Speed test matrix
Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16 on RTX PRO 6000 Blackwell Server Edition 96GB
vllm · 0.30.1rc1.dev452+g73a583112 · BF166 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
- BF16
- 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-BF16 --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}}' --attention-backend FLASHINFER --moe-backend tritonPrefill
prompt tok/s by context length8K context30.9K265 ms TTFT
16K context31.1K522 ms TTFT
32K context32.1K1,006 ms TTFT
64K context29.4K2,193 ms TTFT
128K context23.7K5,420 ms TTFT
251K context17K14,794 ms TTFT
256K context16.7K15,348 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 | 204.275 ms TTFT | 296.6196 ms TTFT | 506.7148 ms TTFT | 142.7197 ms TTFT | 311.8344 ms TTFT | 668.1602 ms TTFT | 1,051.0855 ms TTFT | 1,602.91,248 ms TTFT |
| 16K | 202.872 ms TTFT | 319.390 ms TTFT | 522.2133 ms TTFT | 163.6200 ms TTFT | 430.8286 ms TTFT | 543.3423 ms TTFT | 1,155.6667 ms TTFT | Over capacity |
| 32K | 182.990 ms TTFT | 304.2110 ms TTFT | 207.9161 ms TTFT | 335.1255 ms TTFT | 450.1403 ms TTFT | 704.3692 ms TTFT | Over capacity | Over capacity |
| 64K | 179.2125 ms TTFT | Looped | 208.3245 ms TTFT | Looped | Looped | 798.01,220 ms TTFT | Over capacity | Over capacity |
| 128K | Looped | Looped | Looped | Looped | 736.51,473 ms TTFT | Over capacity | Over capacity | Over capacity |
| 251K | Looped | Looped | Looped | Looped | Over capacity | Over capacity | Over capacity | Over capacity |
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 | 30.9K | — | 265 ms | — | — | 1 | — | — |
| Prefill | 16K | 1 | Valid | 31.1K | — | 522 ms | — | — | 1 | — | — |
| Prefill | 32K | 1 | Valid | 32.1K | — | 1,006 ms | — | — | 1 | — | — |
| Prefill | 64K | 1 | Valid | 29.4K | — | 2,193 ms | — | — | 1 | — | — |
| Prefill | 128K | 1 | Valid | 23.7K | — | 5,420 ms | — | — | 1 | — | — |
| Prefill | 251K | 1 | Valid | 17K | — | 14,794 ms | — | — | 1 | — | — |
| Prefill | 256K | 1 | Valid | 16.7K | — | 15,348 ms | — | — | 1 | — | — |
| Sustained decode | 8K | 1 | Valid | 204.2 | 1435.9 | 75 ms | 2.9 ms | 1.0 | 1 | 4,141 | 20.3 s |
| Sustained decode | 8K | 2 | Valid | 296.6 | 144.3 | 196 ms | 6.9 ms | 2.0 | 2 | 8,896 | 30.0 s |
| Sustained decode | 8K | 4 | Valid | 506.7 | 115.5 | 148 ms | 8.7 ms | 4.0 | 4 | 15,197 | 30.0 s |
| Sustained decode | 8K | 8 | Valid | 142.7 | 75.1 | 197 ms | 13.3 ms | 8.0 | 0 | 4,282 | 30.0 s |
| Sustained decode | 8K | 16 | Valid | 311.8 | 55.4 | 344 ms | 18.1 ms | 16.0 | 0 | 9,352 | 30.0 s |
| Sustained decode | 8K | 32 | Valid | 668.1 | 40.4 | 602 ms | 24.8 ms | 32.0 | 0 | 20,048 | 30.0 s |
| Sustained decode | 8K | 64 | Valid | 1,051.0 | 34.1 | 855 ms | 29.3 ms | 64.0 | 0 | 31,522 | 30.0 s |
| Sustained decode | 8K | 128 | Valid | 1,602.9 | 25.6 | 1,248 ms | 39.0 ms | 128.0 | 0 | 48,052 | 30.0 s |
| Sustained decode | 16K | 1 | Valid | 202.8 | 1446.1 | 72 ms | 2.9 ms | 1.0 | 1 | 4,311 | 21.3 s |
| Sustained decode | 16K | 2 | Valid | 319.3 | 417.9 | 90 ms | 4.2 ms | 2.0 | 2 | 8,022 | 25.1 s |
| Sustained decode | 16K | 4 | Valid | 522.2 | 178.1 | 133 ms | 6.8 ms | 4.0 | 4 | 15,667 | 30.0 s |
| Sustained decode | 16K | 8 | Valid | 163.6 | 75.0 | 200 ms | 13.3 ms | 8.0 | 0 | 4,907 | 30.0 s |
| Sustained decode | 16K | 16 | Valid | 430.8 | 53.8 | 286 ms | 18.6 ms | 16.0 | 0 | 12,924 | 30.0 s |
| Sustained decode | 16K | 32 | Valid | 543.3 | 40.0 | 423 ms | 25.0 ms | 32.0 | 0 | 16,293 | 30.0 s |
| Sustained decode | 16K | 64 | Valid | 1,155.6 | 30.7 | 667 ms | 32.5 ms | 64.0 | 0 | 34,668 | 30.0 s |
| Sustained decode | 16K | 128 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 32K | 1 | Valid | 182.9 | 184.5 | 90 ms | 5.4 ms | 1.0 | 1 | 3,726 | 20.4 s |
| Sustained decode | 32K | 2 | Valid | 304.2 | 310.7 | 110 ms | 4.7 ms | 2.0 | 2 | 9,125 | 30.0 s |
| Sustained decode | 32K | 4 | Valid | 207.9 | 95.6 | 161 ms | 10.5 ms | 4.0 | 0 | 6,234 | 30.0 s |
| Sustained decode | 32K | 8 | Valid | 335.1 | 74.6 | 255 ms | 13.4 ms | 8.0 | 0 | 10,053 | 30.0 s |
| Sustained decode | 32K | 16 | Valid | 450.1 | 52.6 | 403 ms | 19.0 ms | 16.0 | 0 | 13,497 | 30.0 s |
| Sustained decode | 32K | 32 | Valid | 704.3 | 40.3 | 692 ms | 24.8 ms | 32.0 | 0 | 21,128 | 30.0 s |
| Sustained decode | 32K | 64 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 32K | 128 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 64K | 1 | Valid | 179.2 | 180.0 | 125 ms | 5.6 ms | 1.0 | 1 | 5,375 | 30.0 s |
| Sustained decode | 64K | 2 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 64K | 4 | Valid | 208.3 | 100.6 | 245 ms | 10.0 ms | 4.0 | 0 | 6,247 | 30.0 s |
| Sustained decode | 64K | 8 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 64K | 16 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 64K | 32 | Valid | 798.0 | 36.0 | 1,220 ms | 27.8 ms | 32.0 | 0 | 23,927 | 30.0 s |
| Sustained decode | 64K | 64 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 64K | 128 | Over capacity | — | — | — | — | — | — | — | — |
| 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 | Valid | 736.5 | 45.5 | 1,473 ms | 22.0 ms | 16.0 | 0 | 22,095 | 30.0 s |
| Sustained decode | 128K | 32 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 128K | 64 | Over capacity | — | — | — | — | — | — | — | — |
| 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 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 251K | 32 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 251K | 64 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 251K | 128 | Over capacity | — | — | — | — | — | — | — | — |
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.