NVIDIA makes 8-node DGX Spark clusters official: 97.4% scaling efficiency on 900M training
TheZachMueller · x · 2026-10-24
With help from Your Local AI Lab, NVIDIA has officially supported 8-node DGX Spark clusters (ConnectX-7 interconnect), publishing two scaling results on the same hardware:
- Training: a 900M-parameter model trained from scratch hits 7.79x speedup with 97.4% scaling efficiency (near-linear)
- Serving: a 27B model across the same eight nodes gets 1.94x prefill and 4.43x decode speedup
The accompanying chart shows decode scaling on Qwen3.8-27B-NVFP4 with speculative decoding disabled; with ROCEnante, 8-node decode improves from 3.05x to 4.43x. Near-linear training scaling but much weaker inference scaling is the signature of these small clusters.
More from Infra
- RTX 3070 runs quantized Qwen3.8 at 1000 tps prefill with Strata — full config shared — SetNo7692 · 2026-10-24
- OpenAI resumed frontier RL training on Oct 1, 11 days after the containment pause — scaling01 · 2026-10-24
- Cloudflare cuts CPU use 30% on D1 backend with new profiler, Workers Observability drops memory/CPU ~50% — irvinebroque · 2026-10-24
- One NVIDIA Vera Rubin rack would set you back at least ~$7M — LinusEkenstam · 2026-10-24
- DSF4.1 hits 280 tok/s on a single RTX 6000 — community chases 500 tok/s setups — capetorch · 2026-10-24
- PS5 Pro sells out in Toronto as memory shortage hits consumer hardware — Rasmic · 2026-10-24