Guidance on Intel Scalable HPC optimized processors

[whitey]

Looking to gain feedback/guidance if there is a guide out there that covers Intel scalable targeted for HPC workloads. Have a request for 240 HPC nodes and would greatly appreciate a point in the right direction as I have not been nearly as up to date on HW advances in the HPC arena for the last couple of years as I should have been.

Cluster is used for highly parallel operations
Best memory bandwidth per core is desirable
What else am I forgetting to mention.

TIA, whitey

 

[Patrick]

OPA or no? Omni-Path enabled the popular ones I believe are the Gold 6148 and Platinum 8160's.

Let me know if you need me to ping someone.

 

[whitey]

Let's say non-OPA for now. We are a pretty big IB shop for HPC, currently at 100G IB. Looking more for a general reference/guideline that maybe has a small/medium/large or something along the lines of 'If your HPC cluster runs X types of workloads use this, else use Y' maybe more in line w/ more/less cores or if there is a sweet spot for HPC nodes. Wanna future proof us up a bit while ensuring as much compatibility as reasonably possible. The CPU is the thing that eludes me right now, is there a 'best of breed' for HPC workloads that jumps out at anyone that lives/breaths/sleeps this IT discipline?

 

[Patrick]

Well, I think Intel will just tell you get all Platinum 8180's. I think that shops usually find they hit limits beyond just cores.

 

[Evan]

I would look at 2 options that I think are a bit better value if no omnipath requirement.

Platinum 8168 (24 core @ 2.7ghz, 205w tdp)
Gold 6150 (18 core @ 2.7ghz, 165w tdp)

The gold being a much more cost effective per core cost.

Thing is you need to figure out your base cost of the rest of server and per port connection costs etc (+ of course real estate and power costs)
You also need to carefully consider how your exact workload behaves.

Blade or rack servers out of interest ?