Will the AMD RV790 Architecture be Something Different?
A Bit of History
The other day Fudzilla reported that the RV790 was far more than a faster clocked version of the current RV770 which powers the Radeon 4800 series. Initial reports were that the RV790 was to be a slightly optimized RV770 which would run around 150 to 200 MHz faster in core clock speed. Nothing terribly exciting to see here. Or so we were led to believe.
Fudzilla’s claims are not out of this world, especially if we take a look at AMD’s direction when it comes to graphics for the past two years. The R600 (HD 2900 XT) was the last generation of product from AMD which would be “over-the-top”. Since that time AMD has concentrated on making their products smaller, more efficient, and a bit more scalable.
Recovering From the R600
During development of the R600, the decision was made to stop going for “top spot” with a large, complex die which would be expensive to produce and only exist as a $400+ product. Considering how well the original Radeon HD 2900 XT did, we can see that the stakes for this top end game are high, and the price of failure (or coming in second best) is steep. The next iteration of the R600 architecture was far more successful.
Going with a R600 style of chip has proven to be risky, and if not done correctly can be quite costly for a company and their reputation.
Seven months after the overhyped release of the HD 2900 series of cards, AMD unleashed the Radeon HD 3800 series to great acclaim. The performance of the high end HD 3870 was not all that different from the older HD 2900 XT, and it still took a large hit when anti-aliasing was enabled, but the combination of a much lower price, lower heat production, and FAR lower power consumption made this a much more competitive part to the then current offerings of NVIDIA. Because of the lower heat and power consumption of these new parts, we saw the rise of the X2 cards from AMD and their overall performance allowed them to more adequately compete with NVIDIA at the top end. Also consider that AMD was able to achieve comparable performance by using a 256 bit bus rather than the more expensive 512 bit bus.
The decision to go with far less complex chips which perform in the range of 80% to 90% of much larger chips, and then being able to easily double the chip count on a single card to address the high end market was a wise decision. In fact, if we look a little further back we see that NVIDIA embraced that philosophy (whether it was intentional or not) during the 90 nm GeForce 7000 series. Remembering back then, the 7900 GTX was a smaller and less complex part than the competing Radeon X1800/X1900 products from ATI. It did not perform as well as the X1900 in most situations (especially shader heavy applications) but it pulled less power, ran cooler, and was less expensive to produce. Then NVIDIA abandoned that idea and went with the giant G80 chip at the high end. Luckily for NV they had no competition and the 8800 GTX was the fastest single video card for about 1.5 years.
Going with smaller, simpler designs has allowed AMD to be much quicker and effective in countering pretty much all of NVIDIA's moves as of late.
Going back to the RV670 (HD 3800 series), we see that it was essentially a shrunk R600. It contained the same number of stream units, RBEs, and texture units as the older R600, but it was a slightly optimized design that was produced one full node shrink below what the R600 was. The R600 was produced on a high speed 80 nm process from TSMC, while the RV670 was produced on TSMC’s then brand-new 55 nm process. The die size was 50% smaller, and the advantages of the smaller process and its new material technologies allowed it to clock at speeds near the 800 MHz of the original HD 2900 XT without requiring 225 watts+ of power. The 3870 in fact rarely went over 120 watts at full throttle.