[Rumor] First performance figures Interlagos?

Printable View

03-07-2011, 06:45 AM
Musho

[Rumor] First performance figures Interlagos?

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! :up:

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
03-07-2011, 07:17 AM
spicypixel

Am I to assume 500mhz turbo on this sku?
03-07-2011, 07:26 AM
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
03-07-2011, 07:30 AM
Musho

Quote:

Originally Posted by spicypixel

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.
03-07-2011, 07:50 AM
FlanK3r

whau, hard to say, if is it true or not
03-07-2011, 08:01 AM
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 ?
03-07-2011, 08:15 AM
Particle

Do current builds of FAH already take advantage of AVX instructions?
03-07-2011, 08:35 AM
xVeinx

No, the version of gromacs used by F@H does not use AVX.
03-07-2011, 08:36 AM
Musho

Quote:

Originally Posted by Particle

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

AFAIK, they don't. Don't quote me on it, though
03-07-2011, 08:42 AM
Stukov

Is this an 8 module vs a 12 core?
03-07-2011, 08:44 AM
xBanzai89

Quote:

Originally Posted by Stukov

Is this an 8 module vs a 12 core?

Looks it "Bulldozer "Interlagos" 16x4." 16 Cores = 8 modules
03-07-2011, 08:45 AM
Stukov

Quote:

Originally Posted by xBanzai89

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?
03-07-2011, 08:46 AM
madcho

Quote:

Originally Posted by xVeinx

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 :D
03-07-2011, 08:49 AM
red

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

Here's previous Magny Cours folding@home performance relative to others.
03-07-2011, 08:51 AM
sutyi

So after doing some math, a BD core is about ~52% faster C2C compared to a K10.5 one. Hmm.
03-07-2011, 08:53 AM
~CS~

Total BS.
03-07-2011, 08:55 AM
Stukov

Quote:

Originally Posted by ~CS~

Total BS.

Such a strong counter-point. +1 internet points for you.
03-07-2011, 09:08 AM
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.
03-07-2011, 09:21 AM
LightSpeed

Quote:

Originally Posted by Stukov

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
03-07-2011, 09:25 AM
Hornet331

Quote:

Originally Posted by 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.

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.
03-07-2011, 09:33 AM
informal

This was already posted...
03-07-2011, 09:52 AM
xBanzai89

Quote:

Originally Posted by Stukov

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.
03-07-2011, 09:57 AM
Manicdan

Quote:

Originally Posted by xBanzai89

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.
03-07-2011, 10:38 AM
JF-AMD

Quote:

Originally Posted by xBanzai89

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

Quote:

Originally Posted by Manicdan

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

Each module has 2 cores.
03-07-2011, 10:54 AM
Stukov

Quote:

Originally Posted by JF-AMD

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?
03-07-2011, 11:07 AM
Particle

The problem arises in that "core" doesn't really describe the new architecture. We have to stretch the term around to make it fit. We'd be better off just calling these things modules.

A module has two integer units and one big FP unit that can either function as two smaller FP units or one big one. A module is very much two traditional "cores". FP performance is just variable depending on how the module is being used at a given moment. You could say the module is a dual integer + single floating point core or a dual integer + dual floating point core entirely depending on the instructions currently being executed.
03-07-2011, 11:12 AM
Mech0z

It is 2 cores in one module, but each module shares some stuff (which saves a lot of space or cost) and it does so without a huge performance penalty per core (Think it was 90-95% performance of 2 individual cores) this is just what I remember reading on JFs blog though, might be kinda wrong.
03-07-2011, 11:57 AM
DarthShader

Quote:

Originally Posted by 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

You have to remember that this is only one example of application. (assuming all this is true ofc). This particular software could be a "best case scenario" and average IPC increases could be lower among a variety of workloads.

Similary like SB had a 25% FPS increase in Starcraft c2c, but on average it was more like 7-10% over last generation.
03-07-2011, 12:45 PM
xBanzai89

Quote:

Originally Posted by Manicdan

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

I am confused what you are saying here. Do you mean that the IPC on single bulldozer core is slower then a single Phenom II core? If so that means a 2 module 4 core Bulldozer is slower then the current Phenom II X4. That would make no sense.
03-07-2011, 12:49 PM
informal

Quote:

Originally Posted by Stukov

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?

Each core is a full core.Each integer unit has own FMAC. Nothing in BD module is half good,or makes 2 cores act as 1.5 core. You need to go back and read up on Bulldozer.

Quote:

Originally Posted by xBanzai89

I am confused what you are saying here. Do you mean that the IPC on single bulldozer core is slower then a single Phenom II core? If so that means a 2 module 4 core Bulldozer is slower then the current Phenom II X4. That would make no sense.

No. Module has 1.8x the throughput of a hypothetical Bulldozer based dual core with non shared parts.Nothing to do with previous Family 10h or any other design from AMD.
03-07-2011, 01:45 PM
Manicdan

Quote:

Originally Posted by xBanzai89

I am confused what you are saying here. Do you mean that the IPC on single bulldozer core is slower then a single Phenom II core? If so that means a 2 module 4 core Bulldozer is slower then the current Phenom II X4. That would make no sense.

informal did a decent job explaining

basically if you have 2 threads on 2 modules, you get 2x perf (or like 1.95x in reality it seems)

with 2 threads on the same module, its 1.8x

so while there are 2 int cores per module, you will only 80% of the second one in terms of expected perf. although we have no idea how much this can vary or what cases it will be noticed.

i added the extra bit about only comparing between the same architecture people some people are worried that 1 module will be slower than 2 thuban/deneb cores, which is probably way, way off.