As you know, Radeon HD 6970 failed to become the fastest graphics card, taking the second place after GeForce GTX 580. So it was only natural that AMD didn't rest until it rolled out another solution to challenge the current champion. So, according to the company's tradition to make top-end products with two GPUs, AMD rolled out Radeon HD 6990, essentially featuring two Radeon HD 6970.

AMD Radeon HD 6990 specifications

• Codenamed "Antilles"
• 40nm process technology
• 2 GPUs, 2.64 billion transistors in each
• 2 dies, each measuring 389mm²
• Unified architecture with an array of stream processors for handling vertices, pixels, etc.
• Hardware DirectX 11 support, including Shader Model 5.0
• Dual 256-bit bus, 2 x 4 64-bit controllers, GDDR5 support
• 830-880 MHz core clock (see explanations below)
• 2 x 24 SIMD cores, 3072 stream processors, supporting FP32 and FP64 precision of the IEEE 754 standard
• 2 x 96 texture units, as many bilinear filtering units, supporting FP16 texture filtering at full speed as well as all trilinear and anisotropic filtering modes for all texture formats
• 2x32 ROP units, supporting antialiasing with over 16 samples per pixel, FP16/FP32 frame buffers included
• A RAMDAC for each GPU, supporting 6 Single-Link DVI or 3 Dual-Link DVI ports, HDMI 1.4a, DisplayPort 1.2

AMD Radeon HD 6990 (HD 6990 OC) graphics card specifications

• 830(880) MHz core clock
• 3072 stream processors
• 2 x 96 texture units, 2 x 32 blending units
• 4 x 1250 MHz (5000 MHz) memory clock
• 2 x 2GB of GDDR5 memory
• 2 x 160GB/s memory bandwidth
• 53(56) Gpixel/s theoretical peak fill rate
• 159(169) Gtexel/s theoretical peak texture fetch performance
• CrossFireX connector
• PCI Express 2.1
• Dual-Link DVI, 4 x mini DisplayPort 1.2
• Consumes from 37W to 375W(450W)
• Typically consumes in games up to 350W(415W)
• 2 x 8-pin supplementary power connectors
• Dual-slot design
• MSRP in the U.S. $699

As you know, AMD changed product naming policy in this generation of products: HD 5870 and HD 5850 were replaced by HD 6800 and HD 6900. And the latter got the fastest GPU, too. So it's quite logical that a new dual-GPU solution is a part of the HD 6900 series as well. But since the "6970" index was already taken, the today's novelty got "6990".

As for comparing HD 6990 with rivals, it has none as yet. But we've been expecting a similarly dual-GPU product from NVIDIA for a long time, so perhaps they'll roll out one this spring. GPU makers don't have much choice these days, all they can do with the 40nm process technology is release dual-GPU solutions based on good old processors.

Moving on. AMD Radeon HD 6990 features up to 2 x 2GB of GDDR5 memory, which is justified for a top-end product. All the more so because some games run at maxed out settings need more than 1GB per GPU. This is even truer for stereo modes or high-resolution Eyefinity configurations.

As you would expect, the novelty has a dual-slot cooler, quite long and protected by a plastic cover. Power consumption is high for obvious reasons, so the card has two 8-pin supplementary power connectors previously unseen in reference cards (although seen in certain custom products). Also note that AMD has recently started providing typical consumption in games aside from the maximum one.

Architecture

AMD engineers were to create an efficient graphics and computing architecture with advanced GPGPU capabilities, geometry units working in parallel and improved texture filtering and fullscreen antialiasing.

The Cayman architecture became an intermediate between the previous Cypress architecture and the unreleased 32nm one which didn't make it to the market. But the new GPU still has some features from the latter. Compared with Cypress, extra transistors were spent on new computing and graphical capabilities, some of which you can see on the flowchart below.

The two graphics engines are the most important here. Each includes a rasterizer, a tesselator and other geometry units. Also note two dispatch processors. The two geometry units can now process two primities per clock, doubling transformation and culling performance compared with Cypress (even tripling it in some cases thanks to improved bufferization).

Another very important architectural novelty is the VLIW4 superscalar architecture (vs. the older VLIW5). Each stream processor has 4 ALUs instead of 5. This improves stream processor efficiency, although it also reduces potential peak performance.

Write a comment below. No registration needed!