« Return to Thread: AMD Phenom: phenomenally slow
Re: AMD Phenom: phenomenally slow
by Clint Whaley
::
Rate this Message:
Guys,
>BIOS? Motherboard? RAM? Or do you just need to go ask them for a
>different one?
I'm trying to get some understanding before attempting to isolate such a
problem. I.e, I don't want to try swapping out all these components, only
to find that the Phenom really has a 2/4/1 peak. I was hoping someone
had actually used a Phenom, and if they got 4/8/2, then I would know
something is wrong. The only thing even possible is BIOS, it seems to me:
there's no new instructions that need to be enabled, etc. It is difficult
to see how even BIOS could cause half the FPU to be unused, once the mobo
recognizes the chip at all (as mine does) . . . I get these number with
in-cache timings, so it is not a memory issue . . .
>There's quite a bit of discussion online about "erratum" for the
>Phenom. The fix is just to disable to TLB
>[http://en.wikipedia.org/wiki/Phenom_(processor)] and it causes
>significantly a performance hit for some applications
>[http://techreport.com/articles.x/13741,http://www.legitreviews.com/article/618/1/].
>
>I'm not an expert, but
>http://www.hardcoreware.net/amd-phenom-is-broken-costs-a-14-performance-hit-to-fix/
>contains this gem: "The results were even worse than expected. When
>looking directly at cache memory performance, bandwidth dropped by as
>much as 38.7%, and latency slowed down by over 50%!"
>
>Since BLAS utilizes cache more than the desktop applications I found
>results for, it could be that this is sufficient to explain the
>problems you see.
Nope, the problem occurs on in-cache data. My K10h Opteron had that bug,
and I applied the BIOS fix, and it didn't change my peak numbers, which
are 4/8/2 for Opteron K10h. Supposedly, Phenom with numbers ending in 50
don't have this bug, and mine is a 9750 . . .
>AMD says about "Wide FP Accelerator" for both Phenom X3/X4 and
>Opteron.
>For Opteron it's written directly - about 4 FLOP/cycle. For Phenom I
>didn't find this data explicitly.
This is the same experience I'm having: no confirmation on Phenom FPU peak.
I have a help request out to AMD, but no response so far (only been 1 day).
>Is Phenom a 10h Processor ? (May be 10h is a part of CPUID result ?).
ATLAS & Goto think so :) Here's the snippit from /proc/cpuinfo:
vendor_id : AuthenticAMD
cpu family : 16
model : 2
model name : AMD Phenom(tm) 9750 Quad-Core Processor
stepping : 3
cpu MHz : 2400.955
cache size : 512 KB
physical id : 0
siblings : 4
core id : 3
cpu cores : 4
fpu : yes
fpu_exception : yes
cpuid level : 5
wp : yes
flags : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm 3dnowext 3dnow constant_tsc rep_good pni cx16 popcnt lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs
Notice it has the misalignsse, which AFAIK is a k10h feature . . .
>Atlas (make time) says me about 323.6% (max) on some dgemm on Opteron
>2350/2 GHz, this corresponds to 6472 MFLOPS and can't be 2 DP FLOP per
>cycle it's clear that chip has 8 GFLOPS.
Yeah, my K10h Opteron has 4/8/2 as well, it is my Phenom that is showing
2/4/1 . . .
Thanks,
Clint
**************************************************************************
** R. Clint Whaley, PhD ** Assist Prof, UTSA ** www.cs.utsa.edu/~whaley **
**************************************************************************
-------------------------------------------------------------------------
Sponsored by: SourceForge.net Community Choice Awards: VOTE NOW!
Studies have shown that voting for your favorite open source project,
along with a healthy diet, reduces your potential for chronic lameness
and boredom. Vote Now at http://www.sourceforge.net/community/cca08
_______________________________________________
Math-atlas-devel mailing list
Math-atlas-devel@...
https://lists.sourceforge.net/lists/listinfo/math-atlas-devel
>BIOS? Motherboard? RAM? Or do you just need to go ask them for a
>different one?
I'm trying to get some understanding before attempting to isolate such a
problem. I.e, I don't want to try swapping out all these components, only
to find that the Phenom really has a 2/4/1 peak. I was hoping someone
had actually used a Phenom, and if they got 4/8/2, then I would know
something is wrong. The only thing even possible is BIOS, it seems to me:
there's no new instructions that need to be enabled, etc. It is difficult
to see how even BIOS could cause half the FPU to be unused, once the mobo
recognizes the chip at all (as mine does) . . . I get these number with
in-cache timings, so it is not a memory issue . . .
>There's quite a bit of discussion online about "erratum" for the
>Phenom. The fix is just to disable to TLB
>[http://en.wikipedia.org/wiki/Phenom_(processor)] and it causes
>significantly a performance hit for some applications
>[http://techreport.com/articles.x/13741,http://www.legitreviews.com/article/618/1/].
>
>I'm not an expert, but
>http://www.hardcoreware.net/amd-phenom-is-broken-costs-a-14-performance-hit-to-fix/
>contains this gem: "The results were even worse than expected. When
>looking directly at cache memory performance, bandwidth dropped by as
>much as 38.7%, and latency slowed down by over 50%!"
>
>Since BLAS utilizes cache more than the desktop applications I found
>results for, it could be that this is sufficient to explain the
>problems you see.
Nope, the problem occurs on in-cache data. My K10h Opteron had that bug,
and I applied the BIOS fix, and it didn't change my peak numbers, which
are 4/8/2 for Opteron K10h. Supposedly, Phenom with numbers ending in 50
don't have this bug, and mine is a 9750 . . .
>AMD says about "Wide FP Accelerator" for both Phenom X3/X4 and
>Opteron.
>For Opteron it's written directly - about 4 FLOP/cycle. For Phenom I
>didn't find this data explicitly.
This is the same experience I'm having: no confirmation on Phenom FPU peak.
I have a help request out to AMD, but no response so far (only been 1 day).
>Is Phenom a 10h Processor ? (May be 10h is a part of CPUID result ?).
ATLAS & Goto think so :) Here's the snippit from /proc/cpuinfo:
vendor_id : AuthenticAMD
cpu family : 16
model : 2
model name : AMD Phenom(tm) 9750 Quad-Core Processor
stepping : 3
cpu MHz : 2400.955
cache size : 512 KB
physical id : 0
siblings : 4
core id : 3
cpu cores : 4
fpu : yes
fpu_exception : yes
cpuid level : 5
wp : yes
flags : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm 3dnowext 3dnow constant_tsc rep_good pni cx16 popcnt lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs
Notice it has the misalignsse, which AFAIK is a k10h feature . . .
>Atlas (make time) says me about 323.6% (max) on some dgemm on Opteron
>2350/2 GHz, this corresponds to 6472 MFLOPS and can't be 2 DP FLOP per
>cycle it's clear that chip has 8 GFLOPS.
Yeah, my K10h Opteron has 4/8/2 as well, it is my Phenom that is showing
2/4/1 . . .
Thanks,
Clint
**************************************************************************
** R. Clint Whaley, PhD ** Assist Prof, UTSA ** www.cs.utsa.edu/~whaley **
**************************************************************************
-------------------------------------------------------------------------
Sponsored by: SourceForge.net Community Choice Awards: VOTE NOW!
Studies have shown that voting for your favorite open source project,
along with a healthy diet, reduces your potential for chronic lameness
and boredom. Vote Now at http://www.sourceforge.net/community/cca08
_______________________________________________
Math-atlas-devel mailing list
Math-atlas-devel@...
https://lists.sourceforge.net/lists/listinfo/math-atlas-devel
« Return to Thread: AMD Phenom: phenomenally slow
| Free embeddable forum powered by Nabble | Forum Help |
