I mentioned yesterday that I was getting some erratic findings. Let me give you some idea
as to what they were.
First, choice of OS has a huge bearing on performance. Running a TBird at 1200Mhz (A7V133, 9X133, SDRAM fast memory settings,
1024X768 resolution) gets you a Photoshop SysMark2000 score of 135 in Win98. Same machine, same
everything else except running in Win2K gets you a score of about 200, or about a 50% improvement.
Second, the boost going to MP varies quite a bit between OSs, too. Going from a TBird to an MP gets you about
a 24% improvement in Win98. Doing the same in Win2K gets you more in the neighborhood of 40% improvement.
John R. Heritage brought up the point that the MP has a few other enhancements besides SSE, which might affect the score.
He pointed me to a program H. Oda wrote a while back which turns SSE on and off for the MP. You can get it here.
Unfortunately, it just gives you a BSOD in Win2K, but it works in Win98. WCPUID verifies that SSE is on or off.
I ran the Photoshop section of SysMark with it turned on and off, and with it turned off, I got the same kind of scores I got with the TBird.
So it’s SSE-enabling that’s increasing the scores, at least in Win98.
Whether it’s 24% or 40%, that’s quite a bit of an improvement. Does SSE really help Photoshop that much, or were the filters just cherry-picked.
I’ve done some preliminary testing using PS5Bench, which was a series of tests designed by someone who wanted to compare Macs to PCs.
While there are elements within that test which show considerable performance improvement, most don’t.
Does that mean we can go hang Bapco now? Not quite yet. Not even Intel chips showed a whole lot of improvement using PS5Bench. PS5Bench may not
be representative of what SSE can do with filters.
PS5Bench is also erratic. I’ve gotten variations of 50% on some tests. Additionally, how it uses Photoshop’s timer is very suspect.
Photoshop has a built-in timer for actions. If I don’t use PS5Bench, but something else, it takes the action, gives me a time, and that’s that.
With PS5Bench, though, it gives me an initial time, then another one a moment later, which adds as much as a half-second to the total. This is pretty significant if you’re measuring 3 second actions.
So what I’m going to do is use the PS5Bench test file, but not the program, and run all the filters that come with the program with and without SSE.
The advantage of doing it that way is that we’ll get definitive results on what SSE really does with Photoshop filters.
If we find that half the filters benefit a ton from SSE, that’s one thing. If we find out only 5% benefit a lot, and those were the only ones Bapco used, then you can hang them. 🙂
Do you see how and why this is a better approach?