[Rumor] First performance figures Interlagos?

[Musho]

Is this perhaps the first real Interlagos benchmark? From this post and on there are all the details!

F@H Benchmarks Interlagos (Without(!) Turbo core 2.0):
ubuntu 10.10 server x64
512G DDR3-1333
P6901
Average time/frame: 00:03:52

[09:15:07] Completed 0 out of 250000 steps (0%)
[09:18:59] Completed 2500 out of 250000 steps (1%)
[09:22:42] Completed 5000 out of 250000 steps (2%)
[09:26:16] Completed 7500 out of 250000 steps (3%)
[09:30:08] Completed 10000 out of 250000 steps (4%)
[09:34:06] Completed 12500 out of 250000 steps (5%)


For comparison:
Bulldozer "Interlagos" 16x4@ 1.8GHz* = 00:03:52
Opteron "Magny Cours" 12x4@ 2.2GHz = 00:06:40
*Look at the thread

Per thread, Bulldozer seems to be 58% faster than a K10 core at equal clock speeds in this F@H benchmark. JF-AMD already mentioned scaling wouldn't be linear with a whole module loaded vs just 1 out of 2 cores in a module loaded, so single threaded performance should be even _better_ than 58% higher IPC than K10 if you aren't loading the other core in a module! Looking good so far!


Calculations:

Bulldozer: 232 seconds with 16 cores at 1,8ghz
232*16*1,8=6681,6


Opteron: 400 seconds with 12 cores at 2,2ghz
400*12*2,2=10560

10560/6681,6=1,58045977 ~ 1,58 higher IPC


Source: http://gathering.tweakers.net/forum/...40992#35640992

[spicypixel]

Am I to assume 500mhz turbo on this sku?

[RPGWiZaRD]

Sounds too good to be true. Anything over 30% single thread boost would be a positive suprise to me. I still think there's some dodgy things about these results. Hope not though ofc. :p JF-AMD where are you? :p

[Musho]

Am I to assume 500mhz turbo on this sku?

No idea. As posted in the first post, Turbo Boost 2.0 was disabled. This means it was running at a constant 1,8ghz on all cores throughout the test. This is a good thing, though. This means we can accurately determine the IPC increase Bulldozer has over K10.

[FlanK3r]

whau, hard to say, if is it true or not

[madcho]

i think avx is a serious boost, and folding use a lot of floating calcs. So it's normal the fp gain seem huge.

How much is i7 ?

[Particle]

Do current builds of FAH already take advantage of AVX instructions?

[xVeinx]

No, the version of gromacs used by F@H does not use AVX.

[Musho]

Do current builds of FAH already take advantage of AVX instructions?

AFAIK, they don't. Don't quote me on it, though

[Stukov]

Is this an 8 module vs a 12 core?

[xBanzai89]

Is this an 8 module vs a 12 core?

Looks it "Bulldozer "Interlagos" 16x4." 16 Cores = 8 modules

[Stukov]

Looks it "Bulldozer "Interlagos" 16x4." 16 Cores = 8 modules

I wasn't under the impression that 8 modules = 2 "cores" per module, more like 1.25-1.5 cores per module?

[madcho]

No, the version of gromacs used by F@H does not use AVX.

avx don't improve so much perf, but the new fpu is really improved, so sse code should be way much faster.

i said avx, i was thinking about the 2x128bits amazing fpu

[red]

http://www.bit-tech.net/hardware/cpu...x5650-review/9

Here's previous Magny Cours folding@home performance relative to others.

[sutyi]

So after doing some math, a BD core is about ~52% faster C2C compared to a K10.5 one. Hmm.

[~CS~]

Total BS.

[Stukov]

Total BS.

Such a strong counter-point. +1 internet points for you.

[xVeinx]

In order for a true comparison, the exact same unit needs to be run the separate machines, as small differences in PPD between units of the same type (ie, 6701, etc.) can occur. Either way, it would potentially be a nice speedup. Also realize though that Gromacs is a VERY optimized software, so the speedup seen there would only be seen in HPC type applications, not as much in day-to-day applications.

[LightSpeed]

I wasn't under the impression that 8 modules = 2 "cores" per module, more like 1.25-1.5 cores per module?

More like ~1.8x

[Hornet331]

In order for a true comparison, the exact same unit needs to be run the separate machines, as small differences in PPD between units of the same type (ie, 6701, etc.) can occur. Either way, it would potentially be a nice speedup. Also realize though that Gromacs is a VERY optimized software, so the speedup seen there would only be seen in HPC type applications, not as much in day-to-day applications.

Indeed, its very hard to replicate things in F@H and if there are so many different variables it quite pointless to brag about it, it just gives a general direction of where performance is. But so far its seems very promising.

[informal]

This was already posted...

[xBanzai89]

I wasn't under the impression that 8 modules = 2 "cores" per module, more like 1.25-1.5 cores per module?

If I am not mistaken JF-AMD said the module has TWO real functioning cores. Its not like 2 .75 cores.

[Manicdan]

If I am not mistaken JF-AMD said the module has TWO real functioning cores. Its not like 2 .75 cores.

1 module is ~1.8 cores, ONLY if IPC was identical between generations, which is not true.

[JF-AMD]

If I am not mistaken JF-AMD said the module has TWO real functioning cores. Its not like 2 .75 cores.

Each module has 2 cores

1 module is ~1.8 cores, ONLY if IPC was identical between generations, which is not true.

Each module has 2 cores.

[Stukov]

Each module has 2 cores




Each module has 2 cores.

I was under the impression from what AMD has said is that the OS system see's them as "cores", but weren't real cores other than they had an integer core...in addition AMD wanted to get away from the "core" terminology with the debut BD. So is that not true then?