[Phoronix] Benchmarking The AMD EPYC 9V64H: Azure HBv5's Custom AMD CPU With HBM3 (MI300C)
Posted by Noble00_@reddit | hardware | View on Reddit | 8 comments
Qesa@reddit
It's worth noting that the comparison is 368 vs 176 cores, so you're only seeing HBM having a particular benefit in the benchmarks where the performance is well over 2x
ComfortableTomato807@reddit
That’s true if we assume the workload performance scales perfectly with the number of cores, ignoring the possibility that the workload is already bottlenecked by the memory despite the core count.
Qesa@reddit
If it's already bandwidth constrained on the HBv4 then the HBv5 (with 7.5x the bandwidth) should be much more than twice as fast.
hurtfulthingsourway@reddit
that's the issue i been seeing with high core count chips they don't have the memory bandwidth to keep the cores fed, some develops are removing the need to load stuff into memory within the Linux kernel and we been seeing a 800%+ increase in throughput with some task, mostly networking.
1731799517@reddit
Oh, thats a BIG caveat.
hardware2win@reddit
Those Azure CPUs are crazy good
ProfHansGruber@reddit
The Algebraic Multi-Grid Benchmark performance is amazing.
Noble00_@reddit (OP)
The elusive MI300C which was announced around last year is available for general availability. Called their, "Azure HBv5" houses the EPYC 9V64H. 92 Zen4 Cores no SMT per CPU with a total of 368 cores with \~400-450GB of HBM3. Benchmarked against 3D v-cached, HBv4, EPYC 9V33X (Genoa-X) for HPC workloads the performance gains are massive.