Hi all,
I am thinking of switching to AMD’s new EPYC processors for a new workstation build. The main use-case for it is CryoEM data processing.
Since it will be mainly used for GPU bound jobs, I am currently looking at ASRocks
EPYC boards. The workstation will have 4 gpus and 256-512 gigs of memory.
On paper, it looks terrific. But I have never used ASrocks boards before.
Any gotchas?
In a university/work setting:
When you build the system, you support it.
When Supermicro, Dell, HPE, etc. ship it, they support it.
At home, your choice.
Newer board would natively support pci-e 4 and faster ram speeds with Rome, not sure if the old gen boards will still do pci-e 4, not sure if gen 4 nvme is important or not
Don’t count on PCIe4 support for older boards, also if it’s just GPU workloads you don’t really need it.
I can understand the need for the extreme core count, but I can’t help but think you could get the same bang with a lower buck with threadripper.
Not up to date on epyc offerings could totally be speaking out of my ass, but if ECC is your worry I can attest that even X370 has ECC support and I’ve only heard of even better support for ECC on threadripper. At least if you have any interest in shaving down the cost.
Otherwise uncharted territory for me, but I imagine that once you’re aiming at epyc the chipsets won’t be limiting you much in any way in terms of feature sets…Unless AMD went a more intel approach with the most recent line up.
Very good point.
Also, with the job this workstation has to do and the amount of ram, you must have ECC. Rule of thumb for memory is 1 error per gigabyte of memory every 24 hours.
I’ve had several asrock boards in my life, all ran for many years or is still running. Actually, can’t remember that I ever removed a board because it didn’t work anymore. This is however on Intel platform, so cannot say if it is so when it comes to their AMD platform, but have no complaints on build quality.
My next workstation build will be AMD, without a doubt, probably also an epyc. If it would be now, I’d take a closer look at gigabyte and probably also MSI for motherboards. Gigabyte for one, makes some nice stuff on server side.
Thanks for the response folks.
I have built workstations/servers using boards from Supermicro/Asus before. Didn’t get a chance to work with Asrock, hence my hesitation.
Given these machines are mostly used for GPU bound jobs, PCIe 4.0 support should not be an issue.
Why EPYC over thread-ripper? Memory.
Unless I am wrong, the existing thread-ripper platforms are limited to 128gb of memory.
And yes, I will be sticking to ECC.
ThreadRipper can do 2TB of ram but you need to use expensive ram for higher capacities
Thread ripper can have 8 dimm slot boards but it’s only quad channel so if you fill in say all 8 slots and those are dual rank dimms (4 ranks per channel max) then you might have a hard time running the rated speed
Where as a 8 dimm slot epyc board is just 8 channels so even with dual rank dims that’s just 2 ranks per channel
Epyc will generally have twice as much L3 Cache, IDK if thats relevant
Good point. But in two of the board specifications I looked at, Asrock Taichi X399 and Asus ROG X399-E, they mention the maximum amount of memory supported is 128. Weird.
Maybe its not officially supported.
One of our neighbouring labs have a a thread-ripper based workstation, I’d be happy to ask them more details and report back.
