Author: Robert Hyatt
Date: 07:04:21 08/13/00
Go up one level in this thread
On August 13, 2000 at 04:28:26, David Blackman wrote: >On August 12, 2000 at 23:37:00, Robert Hyatt wrote: > > >>I haven't seen _any_ latency improvements since 1995. Don't get fooled by >>SDRAM and the like streaming memory into the L1 cache. Write a program >>that addresses words that are scattered randomly over memory. Then that >>fancy streaming/buffering is worthless and you get a feel for true memory >>latency. And you'll see that 100us number is pretty close to the truth. > >The speed-ups i have seen were measured numbers, not advertising numbers. And i >certainly know there is a big difference. My measurements were genuine latency >for random addressing of large arrays, using pointer chasing to make sure >nothing got hidden in the pipeline. 300 to 1000 ns in 1995. 100 to 200 now. (and >200 is getting rarer). There were probably a few things back in 1995 that could >do 100 but they weren't cheap or easy to find. Note that you have to start testing probably at the first pentiums, so that at least the bus is the same width. However, I have some numbers at the office for some memory benchmarks, and they are pretty level, unless you do the copy- type operations that can use the streaming reads of EDO/SDRAM/etc to ramp up the bandwidth. The right way to test is to access a word in the bottom half of memory, then a word in the top half, and bounce back and forth for a while, making sure that you have a reasonably large stride between successive accesses to the same half. That will cook the machine and show how bad memory latency really is. You can run the same test on the alpha, and although the memory copy is many times faster (more overall bandwidth) latency doesn't look much different. I can run the same test for you on a Cray if you want.. but I already know what the latency is there as I fool with them regularly. Still takes 100nx to get a word there as well, although with the huge level of interleaving they do, it has memory bandwidth untouched by any other machine I know of.
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.