Tinygrad Driver testing!
Nearly 2TB of memory and a cross-architecture cluster—what benches do you want?
Deep Dive
A developer is benchmarking MoE speeds on a Blackwell + M3 Ultra RDMA cluster with a bit less than 2TB of RAM. They want ideas for experiments and will stream the tests on the subreddit’s Discord. Follow @mlx_reaper on X.
Key Points
- Tests MoE (Mixture-of-Experts) throughput on a hybrid Blackwell + M3 Ultra RDMA cluster with ~2TB RAM.
- Live streaming of benchmarks on the Tinygrad subreddit Discord; community invited to suggest specific tests.
- Explores cross-vendor AI acceleration using Tinygrad's compiler, potentially lowering hardware dependency costs.
Why It Matters
Demonstrates scalable, open-source multi-architecture AI compute—critical for democratizing large model research beyond single-vendor clusters.