Speed test matrix
Qwen3.6-35B-A3B-FP8 on CMP 170HX 64GB
vllm · 0.32.0rc1.dev249+g18305948a · FP86 contexts (0–248K) × 8 concurrency levels (1–128) · prefill at 8K, 16K, 32K, 64K, 128K
Hardware
- Type
- Discrete GPU
- GPU
- CMP 170HX
- VRAM
- 64 GB
- CPU
- 2x Intel Xeon E5-2630 v3
- RAM
- 64 GB
- OS
- Ubuntu 26.04.1
Engine
- Engine
- vllm
- Version
- 0.32.0rc1.dev249+g18305948a
- Quantization
- FP8
- Backend
- cuda
- Repository
- github.com/vllm-project/vllm
- Build
- ghcr.io/morrowmake/vllm-cmp170hx@sha256:5b45c4d6feafefe2eba962375b0a9a0b64fd22b9e05abb4d1966f14a32a2db79
Run
- Served model
- bench
- Revision
- main
- Benchmark tool
- llm-inference-bench v0.8.0
- Decode test
- Fixed duration
- Prefill mode
- Measured during decode warm-up
- Time per measurement
- 30 s
- Max output
- 8,192 tokens
Launch command
# Untuned out-of-the-box baseline. Spec decode: MTP, 2 tokens lookahead
# 1x NVIDIA CMP 170HX 64GB (cmpunlocker), 180 W limit, PCIe Gen2 x4 | 2x Intel Xeon E5-2630 v3 | 64 GB DDR4 | Ubuntu 26.04.1, NVIDIA driver 610.57.04 | vLLM: Morrowmake vllm-cmp170hx image
vllm serve Qwen/Qwen3.6-35B-A3B-FP8 --max-model-len 262144 --gpu-memory-utilization 0.94 --max-num-seqs 128 --enable-prefix-caching --trust-remote-code --language-model-only --speculative-config {"method":"mtp","num_speculative_tokens":2}Prefill
prompt tok/s by context length8K context5,6991,429 ms TTFT
16K context6,6762,414 ms TTFT
32K context6,2665,120 ms TTFT
64K context5,38811,882 ms TTFT
128K context4,01831,827 ms TTFT
Sustained decode
total tok/s · rows are context length, columns are simultaneous requests| Context | 1 | 2 | 4 | 8 | 16 | 32 | 64 | 128 |
|---|---|---|---|---|---|---|---|---|
| 0 | 200.9164 ms TTFT | 351.3279 ms TTFT | 570.4288 ms TTFT | 887.1417 ms TTFT | 1,262.7471 ms TTFT | 1,677.7638 ms TTFT | 2,234.01,053 ms TTFT | Over capacity |
| 16K | 131.5281 ms TTFT | 227.0457 ms TTFT | 390.8757 ms TTFT | 494.11,082 ms TTFT | 597.21,904 ms TTFT | 722.53,573 ms TTFT | Over capacity | Over capacity |
| 32K | 90.8614 ms TTFT | 169.9823 ms TTFT | 283.91,329 ms TTFT | 341.82,141 ms TTFT | 391.43,923 ms TTFT | Over capacity | Over capacity | Over capacity |
| 64K | 55.6867 ms TTFT | 104.81,084 ms TTFT | 187.41,747 ms TTFT | 206.22,759 ms TTFT | Over capacity | Over capacity | Over capacity | Over capacity |
| 128K | 31.71,257 ms TTFT | 64.21,668 ms TTFT | 112.22,463 ms TTFT | Over capacity | Over capacity | Over capacity | Over capacity | Over capacity |
| 248K | 18.32,095 ms TTFT | 34.72,780 ms TTFT | Over capacity | Over capacity | Over capacity | Over capacity | Over capacity | Over capacity |
Over capacity: not run because this many requests at this context would exceed the server's KV cache.
TTFT is measured after the benchmark warms the server's prompt cache, so it mostly excludes prefill.
All measurements
Latencies are medians. Total tok/s is output across all simultaneous requests.| Phase | Context | Requests at once | Status | Total tok/s | Per request tok/s | TTFT | Time between tokens | Average active requests | Completed requests | Output tokens | Measured for |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Prefill | 8K | 1 | Valid | 5,699.0 | — | 1,429 ms | — | — | 1 | — | — |
| Prefill | 16K | 1 | Valid | 6,676.0 | — | 2,414 ms | — | — | 1 | — | — |
| Prefill | 32K | 1 | Valid | 6,266.0 | — | 5,120 ms | — | — | 1 | — | — |
| Prefill | 64K | 1 | Valid | 5,388.0 | — | 11,882 ms | — | — | 1 | — | — |
| Prefill | 128K | 1 | Valid | 4,018.0 | — | 31,827 ms | — | — | 1 | — | — |
| Sustained decode | 0 | 1 | Valid | 200.9 | 206.4 | 164 ms | 4.8 ms | 1.0 | 0 | 6,026 | 30.0 s |
| Sustained decode | 0 | 2 | Valid | 351.3 | 180.8 | 279 ms | 5.5 ms | 2.0 | 0 | 10,549 | 30.0 s |
| Sustained decode | 0 | 4 | Valid | 570.4 | 145.7 | 288 ms | 6.9 ms | 4.0 | 0 | 17,110 | 30.0 s |
| Sustained decode | 0 | 8 | Valid | 887.1 | 113.7 | 417 ms | 8.8 ms | 8.0 | 0 | 26,606 | 30.0 s |
| Sustained decode | 0 | 16 | Valid | 1,262.7 | 82.2 | 471 ms | 12.2 ms | 16.0 | 0 | 37,859 | 30.0 s |
| Sustained decode | 0 | 32 | Valid | 1,677.7 | 54.6 | 638 ms | 18.3 ms | 32.0 | 0 | 50,321 | 30.0 s |
| Sustained decode | 0 | 64 | Valid | 2,234.0 | 36.3 | 1,053 ms | 27.6 ms | 64.0 | 0 | 67,015 | 30.0 s |
| Sustained decode | 0 | 128 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 16K | 1 | Valid | 131.5 | 132.6 | 281 ms | 7.5 ms | 1.0 | 0 | 3,943 | 30.0 s |
| Sustained decode | 16K | 2 | Valid | 227.0 | 115.2 | 457 ms | 8.7 ms | 2.0 | 0 | 6,807 | 30.0 s |
| Sustained decode | 16K | 4 | Valid | 390.8 | 98.4 | 757 ms | 10.2 ms | 4.0 | 0 | 11,721 | 30.0 s |
| Sustained decode | 16K | 8 | Valid | 494.1 | 62.2 | 1,082 ms | 16.1 ms | 8.0 | 0 | 14,821 | 30.0 s |
| Sustained decode | 16K | 16 | Valid | 597.2 | 37.0 | 1,904 ms | 27.1 ms | 16.0 | 0 | 17,884 | 29.9 s |
| Sustained decode | 16K | 32 | Valid | 722.5 | 22.0 | 3,573 ms | 45.4 ms | 32.0 | 0 | 21,658 | 30.0 s |
| Sustained decode | 16K | 64 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 16K | 128 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 32K | 1 | Valid | 90.8 | 91.9 | 614 ms | 10.9 ms | 1.0 | 0 | 2,723 | 30.0 s |
| Sustained decode | 32K | 2 | Valid | 169.9 | 85.5 | 823 ms | 11.7 ms | 2.0 | 0 | 5,094 | 30.0 s |
| Sustained decode | 32K | 4 | Valid | 283.9 | 71.3 | 1,329 ms | 14.0 ms | 4.0 | 0 | 8,511 | 30.0 s |
| Sustained decode | 32K | 8 | Valid | 341.8 | 42.2 | 2,141 ms | 23.7 ms | 8.0 | 0 | 10,238 | 30.0 s |
| Sustained decode | 32K | 16 | Valid | 391.4 | 23.4 | 3,923 ms | 42.7 ms | 16.0 | 0 | 11,716 | 29.9 s |
| Sustained decode | 32K | 32 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 32K | 64 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 32K | 128 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 64K | 1 | Valid | 55.6 | 55.7 | 867 ms | 18.0 ms | 1.0 | 0 | 1,666 | 30.0 s |
| Sustained decode | 64K | 2 | Valid | 104.8 | 53.0 | 1,084 ms | 18.9 ms | 2.0 | 0 | 3,144 | 30.0 s |
| Sustained decode | 64K | 4 | Valid | 187.4 | 47.1 | 1,747 ms | 21.2 ms | 4.0 | 0 | 5,614 | 30.0 s |
| Sustained decode | 64K | 8 | Valid | 206.2 | 25.4 | 2,759 ms | 39.4 ms | 8.0 | 0 | 6,175 | 29.9 s |
| Sustained decode | 64K | 16 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 64K | 32 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 64K | 64 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 64K | 128 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 128K | 1 | Valid | 31.7 | 32.3 | 1,257 ms | 31.0 ms | 1.0 | 0 | 952 | 30.0 s |
| Sustained decode | 128K | 2 | Valid | 64.2 | 32.4 | 1,668 ms | 30.9 ms | 2.0 | 0 | 1,922 | 29.9 s |
| Sustained decode | 128K | 4 | Valid | 112.2 | 28.3 | 2,463 ms | 35.3 ms | 4.0 | 0 | 3,359 | 29.9 s |
| Sustained decode | 128K | 8 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 128K | 16 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 128K | 32 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 128K | 64 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 128K | 128 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 248K | 1 | Valid | 18.3 | 18.7 | 2,095 ms | 53.4 ms | 1.0 | 0 | 548 | 30.0 s |
| Sustained decode | 248K | 2 | Valid | 34.7 | 17.7 | 2,780 ms | 56.5 ms | 2.0 | 0 | 1,038 | 29.9 s |
| Sustained decode | 248K | 4 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 248K | 8 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 248K | 16 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 248K | 32 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 248K | 64 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 248K | 128 | Over capacity | — | — | — | — | — | — | — | — |
Over capacity: not run because this many requests at this context would exceed the server's KV cache.