Qwen3.6-27B-FP8 Inference Benchmarks on Single RTX 6000 Ada
Temporary-Owl1725 · reddit · 2026-07-06
Benchmark results for deploying Qwen3.6-27B-FP8 on a single RTX 6000 Ada (48GB) using vLLM 0.19. Tested with ShareGPT samples at a concurrency of 8-16, it achieved a TTFT p50 of 0.48s, TPOT p50 of 29.2ms/token, a peak throughput of 668.5 tok/s, and a max KV cache usage of 32.7%.
More from Infra
- DeepSpeed shows AutoEP/AutoTP/AutoSP parallelism gains, hosts office hours — PyTorch · 2026-10-27
- Cloudflare Durable Objects validated as top architecture for AI agent WebSocket servers — tristanbob · 2026-10-27
- Gary Marcus: With Q3 data added, Anthropic's compute-efficiency growth shows deceleration — GaryMarcus · 2026-10-27
- SpaceX locks in $40B GPU financing deal with Apollo, the largest ever committed — TechCrunch AI · 2026-10-27
- Inference is just Kubernetes with a multi-trillion-dollar market, and the communities overlap — matt_slotnick · 2026-10-27
- vLLM eyes Rust migration as inference engines race to rewrite the stack — AccBalanced · 2026-10-27