P3 data needed - Page 2

Dresdenboy · 2003-09-04, 05:34

Someone should try to run it on a Barton. I have a friend who could do this if required.

Xyzzy's measurement of mprime with and w/o SSE2 showed very little differences. I assume that there is some unidentified bottleneck when it runs on Opteron.

BTW there are new Sisoft Sandra results which indicate a significant difference in SSE2 performance between Opteron revision B and C.

(http://www.amdmb.com/article-display...&PageID=11
and more info at http://amdzone.com/forums/messagevie...threadid=69071)

Sandra names it "whetstones" (which usually includes trigonometric functions and other complex stuff) but it seems (from the resulting number of MFLOPS) that it's only a tight loop of simple FP operations (else it couldn't use SSE2).

And here the resulting IPC between 2 Opteron revisions:

revision B:
x87: ~1.5
SSE2: ~1.6

revision C:
x87: ~1.5
SSE2: ~2.0 (!)

What I still down't know is how this affects other SSE2 code. But it is very possible that the Sandra benchmark just doesn't take care of any optimization hints for pre revision C Opterons - which could be the case for a lot of software. There could for example be a penalty when the lower half of a SSE2 reg is used in scalar ops while the upper half is in an unknown state (not zeroed or loaded with something). I didn't dig deeply into this stuff.

Unfortunately the bought 140 is rev. B. But I will try to get a revision C. At least the amdmb-review showed that the new revision is migrating into lower model ranges.

QuintLeo · 2003-09-04, 20:36

OS stuff was probably getting into the cache at least part of the time - especially in a Windows environment.

battlemaxx · 2003-10-02, 04:10

Are you still needing to run tests?

I have a dual 733/133 and dual 800/100 and a single 733/133 w/ 100 memory I could probably donate and spin up for a while.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Corrupted data?	Oops!	Information & Answers	2	2013-10-22 03:48
GPU TF vs DC/LL data	bcp19	GPU to 72	0	2011-12-02 16:41
Data available?	Prime95	LMH > 100M	10	2007-06-22 23:55
New "Data" forum might be needed	GP2	Data	8	2003-09-13 02:34
Data needed... Please Help!	guido72	Lounge	21	2003-04-04 15:48

2003-09-04, 05:34	#12
Dresdenboy Apr 2003 Berlin, Germany 169₁₆ Posts	Someone should try to run it on a Barton. I have a friend who could do this if required. Xyzzy's measurement of mprime with and w/o SSE2 showed very little differences. I assume that there is some unidentified bottleneck when it runs on Opteron. BTW there are new Sisoft Sandra results which indicate a significant difference in SSE2 performance between Opteron revision B and C. (http://www.amdmb.com/article-display...&PageID=11 and more info at http://amdzone.com/forums/messagevie...threadid=69071) Sandra names it "whetstones" (which usually includes trigonometric functions and other complex stuff) but it seems (from the resulting number of MFLOPS) that it's only a tight loop of simple FP operations (else it couldn't use SSE2). And here the resulting IPC between 2 Opteron revisions: revision B: x87: ~1.5 SSE2: ~1.6 revision C: x87: ~1.5 SSE2: ~2.0 (!) What I still down't know is how this affects other SSE2 code. But it is very possible that the Sandra benchmark just doesn't take care of any optimization hints for pre revision C Opterons - which could be the case for a lot of software. There could for example be a penalty when the lower half of a SSE2 reg is used in scalar ops while the upper half is in an unknown state (not zeroed or loaded with something). I didn't dig deeply into this stuff. Unfortunately the bought 140 is rev. B. But I will try to get a revision C. At least the amdmb-review showed that the new revision is migrating into lower model ranges.

2003-09-04, 20:36	#13
QuintLeo Oct 2002 Lost in the hills of Iowa 2⁶×7 Posts	OS stuff was probably getting into the cache at least part of the time - especially in a Windows environment.

2003-10-02, 04:10	#14
battlemaxx Nov 2002 43 Posts	Are you still needing to run tests? I have a dual 733/133 and dual 800/100 and a single 733/133 w/ 100 memory I could probably donate and spin up for a while.