How Many Cores?!?!?!

[s7e9h3n]

seem like wprime was using 8 cores maybe you can try setting it manually next time.
unless your time is up.

Nope...looks as if Wprime is using all cores. Just ran again:



It seems some form of "Cool and Quiet" is enabled on the board or procs by default - there is no way to shut it off. The cpu's take quite awhile to ramp up to full speed and they apparently ramp down way before the bench is over. The board/cpu's are buggy but at least they're working I'm seeing if I can get a bios update today.

BTW, there is no easy way to overclock this setup. The BIOS is locked down - no real options (besides debugging) to speak of. In fact, there aren't even any options to tweak the memory. The only way I can really think of to possibly push these procs is to edit the cpu registers. For this, I'd probably need Kyosen's input....Kyonsen-San....where are you?

[chew*]

Stephen I wonder if you could bump 1 core with K10 stat.

[Mechromancer]

Stephen I wonder if you could bump 1 core with K10 stat.

Yes, it would be good to know if K10stat works for this CPU. I have a feelings it will work exactly like it does on Particle's Istanbul setup . I think all these server CPUs are hard wired for their clocks and voltages just like Istanbul is. Probably won't be able to overclock unfortunately.

[s7e9h3n]

Stephen I wonder if you could bump 1 core with K10 stat.

Yes, it would be good to know if K10stat works for this CPU. I have a feelings it will work exactly like it does on Particle's Istanbul setup . I think all these server CPUs are hard wired for their clocks and voltages just like Istanbul is. Probably won't be able to overclock unfortunately.

Interesting...never seen that program before, but it does seem to work....at least on one of the CPU's. How do I get it to work on the other?

@ 1800 on one cpu, 32m time drops to sub 6 seconds in Wprime

1024 seems slower

[justapost]

Interesting...never seen that program before, but it does seem to work....at least on one of the CPU's. How do I get it to work on the other?

@ 1800 on one cpu, 32m time drops to sub 6 seconds in Wprime

1024 seems slower

Dunno how to use k10stats for the other cores, but you can use rw-everything to modify the p-states. What have you done in k10stats? Do the p-state-0 modifications work ?

[Lightman]

Oh that was quick, thank you. Guess there is still some room, I assume the NB/IMC is running at <=1600MHz with 1650GHz core speed. The results look like IMC/L3 bottlenecked, can you run with DDR3 at 1066 which should roughly equal DDR2 800. Max possible for DDR3 1333 eight channel is ~85GB/s so this is ~48% efficiency atm, in comparison to ~80% on the quad socket istanbul system with DDR2 800 (theor. max ~51GB/s).
Particle here has an dual istanbul setup, maybe he can run a few benchmarks for comparison. Maybe stream at the same cpu and nb clock speed.

I doubt system has populated all memory channels - kinda hard with only 8GB of RAM (unless 1GB sticks are used). Best to clarify ...

s7e9h3n can you clarify memory configuration? How many DIMMs and which configuration (Single, dual, quad channel per socket)?

NINJA EDIT: Just saw 3.2GHz Magny-Cours CPU-Z

[s7e9h3n]

I doubt system has populated all memory channels - kinda hard with only 8GB of RAM (unless 1GB sticks are used). Best to clarify ...

s7e9h3n can you clarify memory configuration? How many DIMMs and which configuration (Single, dual, quad channel per socket)?

NINJA EDIT: Just saw 3.2GHz Magny-Cours CPU-Z

Board has 16 Slots...only 8 are populated with 1GB sticks. Will upgrade to at least 32GB in the near future

[Lastviking]

Board has 16 Slots...only 8 are populated with 1GB sticks. Will upgrade to at least 32GB in the near future

Looks great , have you tryied to change NB clocks with k10stats?...what is default core on that sick cpu?

[Stefan Payne]

Board has 16 Slots...only 8 are populated with 1GB sticks. Will upgrade to at least 32GB in the near future

That's a little low for a quad core system.
I was hoping to see 2.4GHz at least but at least there is hope for more.

If you're back, try something to bring the NB Clock up.

And interessting to see that it seems that AMD uses the same (crappy) Coolers they use since they broght the socket 754 and 940...

[chew*]

Interesting...never seen that program before, but it does seem to work....at least on one of the CPU's. How do I get it to work on the other?

@ 1800 on one cpu, 32m time drops to sub 6 seconds in Wprime

1024 seems slower

Hmm beats me maybe you can assign the program to 2nd cpu with 2 instances open?

Maybe we need a new version with 12 registers

When a 965BE shows up at your door make friends with that program

[chew*]

Steve try Wprime v 1.55 please.

Oh and while your at it just snatch up the AMD wprime 32m WR

Boot diagnostic mode and run bench 2nd run will be faster usually.

I need to take a emergency flight and bring 2 ln2 pots it will scale with no need for volts if you run into a board limit.

[Particle]

Oh that was quick, thank you. Guess there is still some room, I assume the NB/IMC is running at <=1600MHz with 1650GHz core speed. The results look like IMC/L3 bottlenecked, can you run with DDR3 at 1066 which should roughly equal DDR2 800. Max possible for DDR3 1333 eight channel is ~85GB/s so this is ~48% efficiency atm, in comparison to ~80% on the quad socket istanbul system with DDR2 800 (theor. max ~51GB/s).
Particle here has an dual istanbul setup, maybe he can run a few benchmarks for comparison. Maybe stream at the same cpu and nb clock speed.

Yes, it would be good to know if K10stat works for this CPU. I have a feelings it will work exactly like it does on Particle's Istanbul setup . I think all these server CPUs are hard wired for their clocks and voltages just like Istanbul is. Probably won't be able to overclock unfortunately.

Whoops--almost forgot about this. Here's a run at default for me. I think I must be doing something wrong. Not sure how strongly effected by ECC this is, either. I run with ECC set to maximum protection. Come to think of it, this may be with underclocked memory. Let me check...



Default: 12x2200, DDR2-800, ECC Off...not a whole lot better

[justapost]

Whoops--almost forgot about this. Here's a run at default for me. I think I must be doing something wrong. Not sure how strongly effected by ECC this is, either. I run with ECC set to maximum protection. Come to think of it, this may be with underclocked memory. Let me check...

Default: 12x2200, DDR2-800, ECC Off...not a whole lot better

Ohh thanks, those results look like single socket vs. dual channel.
Do you really have all four channels populated and ran with 12 threads? Beside that with six cores sharing two channels this benchmark is heavi imc speed limited. My 955be and also the AII 240 and 250 get best results if the nb runs one notch below the cpu. I hit a limit once the nb speed reaches 2xmem speed, so the optimum with ddr3 1333 is 2.8GHz/2.6GHz cpu/nb. If it's the same with DDR2 you should not be limited by the stock nb speed however since it's already above 1.6GHz.

Beside that I think only this ES has unlocked multis and voltages and the retail chips will be locked so using k10stats or similar to modify these settings on the chips msr register will not work and oc capability will be limited by the max possible ref HT the boards will offer.

@s7: I'm curious, what msr register did you modify? Was it the controll register or the p-state-0 register?

[SocketMan]

FASN8ing (times 3).
I really hope some of these new G34/C32 boards will be made with "XS
people" in mind.
Definitely sounds like there is a 2P market here.

Many thanks for sharing S7.

[Particle]

Ohh thanks, those results look like single socket vs. dual channel.
Do you really have all four channels populated and ran with 12 threads?

- Dual Opteron 2427s (Istanbul) @ 2.2GHz CPU/2.2GHz NB
- Each CPU has two 2GB DDR2-800 DIMMs installed in dual channel configuration

The numbers are just lower than I'd expect.

[justapost]

- Dual Opteron 2427s (Istanbul) @ 2.2GHz CPU/2.2GHz NB
- Each CPU has two 2GB DDR2-800 DIMMs installed in dual channel configuration

The numbers are just lower than I'd expect.

Check your numa settings in the bios you should get around 20MB/s in that benchmark, just like the other dual socket shanghai/istanbul setups.

[charged3800z24]

it is official, ....where the results