Single 8x 5090 Rig Hits 167k tokens/s Training Throughput, Beating DDP
jon_durbin · x · 2026-07-30
The author demonstrated a training throughput of 167k tokens/s on a single machine equipped with 8x 5090 GPUs, using the Parallax framework on a toy MoE model (5B total / 1B active parameters).
Key performance and optimization metrics:
- High MFU: Achieved an equivalent 60% Model FLOPs Utilization under BF16.
- Better Loss Curve: Beats the DDP (Distributed Data Parallel) baseline at matched steps.
- Efficient Convergence: Reached 3.068 nats at 19B tokens (Llama-3 tokenizer). Compared to the DDP baseline, it reaches the token budget 15% faster and 64% cheaper.
The author speculates this performance gap is largely due to using separate Muon optimizers for experts, incorporating surrogate feedback and fold normalization. They are now preparing for larger-scale training runs.
Related event: 8x5090 Single Node Achieves 167k tokens/s Training Throughput(2 posts)→
More from Infra
- NVIDIA launches RTX Spark for local AI on Windows PCs — ostrisai · 2026-10-16
- SkyRL ships Day-0 RL support for Beam with +62% LoRA decode throughput in vLLM — vllm_project · 2026-10-16
- DFlash2 delivers up to 3.03x decode-only speedup for Beam across five benchmarks — vllm_project · 2026-10-16
- vLLM's Beam serving kernels cut GB200 forward time from 8.15 to 6.47 ms at TP8 batch 128 — vllm_project · 2026-10-16
- Reflection's 501B MoE Beam hits 463 tok/s on 8x GB200 with NVFP4 at day 0 — zhyncs42 · 2026-10-16
- Gemma 4 31B QAT on one 24GB card: 46 tok/s plain, ~80 with draft model on RTX 4090 — Subject_Poet_3861 · 2026-10-16