AWS instances based on Graviton2 processors outperform the Xeon based at a fraction of the cost. These are ARM processors so the software stack has to be built for this architecture.
 |
|---|
| AWS c6gd.16xlarge Graviton2 instance with 64 vcore |
 |
|---|
| AWS c5.24xlarge Xeon instance with 96 vcore |
Here there is a collection of wheels already build for Ubuntu 20.04 arm64:
Benchmark of 2700 simulations of a backtest trading strategy.
| AWS instance | vCore | Cost $/hour | Wall time | Arch |
|---|---|---|---|---|
| c6gd.16xlarge | 64 | 2.4 | 12m53s | aarch64 (ARM) |
| c5.24xlarge | 96 | 4.08 | 13m17s | x86 |
| m5.8xlarge | 32 | 1.53 | 32m20s | x86 |
| z1.12xlarge | 24 | 2.2 | 43m03s | x86 |
| Mac mini M1 (*) | 8 | - | 80m00s | arm64 (ARM) |
(*) Just for fun I have included the benchmark using a Mac mini with the amazingly fast new Apple Silicon chips. In fact it is the fastest per core.
Starting point to build Ray from source in Ubuntu arm64 20.04