Two M3 Ultras linked by Thunderbolt RDMA for local tensor parallelism
antirez · x · 2026-07-20
A first test of DwarfStar distributed tensor parallelism over Thunderbolt RDMA on two M3 Ultra machines.
For DeepSeek V4 Flash Q2, the setup improved prefill throughput from 593 tok/s to 642 tok/s, while decode dropped from 39 tok/s to 33 tok/s. For Q4, prefill rose from 589 tok/s to 677 tok/s (+14.9%), while decode went from 35.5 tok/s on a single M3 Ultra to 31.4 tok/s with tensor parallelism, implying a small decode penalty for the gain in prefill speed.
More from Infra
- Resin lets an agent build standalone LLM inference engines; vLLM and SGLang optimize DeepSeek kernels differently — AccBalanced · 2026-10-27
- Local benchmark: Flint matches Qwen3.6 quality at 50/52 while running 3.6x faster on M4 Pro — RichardSocher · 2026-10-27
- Anthropic S-1: CSP marketplace revenue jumped from 11% to 48% in three years — AccBalanced · 2026-10-27
- Alchemy adds effect/Workflow support across Durable Objects, Workflows and Lambda Durable Functions — samgoodwin89 · 2026-10-27
- Alchemy's Lambda Web Functions run plain Node web servers on Lambda, no OpenNext needed — samgoodwin89 · 2026-10-27
- Alchemy Cuts Cloudflare Bindings CPU Time by Up to 166× Via Profiling — samgoodwin89 · 2026-10-27