Silicon Investor (SI) -- The First Internet Community

STOCKTALK

We've detected that you're using an ad content blocking browser plug-in or feature. Ads provide a critical source of revenue to the continued operation of Silicon Investor. We ask that you disable ad blocking while on Silicon Investor in the best interests of our community. If you are not using an ad blocker but are still receiving this message, make sure your browser's tracking protection is set to the 'standard' level.

Technology Stocks : Advanced Micro Devices - Moderated (AMD) -- Ignore unavailable to you. Want to Upgrade?

To: graphicsguru who wrote (255212)	8/5/2008 5:49:08 PM
From: pgerassi	Read Replies (3) \| Respond to of 275872

Graphicsguru: Going from 320 cores to 800 cores is less a problem in scaling (2.5x) than going from 8 to 32 (4x). The biggest hit in scaling is from 1 to 2. Many programs fail at that initial stage. By the time they (programs that use all cores fairly effectively) are at 8 cores, most of the scaling problems are fairly well understood. That graph that was presented to show nearly linear scaling between 8 and 32 means that the performance at 8 cores must be very bad. Even at 32 cores, it must be bad. Too bad to even get close to utilize the infrastructure. Most modern high performance GPUs like the 3870 and 4870 scale at 40-50% at the top end. They are usually above the "knee" where scaling begins to drop off rapidly. They expect that during its life, games will grow enough GPU resource usage to pull it back below the "knee". To see a GPU where the infrastructure was overbuilt, just look at the original R600 (Radeon 2900). It had the nearly the same BW that the 4870 does, yet the performance was quite a bit less than the 4870. Heck the 3870 with about half the BW, outperformed it in many ways at far less power. The R600 had a 1024 bit ring bus (512 bit each way) that was scaled down and later changed to a much more successful crossbar switch/hub in the R770. The latter used far less die area and gave more performance. So one of the major improvements claimed by Larrabee, has already been tried and rejected by GPU makers. If Larrabee actually comes out in late 2009/early 2010, it will still be three years behind the R600 and likely three generations behind ATI/nVidia flagship GPUs in performance. Being that bad in performance would allow for nearly linear scaling from 8 to 32 (1x to 4x). Going from utterly unusable to practically unusable isn't much of an improvement. Pete