We are prototyping a performance study (followup) on AWS that has the following environments:
- AWS Trainium EKS
- AWS Trainium Parallel Cluster
- AWS EKS with p5/p5en.48xlarge
- AWS Parallel Cluster with p5/p5en.48xlarge
We are dividing the application space in 32/64 bit. We can run 32 bit apps on 64 but not the other way around. Note that Trainium is only 32 bit.
- amg2023
- kripke
- laghos
- lammps-reax
- mixbench
- osu
- pytorch
- pytorch
- inference-perf looks good, but isn't ready yet
- fmperf
- fmwork
- ai-benchmark
- hugging-face from Angel, note deprecated
- DeepGEMM
- gpu-fryer Already has a container ghcr.io/huggingface/gpu-fryer:latest. We might want to rebuild if a common base is desired.
- gpu-burn
- DualPipe
- nccl-tests