Will the AMD RV790 Architecture be Something Different?

Wrappin it Up

    I think that the changes outlined above would nicely cover the increase in die size, along with other fine level optimizations that AMD would include.  But the next question that raises its head is why 55 nm?  Well, with the obvious problems that TSMC’s 40 nm process is currently experiencing, it makes little sense to port a new, complex design to the process when potential yield and heat/power issues could raise their ugly heads.  TSMC’s 55 nm process has been in full production for AMD for around 20+ months now.  The yield curves are well known, the speed bins for a given design are also consistent at this time, and because the process is mature the cost per mm square of die size has gone down over the lifetime of the line.  Given that a 350 mm square die probably costs about the same now as the introductory RV670 die was 18 months ago at 190 mm square, it makes little sense for AMD to gamble on going with a 40 nm product which could have teething issues not unlike TSMC’s 130 nm process that caused so many problems for the introduction of NVIDIA’s GeForce FX series.


    In the future AMD will of course go to 40 nm, but not until it makes sense to do so.  If TSMC suddenly improves its power figures for the 40 nm process, then AMD has the ability to do a dumb shrink of the RV790 and reap the benefits.  But in terms of getting a product out which could conceivably dethrone the fastest single chip card currently out (the GTX 285), plus increase AMD’s visibility in the GPGPU world, then this would be a masterstroke for AMD if they can actually get out a product as I have described above.  Not only would they have a 6 month lead on NVIDIA in getting out a new high end card, but they can do it in a more cost effective manner than what NVIDIA is stuck with in the GT200b part (larger die size, more complex PCB with 512 bit memory bus, reliance on GDDR-3).  If AMD is able to keep the power and heat envelopes in the same general area as the current RV670, then the dual chip cards could be true performance monsters in their space, both in terms of gaming and high performance computing.  3.4 TFLOPs of single precision and 1.6 TFLOPs of double precision in one card sucking up 250 watts of power is almost mind boggling.  Almost.  This is of course assuming that AMD can in fact clock this new chip in the 900 MHz range.

    Given AMD’s renewed focus after the R600 debacle, this speculation does not seem that far out of place of what could actually happen.  With my luck AMD will release a 1200 stream unit chip without any of the GPGPU or efficiency improvements, and it will simply be clocked faster as well.  While it would still make for a fast card, it would be far less interesting than what I have described above.  With OpenCL gaining traction, DirectX 11 having support for general compute, and that particular marketplace opening up and becoming far more interested in GPGPU, I think it would be foolish for AMD to further ignore architectural advances in their chip that would greatly improve their performance in that area.


    So while many people deride Fudzilla for their reporting, I believe they are correct in this case.  The RV790 does appear to be far more than initial rumors have suggested.  Whichever way this product goes, we can assume that it will again be released at a $300 pricepoint for the single GPU card, and a dual chip implementation will be released at a later date in the $450 to $500 range.  If it does turn out as I have described, I hope NVIDIA is ready for more price cuts.

If you are interested in passing along your thoughts on this article, please join the discussion here.