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: dougSF30 who wrote (198073)	5/19/2006 11:59:55 PM
From: pgerassi	Respond to of 275872

Doug: If the IPC gains come from prefetching, then if they were "smart", the gains disappear when it has to stop. But, that is just when the load gets the heaviest and is the reason server customers bought into virtualization in the first place. Replace a bunch of lightly loaded single purpose servers with a few heavily loaded ones. I never said that the designers were dumb. Although PHBs can force engineers to make bad and/or dumb decisions work. Williamette is a good example of many bad dumb decisions being forced by marketing on engineers. Double clocked ALUs (a gee whiz feature that sounds good), narrow decoders (die too big), trace cache (needed by ->), ultra long pipelines (want highest clocks, performance be damned), barrel shifter removed (die too big) and sharing the multiplier between ALU and FPU (die too big). It made the new core perform poorly except in special cases. In many programs at the time, P4 was slower than the older P3 (that happened again with Prescott). A lot of times engineers think up ideas that look good on paper, but when tried, fail for a lot of reasons. The problem is that PHBs (marketing) has already sold customers on that idea that didn't pan out and force the engineers to make it work. Being behind makes this happen more often as management is grasping at straws. As to the numbers, Intel, Newisys (now Saminia), AMD, HP (and Compaq/DEC), Sun, IBM among others were the sources. But even when I give them to you, you ignore them and respout your distorted views backed up with irrelevant information or simply made up facts. Your use of Linpack (long discreditted) as a server benchmark is a case in point. The only time Linpack is used is for the supercomputer list and then that one is made such that larger scores require larger matricies to be used. The length of each side of the matrix, n, is much larger than the one indicated by your scores. That might be the theoretical speed. It also could be the peak score which always occurs when the matrix fits in cache, but the number looks more like the theoretical score. For an example of Linpack sizes: "LINPACK performance of 20-nodes (48000x48000, without Strassen method) was 144.1GFLOPS (90% of peak)" n=48K for 144GFLOPs, thus 38GFlops should use a matrix of 12700x12700. The matrix seen is either theoretical (FPU throughput) or peak (somewhere about 1000x1000, well within the L2). Also see here: spymac.com Because linpack is a really bad, 30 year old benchmark. And PPC970 is practically designed to get high linpack scores. (AFAIK linpack is scheduled to be replaced for the top500 ranking system next year) Pete