Speed test matrix
Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8 on RTX PRO 6000 Blackwell Server Edition 96GB
vllm · 0.30.1rc1.dev452+g73a583112 · FP86 contexts (8K–251K) × 8 concurrency levels (1–128) · prefill at 8K, 16K, 32K, 64K, 128K, 251K, 256K
Hardware
- Type
- Discrete GPU
- GPU
- RTX PRO 6000 Blackwell Server Edition
- VRAM
- 96 GB
- OS
- linux
Engine
- Engine
- vllm
- Version
- 0.30.1rc1.dev452+g73a583112
- Quantization
- FP8
- Backend
- cuda
- Repository
- github.com/vllm-project/vllm
- Build
- cuda-13.2
- Commit
73a5831127a9
Run
- Served model
- nemotron
- Revision
- main
- Benchmark tool
- llm-inference-bench v0.7.6
- Decode test
- Fixed duration
- Prefill mode
- Measured during decode warm-up
- Time per measurement
- 30 s
- Max output
- 4,096 tokens
Launch command
"$HOME/env_rtxpro6k-nvidia-vllm-73a5831127a9/bin/vllm" serve nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8 --served-model-name nemotron --trust-remote-code --tensor-parallel-size 1 --reasoning-parser nemotron_v3 --enable-auto-tool-choice --tool-call-parser qwen3_coder --max-model-len 262144 --gpu-memory-utilization 0.83 --host 0.0.0.0 --port 8000 --api-key <redacted> --video-pruning-rate 0.5 --max-num-seqs 384 --allowed-local-media-path / --media-io-kwargs '{"video":{"fps":2,"num_frames":256}}' --kv-cache-dtype fp8 --attention-backend FLASHINFER --moe-backend tritonPrefill
prompt tok/s by context length8K context45.1K182 ms TTFT
16K context45.3K358 ms TTFT
32K context43.6K741 ms TTFT
64K context36.6K1,757 ms TTFT
128K context27.4K4,694 ms TTFT
251K context18.2K13,870 ms TTFT
256K context17.9K14,365 ms TTFT
Sustained decode
total tok/s · rows are context length, columns are simultaneous requests| Context | 1 | 2 | 4 | 8 | 16 | 32 | 64 | 128 |
|---|---|---|---|---|---|---|---|---|
| 8K | 313.479 ms TTFT | 545.0141 ms TTFT | 617.0178 ms TTFT | 1,168.4Requests queued227 ms TTFT | 2,239.9374 ms TTFT | 940.3669 ms TTFT | 1,285.51,263 ms TTFT | 1,762.02,065 ms TTFT |
| 16K | 295.986 ms TTFT | 399.7160 ms TTFT | 720.4180 ms TTFT | 1,113.2271 ms TTFT | 2,239.4410 ms TTFT | 885.9643 ms TTFT | 1,166.61,144 ms TTFT | 2,188.41,849 ms TTFT |
| 32K | 279.280 ms TTFT | 481.795 ms TTFT | 721.3137 ms TTFT | 1,153.2223 ms TTFT | 967.6514 ms TTFT | 1,209.9677 ms TTFT | 1,464.01,206 ms TTFT | 2,336.32,131 ms TTFT |
| 64K | Looped | Looped | Looped | Looped | Looped | Looped | 2,043.12,400 ms TTFT | 3,072.55,350 ms TTFT |
| 128K | Looped | Looped | Looped | Looped | Looped | Looped | Looped | Over capacity |
| 251K | Looped | Looped | Looped | Looped | Looped | Looped | 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 | 45.1K | — | 182 ms | — | — | 1 | — | — |
| Prefill | 16K | 1 | Valid | 45.3K | — | 358 ms | — | — | 1 | — | — |
| Prefill | 32K | 1 | Valid | 43.6K | — | 741 ms | — | — | 1 | — | — |
| Prefill | 64K | 1 | Valid | 36.6K | — | 1,757 ms | — | — | 1 | — | — |
| Prefill | 128K | 1 | Valid | 27.4K | — | 4,694 ms | — | — | 1 | — | — |
| Prefill | 251K | 1 | Valid | 18.2K | — | 13,870 ms | — | — | 1 | — | — |
| Prefill | 256K | 1 | Valid | 17.9K | — | 14,365 ms | — | — | 1 | — | — |
| Sustained decode | 8K | 1 | Valid | 313.4 | 4894.4 | 79 ms | 0.2 ms | 1.0 | 2 | 8,297 | 26.5 s |
| Sustained decode | 8K | 2 | Valid | 545.0 | 2025.0 | 141 ms | 0.5 ms | 2.0 | 4 | 16,351 | 30.0 s |
| Sustained decode | 8K | 4 | Valid | 617.0 | 497.8 | 178 ms | 3.2 ms | 4.0 | 4 | 18,510 | 30.0 s |
| Sustained decode | 8K | 8 | Requests queued | 1,168.4 | 412.5 | 227 ms | 4.0 ms | 7.9 | 8 | 35,051 | 30.0 s |
| Sustained decode | 8K | 16 | Valid | 2,239.9 | 123.6 | 374 ms | 8.4 ms | 16.0 | 16 | 67,190 | 30.0 s |
| Sustained decode | 8K | 32 | Valid | 940.3 | 69.3 | 669 ms | 14.4 ms | 32.0 | 0 | 28,205 | 30.0 s |
| Sustained decode | 8K | 64 | Valid | 1,285.5 | 51.2 | 1,263 ms | 19.5 ms | 64.0 | 0 | 38,550 | 30.0 s |
| Sustained decode | 8K | 128 | Valid | 1,762.0 | 43.6 | 2,065 ms | 23.0 ms | 128.0 | 0 | 52,834 | 30.0 s |
| Sustained decode | 16K | 1 | Valid | 295.9 | 1974.3 | 86 ms | 0.5 ms | 1.0 | 2 | 8,085 | 27.3 s |
| Sustained decode | 16K | 2 | Valid | 399.7 | 1005.2 | 160 ms | 2.5 ms | 2.0 | 2 | 11,990 | 30.0 s |
| Sustained decode | 16K | 4 | Valid | 720.4 | 622.9 | 180 ms | 3.3 ms | 4.0 | 4 | 21,608 | 30.0 s |
| Sustained decode | 16K | 8 | Valid | 1,113.2 | 146.4 | 271 ms | 6.8 ms | 8.0 | 8 | 32,546 | 29.2 s |
| Sustained decode | 16K | 16 | Valid | 2,239.4 | 284.1 | 410 ms | 5.8 ms | 16.0 | 16 | 67,180 | 30.0 s |
| Sustained decode | 16K | 32 | Valid | 885.9 | 70.5 | 643 ms | 14.2 ms | 32.0 | 0 | 26,570 | 30.0 s |
| Sustained decode | 16K | 64 | Valid | 1,166.6 | 50.5 | 1,144 ms | 19.8 ms | 64.0 | 0 | 34,998 | 30.0 s |
| Sustained decode | 16K | 128 | Valid | 2,188.4 | 39.5 | 1,849 ms | 25.3 ms | 128.0 | 0 | 65,661 | 30.0 s |
| Sustained decode | 32K | 1 | Valid | 279.2 | 281.0 | 80 ms | 3.6 ms | 1.0 | 2 | 8,377 | 30.0 s |
| Sustained decode | 32K | 2 | Valid | 481.7 | 501.0 | 95 ms | 2.7 ms | 2.0 | 4 | 14,450 | 30.0 s |
| Sustained decode | 32K | 4 | Valid | 721.3 | 179.2 | 137 ms | 5.6 ms | 4.0 | 4 | 21,639 | 30.0 s |
| Sustained decode | 32K | 8 | Valid | 1,153.2 | 134.7 | 223 ms | 7.5 ms | 7.9 | 8 | 34,591 | 30.0 s |
| Sustained decode | 32K | 16 | Valid | 967.6 | 90.9 | 514 ms | 11.0 ms | 16.0 | 0 | 29,008 | 30.0 s |
| Sustained decode | 32K | 32 | Valid | 1,209.9 | 68.2 | 677 ms | 14.7 ms | 32.0 | 0 | 36,298 | 30.0 s |
| Sustained decode | 32K | 64 | Valid | 1,464.0 | 48.0 | 1,206 ms | 20.8 ms | 64.0 | 0 | 43,899 | 30.0 s |
| Sustained decode | 32K | 128 | Valid | 2,336.3 | 36.2 | 2,131 ms | 27.6 ms | 128.0 | 0 | 70,065 | 30.0 s |
| Sustained decode | 64K | 1 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 64K | 2 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 64K | 4 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 64K | 8 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 64K | 16 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 64K | 32 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 64K | 64 | Valid | 2,043.1 | 41.7 | 2,400 ms | 24.0 ms | 64.0 | 0 | 61,292 | 30.0 s |
| Sustained decode | 64K | 128 | Valid | 3,072.5 | 29.5 | 5,350 ms | 33.9 ms | 128.0 | 0 | 92,129 | 30.0 s |
| Sustained decode | 128K | 1 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 128K | 2 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 128K | 4 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 128K | 8 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 128K | 16 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 128K | 32 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 128K | 64 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 128K | 128 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 251K | 1 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 251K | 2 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 251K | 4 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 251K | 8 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 251K | 16 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 251K | 32 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 251K | 64 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 251K | 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.