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: pgerassi who wrote (74152)	3/10/2002 1:44:33 PM
From: milo_morai	Read Replies (1) \| Respond to of 275872

<font color=blue>Dual Gigabyte MB watch.impress.co.jp

To: pgerassi who wrote (74152)	3/10/2002 4:24:04 PM
From: Ali Chen	Read Replies (1) \| Respond to of 275872

Dear Pete, "A simple MB swap or BIOS change could alter.." I appreciate you taking my projections under the scrutiny. Let me try to answer your concerns: 1. "simple MB /BIOS change" The data I was using were submitted 2 days apart: spec.org and spec.org Although tweaking BIOS settings is indeed a possibility, it is very unlikely that there were any configuration changes in a conservative organization like Dell. Both submission use the same base system Dell 340, with the same set of optimizations, with the same 010922Z revision of 5.0.1 compiler, with the same hardware and version of OS. 2. "Small reporting errors in either sample could make huge differences in the projection." I prefer scientific approach and follow Hennessy&Patterson formula that "the real measure of computer performance is time". Therefore I have habits to use all reported times from SPEC reports (.asc). The 3 runs of the benchmark contain 36 results, each one typically varies +- 1 second in random direction. The total execution time runs up to 7000 - 8000 seconds. Even if we take the worst case and sum up all errors, the resulting error will be in the range of +-0.5%. In fact, since we sum random errors, the result has to be smaller by a factor of SQRT(36)=6. 3. "And you are assuming linearity rather than a curve in that projection." Your assumption about my assumption is incorrect. Judging from your problems with estimators, you are looking into <Score>-<frequency> trends. This is really a classic error commonly made;-) Things are much simpler in <task completion time>-<clock period> parametric space. If you still prefer frequency-score plots, try the following "estimator": score = 1/(a/f+b). This function is pretty inconvenient to approximate by plain low-order polynomials from Excel ;-) 4. "This is a classic error that is commonly made.... ..That is the other classical error." No one is perfect. I do this kind of stuff for a living for the last 30 years, including last 6 years in the computer industry, and I do not do "classical errors", only non-trivial ones ;-) ;-) ;-) 5. Therefore I stand behind my projection that, assuming the same hardware platform, same binaries, same OS, same microarchitecture, NW at 3000 will have the SPEC2000-int score gain of 32+-1% from NW-2000. After all, it was a theoretical question about typical application speedups as a function of CPU core frequency, which implies identical platform/memory architecture, and I simply objected the excessive generalization of a particular arbitrary-devised performance formula with only 20% scaling while SPECint makes 64% and SPECfp still makes 40%. - Ali