Date: Wed, 9 Apr 1997 15:43:48 -0500 (CDT) From: Tony Kimball <Anthony.Kimball@East.Sun.COM> To: dave@persprog.com Cc: hardware@freebsd.org Subject: Re: Pentuim or Pentuim Pro ? Message-ID: <199704092043.PAA02495@compound.east.sun.com> References: <199704091535.KAA01745@compound.east.sun.com> <199704091755.LAA08240@Ilsa.StevesCafe.com> <199704091826.NAA02162@compound.east.sun.com> <334BF670.5B7E@persprog.com>
next in thread | previous in thread | raw e-mail | index | archive | help
Quoth Dave Alderman on Wed, 9 April: : : Let's not forget that SDRAM may show better performance with the : upcoming faster processors as well, although if I know this industry, : the existing SDRAM will be inadequate in some way or another. It's tied to the bus speed. 66 MHz SDRAM will not reliably deliver a read per clock on an 83 or 100 MHz bus. You can buy 100 MHz SDRAM and put it on a 66 MHz bus and move it to a faster board in the future, but it is a violation of the JIT purchasing methodology (the large 100 MHz cost premium will drop in the future), and may not make sense if you resell the old board anyhow. : That 64Meg limitation can be a real killer for loaded machines and even : for certain mathematical tasks like simulation or modeling. Intel : really wants you to buy the Pentium Pro or Pentium II and this is one of : their "incentives". : : What is "substantially better"? A ballpark figure would be fine. That's the best I can give, since I've never compared systems with similar processor/bus/chipset differing only in RAM architecture. I'm guessing 12-15% win (66MHz SDRAM vs. 60ns EDO, which is approximate cost parity) for my application, a direct ODE solver which is essentially (in terms of architectural performance model) a large number of iterations of a very poorly written 24MB bcopy. Probably not a very *typical* (meaning "REAL WORLD"?) application. I could *strongly* benefit from PPro (primarily because of superscalarization), but today's cost-benefit trade-off favors the smaller incremental benefit of K6-166/SDRAM. If my application were not embarassingly parallel, this would no longer be the case -- for example, when I convert to an indirect solver. (Actually, the real price-performance win in the near future is Alpha. When the AlphaPC boards plateau -- in a month or two, perhaps? -- I'm going to have to seriously consider switching to, ugh, Linux.)
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199704092043.PAA02495>