Theory and architecture

Several days ago we wrote about GeForce GTX 200 - fresh High-End solutions from NVIDIA. And now it's time for new products from AMD. But we'll start with a graphics card, which is not intended to compete with GTX 200.

AMD plans to roll out dual-GPU cards based on the high-frequency RV770 to compete with the most expensive single-GPU card from NVIDIA - GTX 280. As our readers may already know, AMD abandoned its old strategy of launching special GPU versions for the high-end segment. Instead, Upper High-End will be covered by multi-GPU solutions (two GPUs on a single card in CrossFire mode). AMD notes the following problems of big high-end GPUs: too high costs and power consumption for most users, long delays before new fabrication processes are adopted (less complex GPUs are easier to upgrade to thinner processes), etc.

Advantages of the new (multi-GPU) strategy from AMD: optimal solutions for all price ranges (using cards with different number of GPUs), low manufacturing costs, and fast adoption of new technologies. Multi-GPU configurations have their cons, of course. Firstly, the most frequently used method of multi-GPU rendering: Alternate Frame Rendering (AFR). Even if it provides the highest performance gain in framerate, it does not make rendering smoother like the growth of single-GPU FPS. As a result, multi-GPU systems often demonstrate high FPS accompanied by control lags and lack of smoothness.

However, we shall not find out who's the king of the hill until we test Radeon HD 4870 X2 vs. GeForce GTX 280. And now we have a potential leader in another utilitarian segment of the market from $200 to $300. Such graphics cards are powerful enough for modern games in high resolutions and maximum graphics quality settings, but they are not too expensive, they do not consume very much power, they are not very big, and they do not get very hot. Has AMD coped with the task of creating an ideal card for the biggest segment of the market? We'll try to find it out in this article.

First of all, we are pleased to announce a sterling architectural overhaul. That is the theoretical part promises to be very interesting, as these architectural changes are very important, and we've all been looking forward to them. Before you read this article, you should study the baseline theoretical articles - DX Current, DX Next, and Longhorn. They describe various aspects of modern graphics cards and architectural peculiarities of older products from NVIDIA and AMD.

• Longhorn: DirectX 10 accelerators and shaders
• DirectX.Update: Half a step forward for graphics accelerators
• DX.Next: The near and nearest future of hardware graphics acceleration

These articles predicted the current situation with GPU architectures, many assumptions about future solutions were confirmed. Detailed information about the unified architecture of AMD R6xx solutions (including RV670) can be found in the following articles:

• AMD/ATI DirectX 10 series: RADEON HD 2900 XT (R600)
• A detailed analysis of the ATI RADEON HD 2400XT/2600PRO/2600XT (RV610/630)
• ATI RADEON 3850/3870 (RV670): 320 shader processors and 256-bit memory bus

So, let's examine detailed characteristics of two new RADEON HD 4800 graphics cards, based on the new RV770 GPU.

RADEON HD 4800

• Codename: RV770
• Fabrication process: 55 nm
• 956 million transistors
• Unified architecture with an array of common processors for streaming processing of vertices and pixels, as well as other data
• Hardware support for DirectX 10.1, including new Shader Model 4.1, geometry generation, and stream output
• 256-bit memory bus: four 64-bit controllers supporting GDDR3/GDDR5
• Core clock: 625-750 MHz
• 10 SIMD cores, including 800 scalar floating-point ALUs (integer and floating-point formats, support for FP32 and FP64 in compliance with IEEE 754)
• 10 enlarged texture units supporting FP16 and FP32 formats
• 40 texture address units
• 160 texture fetch units
• 40 bilinear filtering units that can filter FP16 textures at full speed, trilinear and anisotropic filtering for all texture formats
• Dynamic branching in pixel and vertex shaders
• 16 ROPs supporting antialiasing with programmable sample patterns (over 16 samples per pixel), including FP16 or FP32 formats of the frame buffer. Peak performance is up to 16 samples per cycle (including MSAA 2x/4x and FP16 buffers), 64 samples per cycle in Z only mode
• Writing results up to eight frame buffers simultaneously (MRT)
• Integrated support for two RAMDACs, two Dual Link DVIs, HDMI, HDTV, DisplayPort

RADEON HD 4870 Specifications

• Core clock: 750 MHz
• Unified processors: 800
• 40 texture units, 16 blending units
• Effective memory frequency: 3600 MHz (4x900 MHz)
• Memory type: GDDR5
• Memory: 512 MB
• Memory bandwidth: 115 GB/sec
• Maximum theoretical fillrate: 12.0 gigapixel per second
• Theoretical texture sampling rate: 30.0 gigatexel per second
• 2 x CrossFireX connectors
• PCI Express 2.0 x16 bus
• 2 x DVI-I Dual Link, 2560x1600 video output
• TV-Out, HDTV-Out, support for HDCP, HDMI, DisplayPort
• Power consumption: up to 160 W (two 6-pin connectors)
• Two-slot design
• Recommended price: $299

RADEON HD 4850 Specifications

• Core clock: 625 MHz
• Unified processors: 800
• 40 texture units, 16 blending units
• Effective memory frequency: 2000 MHz (2x1000 MHz)
• Memory type: GDDR3
• Memory: 512 MB
• Memory bandwidth: 64 GB/sec
• Maximum theoretical fillrate: 10.0 gigapixel per second
• Theoretical texture sampling rate: 25.0 gigatexel per second
• 2 x CrossFireX connectors
• PCI Express 2.0 x16 bus
• 2 x DVI-I Dual Link, 2560x1600 video output
• TV-Out, HDTV-Out, support for HDCP, HDMI, DisplayPort
• Power consumption: up to 110 W (one 6-pin connector)
• Single-slot design
• Recommended price: $199

So, AMD continues to roll out 55nm graphics solutions. Compared to the 65nm fabrication process used for top solutions from NVIDIA, the more advanced process gives certain advantages: smaller core surface, higher frequency potential, higher yield of effective GPUs able to operate at high frequencies, and probably lower production costs. Compared to RV670, the new GPU is even more power efficient. It does not consume much energy, but it offers much higher performance than previous solutions.

As you can see, designations of different solutions remains the same (as in the previous series), only the first digit has changed. In other respects, positioning of the HD 4870 and 4850 is similar to that of the HD 3870 and 3850 at the time of their launch. The HD 3000 family is not discontinued, of course, these cards just go to the lower price ranges.

As usual, AMD rolls out two modifications in a series, which differ in clock rates of the GPU and memory. But this is not the only difference between these models - the top card is equipped with GDDR5 memory, which provides larger bandwidth - it affects performance in complex modes with antialiasing in the first place. What concerns other differences, we can mention higher power consumption and dual-slot cooling system of the more expensive card.

Both reference cards are equipped with 512 MB of memory. Modern games have high memory requirements, but they usually use up to 500-600 MB, higher memory volumes are necessary only on rare occasions. That's why 512 MB of video memory is an optimal memory size for this moment. However, it would have been nice, if some of AMD partners offered RADEON HD 4870 cards with 1 GB of memory. It's important for the most hard-driving games, like Crysis.

Write a comment below. No registration needed!