R1 runs at 150 tps, 7x faster than DeepSeek's serving
teortaxesTex · x · 2026-08-16
Discussion on inference performance: R1 runs at 150 tps on Rubin, about 7x faster than DeepSeek's official serving, reaching 60% of peak throughput and nearly 1 t/s/watt. Also mentions 200k is lower bound for pure inference; throttling Flash (≈GLM 5.2 tier) to 75 tps could support 200 agents per NPU, 1.6M per SuperPoD. Emphasizes perfect fine-grained replayability provides massive training signal, like an infinite data machine.
Related event: GLM's Rapid Updates and DeepSeek's Compute, Inference Speed Spark Debate(3 posts)→
More from Infra
- Three CoreWeave co-founders have sold $3.3B in shares since IPO amid credit worries — SumitGup · 2026-10-16
- Cherry pi on a Raspberry Pi: the tiny-model rumor developers are watching — sull · 2026-10-16
- CoreWeave's debt costs falling: DDTL spreads tightened from S+9.62% to S+2.25% — BenBajarin · 2026-10-16
- SoftBank lines up US IPOs for SB Energy and AI robotics spinoff Roze, valuations top $50B — nordicinst · 2026-10-16
- TRL reproduces Modal's mixed-precision async RL: plain GRPO collapses, SC+MIS holds 0.72 — QGallouedec · 2026-10-16
- Hugging Face scales Indic pretraining with 2.8T tokens and an open data pipeline — Hugging Face Blog · 2026-10-16