|From: Poul-Henning Kamp <firstname.lastname@example.org>
|Subject: Re: Adaptec 2940?
|> > The drivers are identical, so the interupt time should be the same for
|> > driving either card on the same machine. Your benchmark is not really
|> > valid since they were run on different motherboards.
|> To some extent. It is instesting that a good EISA system can best a
|> poor PCI system. Woe to those buying cheap PCI motherboards.
|Oh, it's much worse than that. This is some numbers I have been collecting,
|the represent the bandwidth between CPU and RAM pretty well:
|A RS6000/590, AIX 3.2.5
| # cc -o ram-speed -O3 ram-speed.c
| # ./ram-speed
| 49005fb0 0.116 uS/op 8.61e+06 op/S 32.833 Mb/S
| 8938c0df 0.177 uS/op 5.64e+06 op/S 21.532 Mb/S
I don't know what to make of this. The 590 is capable of 4*8=32 bytes
transferred from main memory per 62.5 MHZ clock. A simple matrix-matrix
multiply will attain this. That's a hellava lot more than 32 MB/s.
This system has had the best memory bandwidth of any non Cray Research
system for the past year.
There is a something like a "-power2" optimization switch that makes a
lot of difference.
Another memory benchmark that has been run on a lot of systems is STREAM,
available from ftp://perelandra.cms.udel.edu/bench/stream. Results
My P5-100-TP4-PB/FreeBSD system is in there, looks pretty good against lower
end SGIs and SUNs, but doesn't come close to the RS6000/590.