on the older models of the family.

AT AMD drivers Catalyst will support custom resolution starting in February. That is, you can set the required resolution depending on the configuration of the displays in Eyefinity. As of Catalyst 12.2, there is an option to set the Start menu to a display that is convenient for you, instead of the far left as it used to be. In addition, Eyefinity 2 supports stereo HD3D output. It supports the combination of three monitors that work in 3D mode.

Improved tessellation

Video cards AMD families The Radeon HD 7000 features the ninth generation tessellator and has received a significant performance boost in geometry processing in modern games. The GCN core still includes two Graphics Engines, but if they previously contained blocks for tessellation and rasterization, now they consist of an arbitrary number of pipelines designed to process geometry and pixels.

AMD Radeon HD 7800 graphics cards support HDMI interface 1.4a, which allows you to display a picture of 120 Hz (60 Hz for each eye), which allows you to display a 3D image. With earlier versions of HDMI, this was not possible. Starting in December, AMD enabled HD3D and Eyefinity to work together in their drivers.

DirectX 11.1

Video cards of the Radeon 7000 family will support the upcoming DirectX 11.1. What this will give in practice is too early to say, since DX 11.1 will be released along with Windows 8. The main advantages of the new API are indicated as follows:

• Independent rasterization;
• Flexible combination of graphics computing and video processing;
• Native Stereo 3D support.

AMD Unified Video Decoder

It is a hardware part of AMD GPUs responsible for video stream decoding. In the Radeon 7000 series, UVF received some improvements. In general, UVD retained all the features of its predecessors, namely support for H.264/AVCHD, MPEG-2, MPEG-4/DivX, VC-1/WMV profile D, Multi-View Codec (MVC), Video Codec Engine ( VCE), AMD Steady Video 2.0. Added support for Dual Stream HD+HD format.

10 Mar. 2016

On this page below there are links to download the latest free AMD graphics card drivers from the Radeon HD 7800 Series, which is part of the Radeon HD Series family. The installation files are taken from the official site and are suitable for: Windows 7, 10, 8, 8.1, XP, Vista 32/64-bit (x86/x64).

For the convenience of choosing the right files, the version of your Windows and its bit depth (“bit depth”) are listed below.

Your computer runs on:

1. Download (153.5 MB / version 16.8.2 (Crimson Edition 16.8.2 Hotfix) / release date 08/12/2016)

   For Windows 7 32-bit

2. Download (239.8 MB / version 16.8.2 (Crimson Edition 16.8.2 Hotfix) / release date 08/12/2016)

   For Windows 7 64-bit

3. Download (134.8 MB / version 16.8.2 (Crimson Edition 16.8.2 Hotfix) / release date 08/12/2016)

   For Windows 10 32-bit

4. Download (208.24 MB / version 16.8.2 (Crimson Edition 16.8.2 Hotfix) / release date 08/12/2016)

   For Windows 10 64-bit

5. Download (205 MB / version 14.4 (Catalyst Software Suite) / release date 04/25/2014)

   For Windows 8 32-bit

6. Download (260 MB / version 14.4 (Catalyst Software Suite) / release date 04/25/2014)

   For Windows 8 64-bit

7. Download (154.21 MB / version 16.8.2 (Crimson Edition 16.8.2 Hotfix) / release date 08/12/2016)

   For Windows 8.1 32-bit

8. Download (239.88 MB / version 16.8.2 (Crimson Edition 16.8.2 Hotfix) / release date 08/12/2016)

   For Windows 8.1 64-bit

9. Download (179 MB / version 14.4 (Catalyst Software Suite) / release date 04/25/2014)

   For Windows XP 32 and 64-bit

10. Download (151 MB / version 13.12 (Catalyst Software Suite) / release date 12/18/2013)

    For Windows Vista 32-bit

11. Download (209 MB / version 13.12 (Catalyst Software Suite) / release date 12/18/2013)

    For Windows Vista 64-bit

Fallback - Obtain Drivers Using AMD Driver Autodetect

This option convenient because the program AMD Driver Autodetect will automatically select and download the latest working drivers, which are suitable for your AMD graphics card and for your version of Windows. The program does not need to be installed, it was created by AMD and the files are downloaded from their official servers.

Instruction:

1. Launch AMD Driver Autodetect and it will immediately automatically select the necessary files for installing drivers.
2. To download files, click on the "Download Now" button.
3. Wait for the files to download and start the installation.