IBT

paradisiac · 20 Jul 2011 at 15:34

i understand that the clock is going to be an indicator in speed, but you cant tell me my 4 core i5 2500k @ 4.8GHz would be the same speed as say a phenom X4 @ 4.8GHz? so there has to be more to it than just the clock speed :confused:

joeyjojo · 20 Jul 2011 at 15:38

Normally yes, however SB supports AVX, effectively doubling the GFLOPS, as it can push 8 operations through per cycle over the Phenom's 4. When AMD release AVX compatible chips they will be on par again.

paradisiac · 20 Jul 2011 at 15:40

joeyjojo said:
Normally yes, however SB supports AVX, effectively doubling the GFLOPS, as it can push 8 instructions through per cycle over the Phenom's 4. When AMD release AVX compatible chips they will be on par again.

so does that mean the 2600k with their HT will be pushing 16 instructions per cycle?

joeyjojo · 20 Jul 2011 at 15:48

Sorry I meant "8 operations" above, changed now.

No, but I don't know enough about micro-architecture to explain why. Hyper threading utilises more of the cores so should move the GFLOP measure closer to the max. However the chip is still only able to do 8 FLOP per core per cycle.

This should all be checked by an expert, this is just my thinking about it as we go along.

paradisiac · 20 Jul 2011 at 15:51

joeyjojo said:
Sorry I meant "8 operations" above, changed now.

No, but I don't know enough about micro-architecture to explain why. Hyper threading utilises more of the cores so should move the GFLOP measure closer to the max. However the chip is still only able to do 8 FLOP per core per cycle.

This should all be checked by an expert, this is just my thinking about it as we go along.

lol ok. i dont actually understand it all, far too complicated. but now i have done my IBT run (124.*** Gflop/sec @4.8GHz

) then i can run a few apps like media converter and see if there is actually any increase in performance

or infact any instability IBT missed

joeyjojo · 20 Jul 2011 at 15:56

Great work, let us know. My guess is yes to the first bit, no to the second

Some links I found interesting:

http://forums.anandtech.com/showthread.php?t=2148217
http://www.guru3d.com/article/phenom-ii-x4-980-be-processor-review/12 <- 50% faster than a phenom without AVX, 100% faster with.

I don't want to confuse anyone with the above posts - SB is obviously faster clock for clock than a Phenom II X4, which clearly is down to the architecture. Only for AVX benchmarks however do we see really huge (double?) the performance.

betaspin · 20 Jul 2011 at 16:39

One interesting observation is that hyperthreading does not work well with IBT and LinX because the background Linpack benchmark runs faster if the number of threads is equal to the number of physical cores. If one starts a recent Linpack benchmark executable on a HT CPU, Linpack initially launches a thread for each virtual core, but after a short while disables half of the threads and continues using the physical cores only. So, for serious floating-point number crunching the i7-2600k appears to have next to no advantages over the i5-2500k.

betaspin · 20 Jul 2011 at 16:45

WingZero30 said:
Edit: Here is a further thread on the Prime95 AVX

http://www.mersenneforum.org/showthread.php?t=15732

It seems some of the FFTs sizes have been coded to use AVX but still seems to be in development phase.

AVX again seems to be almost twice as fast as SSE2. Kind a like doubling Flops from 4 to 8 in double precision.

With cpu solving the FFTs faster, that sure is bound to make the cpu much hotter

This is very good news, thanks, I had not spotted this one.

By the way, the new Intel and PGI C++ and Fortran compilers, as well as even gcc, are supposed to be able to use AVX instructions automatically wherever appropriate, but I have not had the chance to compile a code to see if there are noticeable benefits.

paradisiac · 20 Jul 2011 at 20:13

well its done 2 hours prime testing too

joeyjojo · 20 Jul 2011 at 23:15

betaspin said:
This is very good news, thanks, I had not spotted this one.

By the way, the new Intel and PGI C++ and Fortran compilers, as well as even gcc, are supposed to be able to use AVX instructions automatically wherever appropriate, but I have not had the chance to compile a code to see if there are noticeable benefits.

Indeed, the intel compilers on the cluster at my (now ex-) university had AVX optimisation 6 months ago, but I couldn't do tests as the newest cores they had were Westmeres (dual xeon X5650 boxes, very nice).

I wish they had SB as it would have been significantly faster, as it was almost all floating point arithmetic.

Lunawolfy · 21 Jul 2011 at 10:16

The biggest problem I had with IBT is TIME! I have 8GB of RAM and it takes 120s to do one run! (on max)

Bugger doing 35+ runs

painter24 · 21 Jul 2011 at 10:33

Informative thread. Was only using Prime95 before reading this. Finally stabalised my set up lastnight (with IBT anyway). I'm OC'd to 4.5ghz @ 1.295v and getting 119.*** Gflops after 3 lots of 20 cycles. Max core temp was 69c. Gonna start 8 hrs of Prime later as well for the hell of it.

paradisiac · 21 Jul 2011 at 10:48

Lunawolfy said:
The biggest problem I had with IBT is TIME! I have 8GB of RAM and it takes 120s to do one run! (on max)

Bugger doing 35+ runs

i did 50 runs, at 131ish secs per run and took around 2hours. prime is a minimum recommended of 8hour test, so do the maths :rolleyes:

joeyjojo · 21 Jul 2011 at 10:53

Lunawolfy said:
The biggest problem I had with IBT is TIME! I have 8GB of RAM and it takes 120s to do one run! (on max)

Bugger doing 35+ runs

Compared to a 12 hour prime, 2 hours is pretty rapid :rolleyes:

Also, if it's unstable, it'll almost always find it in the first 5 runs (10 mins).

Martini1991 · 21 Jul 2011 at 10:55

joeyjojo said:
Compared to a 12 hour prime, 2 hours is pretty rapid

Also, if it's unstable, it'll almost always find it in the first 5 runs (10 mins).

I can get IBT to pass 20 runs easily, but I'll BSOD on idle at 4.6GHZ on my 2500k.

joeyjojo · 21 Jul 2011 at 11:07

Martini1991 said:
I can get IBT to pass 20 runs easily, but I'll BSOD on idle at 4.6GHZ on my 2500k.

How many GFLOPS average?

Martini1991 · 21 Jul 2011 at 11:11

joeyjojo said:
How many GFLOPS average?

I can't say I remember, I didn't pay too much attention to it, other than it passed, but messed up idle a while later.
I do know that they were in line with each other, and there wasn't any "outliers", as they, to me, are an indication of instability.

joeyjojo · 22 Jul 2011 at 09:58

Martini1991 said:
I can't say I remember, I didn't pay too much attention to it, other than it passed, but messed up idle a while later.
I do know that they were in line with each other, and there wasn't any "outliers", as they, to me, are an indication of instability.

This might be a mistake I was making a lot before it was explained on this forum. Basically, you need to use enough memory to get the GFLOPS up to a good fraction of the peak GFLOPS (I usually say 80%). At 4.6 on SB with IBT v2.51 you should be getting 4*8*4.6=147.2 peak GFLOPS. 80% of that is about 118 GFLOPS. I suspect that when you stability tested it was doing far less than 118, so wasn't really stressing it, which is why you saw BSOD at idle.

If you fancy testing this out I'd be grateful

paradisiac · 22 Jul 2011 at 10:02

ok, well it did the 50 runs of IBT fine, a good 5hours+ of prime, loads of benchmarks and ran ok for general browsing, office etc. 5mins into crysis and BAM!!!! locked up.

joeyjojo · 22 Jul 2011 at 10:28

paradisiac said:
ok, well it did the 50 runs of IBT fine, a good 5hours+ of prime, loads of benchmarks and ran ok for general browsing, office etc. 5mins into crysis and BAM!!!! locked up.

Graphics card?

Seriously that's pretty surprising.