Ultimate DC rig, Haswell-E? 8 Cores / 16 Threads on the desktop!

[VirtualLarry]

http://forums.anandtech.com/showthread.php?t=2325520

 

[petrusbroder]

If it really happens, it would be very promising!

 

[sangyup81]

Maybe it'll be overclockable unlike server builds!

 

[Sunny129]

i'm more excited about the 40 lanes of PCIe 3.0 (and hopefully a proper arrangement of PCIe x16 slots)...

...although in order to service several GPU tasks, a multitude of CPU cores couldn't hurt.

 

[spencers]

Sunny, what is the optimal ratio of CPU cores to GPU tasks?

I'm guessing it varies by setup, as I run 2x GPU tasks on one GPU, and they show 0.2CPU each.

And.... Is that ultimately pulling CPU power from the existing 4x CPU tasks that are running?

 

[zzuupp]

Sunny, what is the optimal ratio of CPU cores to GPU tasks?

I'm guessing it varies by setup, as I run 2x GPU tasks on one GPU, and they show 0.2CPU each.

And.... Is that ultimately pulling CPU power from the existing 4x CPU tasks that are running?

Which specific project?
AFAIK, the 0.xx GPU is manually set by the project admins, and may or may not have any bearing on what the GPU task actually uses.

 

[spencers]

Which specific project?
AFAIK, the 0.xx GPU is manually set by the project admins, and may or may not have any bearing on what the GPU task actually uses.

Einstein@Home

Interesting. Thanks for the info!

 

[Sunny129]

Sunny, what is the optimal ratio of CPU cores to GPU tasks?

I'm guessing it varies by setup, as I run 2x GPU tasks on one GPU, and they show 0.2CPU each.

And.... Is that ultimately pulling CPU power from the existing 4x CPU tasks that are running?

yup, you guessed it - it really depends on the setup, which can involve a number of factors. it'll depend on the specific project (and possibly the specific application within the project, for instance SETI@Home Multibeam vs SETI@Home Astropulse), the GPU type (AMD vs nVidia), OS (Windows vs Linux vs etc), and possibly some other factors that i'm not thinking of right now.

also, when you see the "0.2 CPUs + 0.5 NVIDIA GPUs" in BOINC, the "0.2 CPUs" part is really just an estimate, not a strictly-adhered-to value. if you monitor one of your Einstein@Home BRP tasks in the Windows Task Manager, you'll see that its CPU consumption fluctuates over a small range of percentages rather than constantly consuming exactly 20% of a single core. for instance, a single BRP task consumes 4-7% of my 6-core CPU - that's more than just 20% consumption of a single core, yet less than 20% consumption of the whole CPU...so take the "x.xx CPUs" value in BOINC w/ a grain of salt, use the Windows Task Manager to see for yourself how much CPU a GPU task is consuming, and adjust accordingly.

so the optimal ratio of CPU core to GPU tasks will require some experimentation on your part, since nobody will have the exact same mix of hardware, software, OS, project/application, etc. as you. and yes, any CPU resources consumed by GPU tasks will not be available to CPU-only tasks. you should be able to monitor exactly how adding or taking away GPU tasks affects your CPU task runtimes.