Prime95 version 28.9 / 28.10 - Page 2

Prime95 · 2016-03-30, 23:43

Prime95 uses SSE2 code for AMD chips. I don't believe you can force prime95 to use FMA FFTs because I'm not sure that AMD supports the Intel FMA3 syntax. I'd have to build a special executable with the AMD FFT4 syntax.

ewmayer · 2016-03-31, 00:02

Quote:

Originally Posted by Prime95

Prime95 uses SSE2 code for AMD chips. I don't believe you can force prime95 to use FMA FFTs because I'm not sure that AMD supports the Intel FMA3 syntax. I'd have to build a special executable with the AMD FFT4 syntax.

According to Wikipedia:

Quote:

AMD introduced FMA3 support in processors starting with Piledriver architecture for compatibility reasons.[2][3] The 2nd generation APU processors based on "Trinity" (32nm) supporting FMA3 instructions were launched May 15, 2012. The 2nd generation Bulldozer processors with Piledriver cores supporting FMA3 instructions were launched October 23, 2012.

So to flip my SSE2-mode query around, is there a way to force Prime95 to run in AVX (both sans and with FMA3) mode on AMD CPUs? It seems to me it is high time to revisit the SSE2-versus-AVX performance issue on their more recent CPUs ... what is the last AMD CPU on which you did comparative timings?

Prime95 · 2016-03-31, 01:46

Try putting "CpuArchitecture=5" in local.txt file.

The last time I investigated AMD performance was Bulldozer.

kladner · 2016-03-31, 03:14

Quote:

Originally Posted by ewmayer

A note and a couple of questions re. your results summary:

o AMD's 8-int-core/4-fpu hybrid arch will likely act like an 8-core for integer-dominated work which is not too memory-intensive, for FPU-and-memory-heavy stuff like Prime95 it's effectively a 4-core.

o Does "FMA" here mean Intel-AVX2-style FMA3, or the AMD-only FMA4 which AMD introduced in late 2011 with their Bulldozer core?

o Has AMD fixed the issue noted by George in which their early AVX offerings performed worse running in AVX mode than they did in SSE2 mode? George, is there any way to force the program to run in SSE2 mode on a platform supporting both?

Indeed, it is quick enough (if P95 and 2 x mfaktc are not running.) However, compared to contemporaneous Intel chips, it is a power-hungry dog. (No offense intended to real, honest dogs.)
I can't answer in detail, but this chip is a Piledriver, the successor to Bulldozer. (Argh! These names!)
https://en.wikipedia.org/wiki/Piledr...rchitecture%29

Fred · 2016-03-31, 13:18

So far, testing v28.9 in comparison to v28.7, using 4096 benchmarks I'm noticing the following:

Crappy old i3 laptop - ~10% drop in performance when I go to 28.9
i5-4670K, DDR3-1600 : Just about dead even, no change in performance
i5-6500, Non-Z OC DDR4-2133 : ~4% increase in performance when I go to 28.9
(interesting sidenote, with 28.7 on my newer systems the benchmarks indicated I should run 4 workers for maximum throughput. with 28.9, I seem to get the best throughput with 4 cores on 1 worker)

So it seems the faster processors or faster memory benefits from the changes, while (at least in my testing) older systems take a hit.

Mark Rose · 2016-03-31, 15:16

The 10% drop in performance is interesting to me. I wonder what caused that.

Prime95 · 2016-03-31, 15:39

Quote:

Originally Posted by Mark Rose

The 10% drop in performance is interesting to me. I wonder what caused that.

There are some strange cache effects going on that I don't understand. The difference between 28.7 and 28.9 is I chose different sizes for pass 1 and pass 2 of the FFT. This changes the working set size for each pass.

I presume the crappy i3 was pre-FMA3. I optimized based on my 4-core Sandy Bridge, the crappy i3 must have much different cache and memory characteristics. At least Fred knows about the issue and can revert to 28.7 -- other users are not as fortunate.

Mark Rose · 2016-03-31, 15:47

Quote:

Originally Posted by Prime95

There are some strange cache effects going on that I don't understand. The difference between 28.7 and 28.9 is I chose different sizes for pass 1 and pass 2 of the FFT. This changes the working set size for each pass.

I presume the crappy i3 was pre-FMA3. I optimized based on my 4-core Sandy Bridge, the crappy i3 must have much different cache and memory characteristics. At least Fred knows about the issue and can revert to 28.7 -- other users are not as fortunate.

The chip he mentioned was an i3-3120M. It has 2 cores, 1 MB of L2, and 3 MB of L3.

tului · 2016-03-31, 18:01

Pentium G4400, two real cores, no HT W10-64. It tells me that I've assigned more workers than cores despite that not being the case. I ignored it and will check performance.

science_man_88 · 2016-03-31, 19:03

my processor is even older I just don't have old results to compare it to http://ark.intel.com/products/53428/...Cache-3_40-GHz based on things I did on spu-Z at one point it's on a foxconn board with 8 gb single channel memory.

ewmayer · 2016-03-31, 22:22

Quote:

Originally Posted by kladner

Indeed, it is quick enough (if P95 and 2 x mfaktc are not running.) However, compared to contemporaneous Intel chips, it is a power-hungry dog. (No offense intended to real, honest dogs.)
I can't answer in detail, but this chip is a Piledriver, the successor to Bulldozer. (Argh! These names!)
https://en.wikipedia.org/wiki/Piledr...rchitecture%29

So did you try the timings with AVX-usage enabled as George noted? I was asking on your behalf ... no AMD in my CPU collection.

2016-03-30, 23:43	#12
Prime95 P90 years forever! Aug 2002 Yeehaw, FL 16267₈ Posts	Prime95 uses SSE2 code for AMD chips. I don't believe you can force prime95 to use FMA FFTs because I'm not sure that AMD supports the Intel FMA3 syntax. I'd have to build a special executable with the AMD FFT4 syntax. Last fiddled with by Prime95 on 2016-03-30 at 23:44

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Prime95 version 27.3	Prime95	Software	148	2012-03-18 19:24
Prime95 version 26.3	Prime95	Software	76	2010-12-11 00:11
Prime95 version 25.5	Prime95	PrimeNet	369	2008-02-26 05:21
Prime95 version 25.4	Prime95	PrimeNet	143	2007-09-24 21:01
When the next prime95 version ?	pacionet	Software	74	2006-12-07 20:30

2016-03-31, 01:46	#14
Prime95 P90 years forever! Aug 2002 Yeehaw, FL 1CB7₁₆ Posts	Try putting "CpuArchitecture=5" in local.txt file. The last time I investigated AMD performance was Bulldozer.

2016-03-31, 13:18	#16
Fred "Ron" Jan 2016 Fitchburg, MA 1100001₂ Posts	So far, testing v28.9 in comparison to v28.7, using 4096 benchmarks I'm noticing the following: Crappy old i3 laptop - ~10% drop in performance when I go to 28.9 i5-4670K, DDR3-1600 : Just about dead even, no change in performance i5-6500, Non-Z OC DDR4-2133 : ~4% increase in performance when I go to 28.9 (interesting sidenote, with 28.7 on my newer systems the benchmarks indicated I should run 4 workers for maximum throughput. with 28.9, I seem to get the best throughput with 4 cores on 1 worker) So it seems the faster processors or faster memory benefits from the changes, while (at least in my testing) older systems take a hit.

2016-03-31, 15:16	#17
Mark Rose "/X\(‘-‘)/X\" Jan 2013 29×101 Posts	The 10% drop in performance is interesting to me. I wonder what caused that.

2016-03-31, 18:01	#20
tului Jan 2013 2²×17 Posts	Pentium G4400, two real cores, no HT W10-64. It tells me that I've assigned more workers than cores despite that not being the case. I ignored it and will check performance.

2016-03-31, 19:03	#21
science_man_88 "Forget I exist" Jul 2009 Dumbassville 10000011000000₂ Posts	my processor is even older I just don't have old results to compare it to http://ark.intel.com/products/53428/...Cache-3_40-GHz based on things I did on spu-Z at one point it's on a foxconn board with 8 gb single channel memory.