AMD is launching two products based on Cayman. The Radeon 6970 is a 250W part that runs at 880MHz, with a full 24 SIMDs and a 5.5GT/s GDDR5 memory interface. The 6950 uses the same die, but is binned out at 800MHz, with 22 SIMDs and a 5GT/s memory interface; it also runs slightly cooler with a 200W TDP. Both cards have an idle power consumption of 20W and use 2GB of GDDR5. The 6970 and 6950 will be priced at roughly $369 and $299 (online), squarely in the enthusiast segment. They are positioned to replace the 5870 and 5850 in the high-end of the market. This leaves Nvidia’s GTX 580 as the highest performance GPU, while the highest performance graphics card is still the dual-GPU Radeon 5970.
For graphics performance data, we rely on Scott Wasson’s excellent Cayman review at the Tech Report. Competitively, both Cayman products are pitted against versions of Nvidia’s Fermi. The Radeon 6970 lines up with Nvidia’s GeForce GTX 570. The overall performance for the two cards is closely matched, with a slight edge (0-5%) for the Radeon. Frankly, the results are close enough to be a wash, with the performance advantage largely determined by the nature of the graphics workload. Despite the close performance, the 6970 runs at roughly 15% higher power than the 570. Ironically, the Radeon 6950 fares much better overall and compares very favorably to the GTX 470. Not only does the 6950 have a clear 10-20% performance advantage, it also has a 10% lower TDP – the best of both worlds. While not a factor initially, driver maturity will influence the relative performance over time. Cayman is the first of a new microarchitecture, while Fermi has been around for 9 months. As a result, Cayman should see larger relative gains as the drivers are tweaked, but this is very difficult to quantify.
One novelty with Cayman is that AMD has started to include ‘typical’ power numbers for their GPUs. The goal is to address what AMD feels is the widening gap between power consumption on real workloads, and product design specifications. The typical power numbers for the two cards are 190W and 140W respectively; roughly 70-75% of the TDP. It would be quite interesting to see how these numbers vary for different architectures and a wide range of workloads (e.g. shader vs. ROP vs. texturing limited games).
While pricing may change over time, the costs for TSMC’s 40nm process should stay relatively constant. The main determinant of GPU cost is the silicon die area and for graphics cards, the quantity of GDDR5 required is also relevant. Both Cayman GPUs are 389mm2, while the original Fermi (GF100, for the 4xx series) is ~550mm2 and the GF110 (5xx series) is ~529mm2. Additionally, AMD has slightly lower board costs because they rely on 8 channels of GDDR5, rather than 12 for the GTX 570 and 10 for the GTX 470.
AMD has slightly departed from the sweet spot strategy with ever larger chips, but they still have a leg up with a significant cost advantage for Cayman. AMD could use this cost advantage to aggressively price their products and take market share from Nvidia, opt for higher margins or a combination of the two. Judging by AMD’s disclosed pricing, the 6970 seems to be aimed at achieving higher margins, rather than taking marketshare. It is priced on-par or slightly above the GTX 570, despite very close performance. The 6950 is much more aggressively priced given the competition and stands a good chance of stealing marketshare around the $300 level. How this plays out in the market really depends on Nvidia though. They could react by reducing prices to hold onto volume or stick to the established prices.