'Balls to the wall' HPC workstation

[ColonelMcNugget]

Hey guys!

 

I'm a researcher working in an infectious disease institute in the UK. We often find ourselves in situations needing big compute resources. I've been tasked with building a (headless) server for use in ML (machine-learning) (CUDA/tensor) applications as well as other things which reqire many CPU threads and RAM. The machine will sit in our office so thermal/noise considerations are key, and it will be in an upright case.

 

The budget is pretty open ended, but I was aiming for the £15,000-£20,000 mark and then some more budget options aswell - but big value for more cash shouldn't be a problem.

 

This is what I have so-far:

 

 

If anybody would like to contribute I'd be very much open to reccommendations on girthier machines.

 

All the best,

 

Dan

 

 

[igormp]

Go with the fatman, or try to up the ram on your gpu-boi, more often than not I eat up all the ram in my systems easily with reasonable-sized models. If you have the budget, I don't see the point on the little boi.

 

Also, have you guys considered doing stuff on the cloud? If you are not running stuff 24/7, it may end up being cheaper and more flexible. Using a V100 anytime you want on gcp really gets you spoiled though haha

[Stu_Bear]

On 1/19/2020 at 5:30 PM, ColonelMcNugget said:

I've been tasked with building a (headless) server for use in ML (machine-learning) (CUDA/tensor) applications as well as other things which reqire many CPU threads and RAM. 

 

Are you sure the software can utilize more than 32 cores on a 7702p?  I only ask cause of AMD cpu architecture on the 64 core models has two banks of cores...some softwares may not play nice with it or not work as expected.

[igormp]

1 hour ago, Stu_Bear said:

 

Are you sure the software can utilize more than 32 cores on a 7702p?  I only ask cause of AMD cpu architecture on the 64 core models has two banks of cores...some softwares may not play nice with it or not work as expected.

Most software meant for actual work usage are usually numa-aware, so that wouldn't be an issue. I can confirm by personal experience that tensorflow works just fine in such cases.

[ColonelMcNugget]

That's amazing guys! Thanks for your replies