Theory and architecture

It's finally the time for another update of GPU series. As for now, there is only one product to discuss. Today AMD has announced new solutions for the top price range. The company had no cards to oppose High-End products from NVIDIA for some time, but now the situation has changed to the opposite (at least temporarily). And it has modified the pricing policy of the company.

Remember our previous articles about the fastest single-GPU solutions from AMD? The company described its market policy as follows: it was going to manufacture single-GPU graphics cards for Mid-End and Lower High-End segments, and offer only dual-GPU solutions for the top segment. But when the company had an excellent opportunity to raise prices for single-GPU solutions in the absence of new graphics cards from NVIDIA, this strategy was temporarily adjusted, and now the company has mastered the new price range of $399 for a top single-GPU card. Where all these buzzwords about affordable solutions for common users? What about the economic recession?

The company will also launch the dual-GPU solution in due time, of course. The market does not need it right now. But the time will come, when NVIDIA rolls out its new graphics cards, and AMD drops prices for the current single-GPU cards. As for now, AMD makes money for being the first in the market to offer... no, not DirectX 11 compatibility (it will come in handy only the next year). The highest performance level of the new single-GPU cards. Only prev-gen dual-GPU solutions can compete with them.

However, AMD hasn't gone far from the approach in its RV670 and RV770. They just overhauled the Mid-End GPU for the price range of $200-300 into a larger GPU with a tad higher power consumption for $300-400. It's generally an adequate solution, if a dual-GPU card does not require lowering GPU parameters to fit in the required maximum power consumption.

Such a powerful GPU has become possible owing to the finally streamlined 40nm process technology of TSMC. RV740 was the first experimental GPU, but it wasn't wide spread in retail stores, and those few products that made it came with inflated price tags. Let's hope that Cypress will have a better fate.

Fortunately, we've got a GPU with the modified architecture (you will learn how modified below). The theoretical part promises to be very interesting due to architectural changes in the GPU, even though it's a descendant of R6xx/R6xx/R7xx. Detailed information about the previous unified architectures, exemplified by older solutions, can be found in the following articles:

• ATI RADEON HD 4890
• ATI RADEON HD 4870 512MB
• ATI RADEON 4850 512MB
• ATI RADEON HD 2400XT/2600PRO/2600XT (RV610/630)
• ATI RADEON HD 2900 XT (R600)

We assume that you are already familiar with the previous architectures from AMD. Now we'll examine the specifications of two new RADEON HD 5800 graphics cards based on the new Cypress GPU.

RADEON HD 5800 graphics cards

• Codename: Cypress
• process technology: 40nm
• 2.15 billion transistors (over twice as many as in RV770)
• Unified architecture with an array of common processors for streaming processing of various data types: vertices, pixels, etc.
• Hardware support for DirectX 11, including new Shader Model 5.0
• 256-bit memory bus: four 64-bit controllers supporting GDDR5
• Core clock: 725-850MHz
• 20 SIMD cores, including 1600 scalar floating-point ALUs (integer and floating-point formats, support for FP32 and FP64 in compliance with IEEE 754)
• 20 enlarged texture units supporting FP16 and FP32 formats
• 80 texture address units and just as many bilinear filtering units that can filter FP16 textures at full speed, trilinear and anisotropic filtering for all texture formats
• 32 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 32 samples per cycle (including MSAA 2x/4x and FP16 buffers), 128 samples per cycle in Z only mode
• Writing results up to eight frame buffers simultaneously (MRT)
• Integrated support for RAMDAC, six Single Link or three Dual Link DVIs, HDMI, as well as HDMI and DisplayPort

RADEON HD 5870 specifications

• Core clock: 850MHz
• Unified processors: 1600
• 80 texture units, 32 blending units
• Effective memory frequency: 4800MHz (4*1200MHz)
• Memory type: GDDR5
• Memory: 1024MB
• Memory bandwidth: 153.6GB/sec
• Maximum theoretical fillrate: 27.2 gigapixel per second
• Theoretical texture sampling rate: 68.0 gigatexel per second
• Two CrossFireX connectors
• PCI Express 2.0 x16
• 2 x DVI-I Dual Link, HDMI, DisplayPort
• Power consumption: from 27W to 188W (two 6-pin power connectors)
• Two-slot design
• Recommended price: $399

RADEON HD 5850 specifications

• Core clock: 725MHz
• Unified processors: 1440
• 72 texture units, 32 blending units
• Effective memory frequency: 4000MHz (4*1000MHz)
• Memory type: GDDR5
• Memory: 1024MB
• Memory bandwidth: 128GB/sec
• Maximum theoretical fillrate: 23.2 gigapixel per second
• Theoretical texture sampling rate: 52.2 gigatexel per second.
• Two CrossFireX connectors
• PCI Express 2.0 x16
• 2 x DVI-I Dual Link, HDMI, DisplayPort
• Power consumption: from 27W to 170W (two 6-pin power connectors)
• Two-slot design
• Recommended price: $299

So, AMD uses the most advanced process technology (40nm) to roll out solutions that enjoy apparent advantages over 55nm GPUs. The die surface hasn't grown much, but the frequency potential has become higher, and almost all characteristics have been doubled versus RV770. The new GPU is certainly much more power-efficient, but these cards consume practically just as much power as the top solutions from the previous generation (HD 4870 and HD 4890). And they offer much higher performance.

We are happy to say that the naming principle remains unchanged (you know who's constantly changing that). Only the first number has changed in comparison with the previous series. However, these cards have slightly different positioning. The HD 4870 and 4850 were designed to replace the HD 3870 and 3850 as soon as they appeared in stores, but the price of HD 5870 indicates that AMD is conquering new price ranges -- unusually high for single-GPU solutions from this company. And the HD 5850 is to replace the HD 4890, the fastest card from the previous generation. The latter will not be discontinued, but it will go down to the lower price range.

As is always the case with AMD, two modifications differ in GPU/memory clock rates. And it's not the only difference between these models, although both cards are now equipped with the same memory type -- GDDR5, as GDDR3 memory bandwidth would have been too small even for the HD 5850. A more important difference between the HD 5850 and the HD 5870 is in the number of ALUs and TMUs -- their number is decreased by 11% in the lower model. It was probably done not to reject GPUs with defected ALUs/TMUs. Besides, the low solution consumes less power (but still much) and has a shorter PCB. However, the HD 5850 still comes with a two-slot reference cooler.

Both cards have the same video memory size in their reference modifications -- one gigabyte. That's the right decision, as modern games (even multiplatform titles) have high requirements to video memory. Memory size of 1024MB is the optimal volume now -- there is no need in more memory, and 512MB is already insufficient for many applications.

Write a comment below. No registration needed!