I think that part of the Aurora supercomputer is getting decommissioned because there’s suddenly a large supply of Intel Xeon Max 9480 Scalable processors available on Ebay and Newegg. This is the CPU that has 64GB of HBM memory that it can boot alone with.
Now does anyone know if this processor would work in a workstation motherboard like the ASUS Pro WS W790-ACE? Its not on the list of supported CPUs but there’s a ServeTheHome Forum post that suggests scalable processors work in it. The highest tested proc though was a 8480 and idk if the HBM component requires a special bios. After more searching it looks like no, the HBM requires particular support but there’s Gigabyte workstation motherboards available
I was today years old when I found out about processors with their own ram memory are a thing . Also what possibly would you need a 52 core processor for?
Wow that deal was short lived, though there’s still more listings for slightly more. We bought two for workstations, one will be dedicated for 3d scanning and the other for flow simulations. I’m very excited as these are two applications that should be greatly accelerated by the HBM cache layer. When not 3d scanning that machine will be my AI training platform as well.
What motherboard are you going to run them on?
I got some and have been having some trouble tuning them to get best performance; There as so many OS tuning knobs to touch I haven’t been able to optimize yet.
Nice, I ended up getting the Asus Z13PE-D16. I purchased from the same seller back in June and got production (non-ES/QS) processors so it is legit.
The HBM memory seems to be fairly high latency, either that or the mesh is severely bogged down by all the memory traffic; if the former is true we should be able to get good performance by running HBM and DRAM in parallel using the “Flat Mode” and making sure programs access the correct memory via numactl.
This has given me the motivation to work on my system more today.
I suppose the memory latency isn’t that high for a “big” system, chipsandcheese clocked the new Turin 9575F’s loaded latency at 155ns, and 118ns unloaded, which is almost exactly the same as the 9480’s.
MLC gives a good report on memory latency, here’s what reads look like for the 9480s using AVX512:
. Also what possibly would you need a 52 core processor for?