Speed test matrix
NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4 on CMP 170HX 64GB
vllm · 0.32.0rc1.dev249+g18305948a · NVFP4 mixed6 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
- NVFP4 mixed
- 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: none # 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 nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4 --max-model-len 262144 --gpu-memory-utilization 0.94 --max-num-seqs 128 --enable-prefix-caching --trust-remote-code
Prefill
prompt tok/s by context length8K context3,9022,140 ms TTFT
16K context3,8434,297 ms TTFT
32K context3,7198,838 ms TTFT
64K context3,47718,860 ms TTFT
128K context3,01543,444 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 | 82.6129 ms TTFT | 143.5227 ms TTFT | 246.3702 ms TTFT | 398.5590 ms TTFT | 598.7544 ms TTFT | 832.0851 ms TTFT | 1,076.51,570 ms TTFT | 1,078.6Requests queued1,749 ms TTFT |
| 16K | 82.01,153 ms TTFT | 141.51,796 ms TTFT | 238.72,954 ms TTFT | 382.15,070 ms TTFT | 571.79,366 ms TTFT | 791.617,846 ms TTFT | 666.6Requests queued35,001 ms TTFT | Over capacity |
| 32K | Looped | 137.8895 ms TTFT | 234.81,419 ms TTFT | Looped | 553.04,322 ms TTFT | 759.48,264 ms TTFT | Over capacity | Over capacity |
| 64K | Looped | 135.31,952 ms TTFT | 228.33,080 ms TTFT | 360.85,078 ms TTFT | 512.99,011 ms TTFT | 687.417,083 ms TTFT | Over capacity | Over capacity |
| 128K | 78.52,901 ms TTFT | Looped | 218.47,343 ms TTFT | 326.711,670 ms TTFT | 451.421,062 ms TTFT | Over capacity | Over capacity | Over capacity |
| 248K | 75.23,173 ms TTFT | 123.54,119 ms TTFT | 196.36,112 ms TTFT | 285.29,663 ms TTFT | Over capacity | Over capacity | Over capacity | Over capacity |
Requests queued: the server could not run all requests at once, so some waited in a queue. Looped: the model kept repeating the same text, so the benchmark discarded this measurement. 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 | 3,902.0 | — | 2,140 ms | — | — | 1 | — | — |
| Prefill | 16K | 1 | Valid | 3,843.0 | — | 4,297 ms | — | — | 1 | — | — |
| Prefill | 32K | 1 | Valid | 3,719.0 | — | 8,838 ms | — | — | 1 | — | — |
| Prefill | 64K | 1 | Valid | 3,477.0 | — | 18,860 ms | — | — | 1 | — | — |
| Prefill | 128K | 1 | Valid | 3,015.0 | — | 43,444 ms | — | — | 1 | — | — |
| Sustained decode | 0 | 1 | Valid | 82.6 | 82.7 | 129 ms | 12.1 ms | 1.0 | 0 | 2,479 | 30.0 s |
| Sustained decode | 0 | 2 | Valid | 143.5 | 71.9 | 227 ms | 13.9 ms | 2.0 | 0 | 4,304 | 30.0 s |
| Sustained decode | 0 | 4 | Valid | 246.3 | 61.7 | 702 ms | 16.2 ms | 4.0 | 0 | 7,388 | 30.0 s |
| Sustained decode | 0 | 8 | Valid | 398.5 | 50.2 | 590 ms | 19.9 ms | 8.0 | 0 | 11,952 | 30.0 s |
| Sustained decode | 0 | 16 | Valid | 598.7 | 37.9 | 544 ms | 26.4 ms | 16.0 | 0 | 17,952 | 30.0 s |
| Sustained decode | 0 | 32 | Valid | 832.0 | 26.4 | 851 ms | 37.8 ms | 32.0 | 0 | 24,960 | 30.0 s |
| Sustained decode | 0 | 64 | Valid | 1,076.5 | 17.1 | 1,570 ms | 58.4 ms | 64.0 | 0 | 32,256 | 30.0 s |
| Sustained decode | 0 | 128 | Requests queued | 1,078.6 | 12.5 | 1,749 ms | 80.0 ms | 87.0 | 0 | 32,364 | 30.0 s |
| Sustained decode | 16K | 1 | Valid | 82.0 | 82.1 | 1,153 ms | 12.2 ms | 1.0 | 0 | 2,459 | 30.0 s |
| Sustained decode | 16K | 2 | Valid | 141.5 | 69.9 | 1,796 ms | 14.3 ms | 2.0 | 0 | 4,246 | 30.0 s |
| Sustained decode | 16K | 4 | Valid | 238.7 | 57.5 | 2,954 ms | 17.4 ms | 4.0 | 0 | 7,160 | 30.0 s |
| Sustained decode | 16K | 8 | Valid | 382.1 | 43.6 | 5,070 ms | 22.9 ms | 8.0 | 0 | 11,464 | 30.0 s |
| Sustained decode | 16K | 16 | Valid | 571.7 | 29.6 | 9,366 ms | 33.8 ms | 16.0 | 0 | 17,152 | 30.0 s |
| Sustained decode | 16K | 32 | Valid | 791.6 | 17.2 | 17,846 ms | 58.0 ms | 32.0 | 0 | 23,776 | 30.0 s |
| Sustained decode | 16K | 64 | Requests queued | 666.6 | 6.7 | 35,001 ms | 149.7 ms | 62.0 | 0 | 19,986 | 30.0 s |
| Sustained decode | 16K | 128 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 32K | 1 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 32K | 2 | Valid | 137.8 | 68.6 | 895 ms | 14.6 ms | 2.0 | 0 | 4,132 | 30.0 s |
| Sustained decode | 32K | 4 | Valid | 234.8 | 57.8 | 1,419 ms | 17.3 ms | 4.0 | 0 | 7,044 | 30.0 s |
| Sustained decode | 32K | 8 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 32K | 16 | Valid | 553.0 | 31.8 | 4,322 ms | 31.5 ms | 16.0 | 0 | 16,576 | 30.0 s |
| Sustained decode | 32K | 32 | Valid | 759.4 | 20.2 | 8,264 ms | 49.6 ms | 32.0 | 0 | 22,784 | 30.0 s |
| Sustained decode | 32K | 64 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 32K | 128 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 64K | 1 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 64K | 2 | Valid | 135.3 | 67.2 | 1,952 ms | 14.9 ms | 2.0 | 0 | 4,058 | 30.0 s |
| Sustained decode | 64K | 4 | Valid | 228.3 | 55.4 | 3,080 ms | 18.0 ms | 4.0 | 0 | 6,848 | 30.0 s |
| Sustained decode | 64K | 8 | Valid | 360.8 | 41.6 | 5,078 ms | 24.0 ms | 8.0 | 0 | 10,824 | 30.0 s |
| Sustained decode | 64K | 16 | Valid | 512.9 | 27.0 | 9,011 ms | 37.1 ms | 16.0 | 0 | 15,376 | 30.0 s |
| Sustained decode | 64K | 32 | Valid | 687.4 | 15.5 | 17,083 ms | 64.7 ms | 32.0 | 0 | 20,608 | 30.0 s |
| Sustained decode | 64K | 64 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 64K | 128 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 128K | 1 | Valid | 78.5 | 78.9 | 2,901 ms | 12.7 ms | 1.0 | 0 | 2,355 | 30.0 s |
| Sustained decode | 128K | 2 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 128K | 4 | Valid | 218.4 | 51.2 | 7,343 ms | 19.6 ms | 4.0 | 0 | 6,553 | 30.0 s |
| Sustained decode | 128K | 8 | Valid | 326.7 | 33.8 | 11,670 ms | 29.6 ms | 8.0 | 0 | 9,800 | 30.0 s |
| Sustained decode | 128K | 16 | Valid | 451.4 | 19.3 | 21,062 ms | 51.7 ms | 16.0 | 0 | 13,536 | 30.0 s |
| Sustained decode | 128K | 32 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 128K | 64 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 128K | 128 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 248K | 1 | Valid | 75.2 | 75.8 | 3,173 ms | 13.2 ms | 1.0 | 0 | 2,257 | 30.0 s |
| Sustained decode | 248K | 2 | Valid | 123.5 | 61.1 | 4,119 ms | 16.4 ms | 2.0 | 0 | 3,702 | 30.0 s |
| Sustained decode | 248K | 4 | Valid | 196.3 | 46.6 | 6,112 ms | 21.5 ms | 4.0 | 0 | 5,888 | 30.0 s |
| Sustained decode | 248K | 8 | Valid | 285.2 | 31.0 | 9,663 ms | 32.3 ms | 8.0 | 0 | 8,552 | 30.0 s |
| Sustained decode | 248K | 16 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 248K | 32 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 248K | 64 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 248K | 128 | Over capacity | — | — | — | — | — | — | — | — |
Requests queued: the server could not run all requests at once, so some waited in a queue. Looped: the model kept repeating the same text, so the benchmark discarded this measurement. Over capacity: not run because this many requests at this context would exceed the server's KV cache.