DeepSeek Hits 243 tok/s on Dual RTX 6000 with Speculative Decoding
TheZachMueller · x · 2026-08-02
A developer tested DeepSeek-V4-Flash-0731 with DSpark speculative decoding on 2× RTX PRO 6000 Blackwell (TP=2).
- Performance: Achieved a median single-stream throughput of 243 tok/s, a 3.1× speedup over the no-spec baseline. Aggregate throughput hit 299 tok/s (c2) and 403 tok/s (c4).
- Stability: Draft acceptance rate was 74–76%, sustaining 237–245 tok/s over a 3-hour mixed-load soak test with zero failures.
- Bottleneck: Avoided the expected SM120 kernel wall. The real issue is a deterministic failure when maxnumseqs > 4 due to a hardcoded prefill chunk size feeding an empty slice to the sparse kernel. Capping it at 4 resolved the issue, making it the author's new daily driver.
Related event: DeepSeek V4 Flash Sparks a Wave of Local Deployment Tests(26 posts)→
More from Infra
- NVIDIA quietly delists the entry-level RTX 5050 from its marketplace, following the RTX 5090 — chemist_slime · 2026-10-16
- Grok claims 816B MoE model runs locally on dual M5 Ultra Mac Studios via MLX — grok · 2026-10-16
- Same 1,000 Cloudflare Durable Objects: $4,050/month cut to $22 with one trick — threepointone · 2026-10-16
- NVIDIA ships FP4-quantized Qwen3.8-Flash for local use, trending on Hugging Face — nvidia · 2026-10-16
- Baidu's Virtual Head Attention tames the KV cache, FLOPs, quality triangle in long-context LLMs — baidu · 2026-10-16
- DuoDiscover agents auto-build fast eval proxies, scoring 2.68x TPU kernel speedup — berkeley_ai · 2026-10-16