Design PhilosophiesThe design team essentially was able to take a dog and make it a competent part in a pretty short amount of time considering when work likely started. While they did not add any huge, new advances (except for PCI-E 2.0 compatibility and DX 10.1 support) they did produce a solid part which helped turn around the graphics division at AMD. The HD 3870 did not break any records, but it did provide a much needed boost for AMD and people were far more willing to adopt AMD technology as their graphics choice.
While this is a marketing slide from AMD, it is a good example of what I am talking about here. The GT200 chip at 65 nm is huge as compared to the RV770. I don’t necessarily agree with the “efficiency” con of the NV chip, but the other 3 cons are certainly true.
Approximately nine months after the introduction of the HD 3800 series we see AMD stake another claim at being the midrange and upper-midrange champion with the release of the HD 4800 series. The RV770 chip that powers these products is still based on the R600 architecture, but significantly improved in every way. There are now 800 stream units, 40 texture units, beefed up RBEs, and a lot of internal optimization. All of this was done with only a slightly larger die size from the RV670. The anti-aliasing hit was addressed by work done on the RBEs and their ability to throughput double the amount of anti-aliased pixels per clock, and we finally have a part that is competitive with the high end products that NVIDIA was offering at that time with the GTX 260/280 parts. AMD also controls costs by keeping the die relatively small and using a 256 bit memory bus. To offset potential bandwidth limitations AMD relies on GDDR-5 to give bandwidth numbers close to what NVIDIA offers with their 512 bit wide memory bus with GDDR-3. The big splash at the time was made primarily by the HD 4870’s price tag, which was right around $300. It again compared quite favorably to NVIDIA’s GTX 260, which was $150+ more expensive.
For a 35% increase in die size from the RV670 (192 mm square) to the RV770 (260 mm square), we see a tremendous performance increase in both shader power and pure pixel throughput (both AA and AF applied). Going from 320 stream units to a whopping 800 is pretty eye opening, especially considering that we get more than double the stream units for a relative paltry 35% increase in die size! And at 260 mm square the RV770 is still less than one-half the size of the original GT200 chip at 65 nm (576 mm square), and still about 145 mm square smaller than the 55 nm GT200b (approx 405 mm square).
We can see the progression from the RV670 to the RV770. For not a huge amount of space used, we get a pretty significant increase in performance and overall usability.
Now we are approaching the release of the RV790 and AMD has of course been very silent about it, and the only rumors we have heard about it up until this week are that it would be merely a speed optimized RV770 chip that would consume about the same amount of power and produce the same amount of heat as the older HD 4870 cards. Looking back at AMD’s design history, we can guess that this is probably so far from the truth that it was intentionally planted by AMD to potentially mislead NVIDIA into thinking that they were going to wait for 40 nm before introducing a new high end part based on a new or improved architecture.
My Guesses and Reasoning
Current estimates point to both AMD and NVIDIA releasing a more midrange part on TSMC’s newly opened 40 nm process. Initial reports peg this process as being somewhat problematic. While the die shrink aspects are nice, apparently TSMC did little to improve the underlying materials used in their 40 nm process. Leakage is a major issue because of the lack of improvements in these materials combined with the finer geometries (stuff being closer together). In comparison companies like Intel and AMD have invested a lot of time and effort into the materials so as to improve switching speed and decrease leakage. AMD uses SOI, both AMD and Intel use Low-k di-electrics, both use compressive and expansive strain technologies, and Intel has also introduced High-k/metal gates. All of these technologies act to decrease leakage and improve transistor switching speed/power. From all indications TSMC has not changed the materials from 55 nm to 40 nm in any significant way. In fact, performance could be slightly worse overall for the 40 nm process as compared to 55 nm, even though the die savings may outweigh some of these performance considerations.
The card that helped to fully recover AMD’s image as a serious graphics producer.
While both AMD and NVIDIA will produce a smaller, more midrange part at 40 nm, they have both been apparently waiting until things improve to release a new high end part at 40 nm. Estimates have NVIDIA pegged at a late August/early September release for a new, high end 40 nm part. Estimates for AMD fall in the same area. So this leaves Spring with the GTX 285/295 release as the new high end for NVIDIA, and most are expecting the speed optimized RV790 to be the new high end for AMD.