Running Llama 3.3 70B on B200
Quick answer
Llama 3.3 70B sustains 6,127 tokens/s per GPU on B200 at 50 tokens/s per user, which works out to $0.078 per million tokens at hyperscaler pricing, served by vLLM. Fastest measured TTFT: 0.0 ms; fastest TPOT: 0.0 ms (each the best across all configs, not one run).
Benchmarked configs
264
Serving engines
TRTLLM, vLLM
Precisions
fp4, fp8
Run dates
2025-10-29 → 2025-10-29
Throughput at every interactivity target
Serving is a trade-off: push more concurrent users through a GPU and each user's tokens arrive slower. The ladder below reads the measured frontier at each per-user speed target on a single-turn chat workload (8k input / 1k output), using the best engine and precision at that point.
| Per-user target | Tokens/s per GPU | $ / 1M tokens | Engine | Precision |
|---|---|---|---|---|
| 30 tok/s | 8,935 | $0.054 | vLLM | fp4 |
| 50 tok/s | 6,127 | $0.078 | vLLM | fp4 |
| 75 tok/s | 3,989 | $0.12 | vLLM | fp4 |
| 100 tok/s | 2,500 | $0.19 | vLLM | fp4 |
| 150 tok/s | - | - | vLLM | fp4 |
| 200 tok/s | - | - | vLLM | fp4 |
What serving actually costs
Converting the 50 tokens/s per user operating point to $ per million total tokens across rental pricing tiers from the SemiAnalysis AI Cloud TCO model.
| Pricing tier | $ / GPU / hr | $ / 1M tokens |
|---|---|---|
| Hyperscaler | $1.73 | $0.078 |
| Neocloud | $2.07 | $0.094 |
| Retail | $2.60 | $0.12 |
Frequently asked questions
- How fast is Llama 3.3 70B on B200?
- At an interactivity target of 50 tokens/s per user on a single-turn chat workload (8k input / 1k output), B200 sustains 6,127 tokens/s per GPU serving Llama 3.3 70B with vLLM in FP4. Peak measured throughput across all configs is 12,981 tokens/s per GPU.
- How much does it cost to serve Llama 3.3 70B on B200?
- $0.078 per million total tokens at hyperscaler $/GPU/hr pricing, at 50 tokens/s per user. Neocloud and retail rental tiers are tabulated above; slower interactivity targets lower the cost further.
- Which serving engines run Llama 3.3 70B on B200?
- The runs behind this page used TRTLLM, vLLM in FP4, FP8. Engines are rebuilt and re-benchmarked continuously, so the best config can change between visits.
- How are these Llama 3.3 70B numbers measured?
- Every number is measured on real B200 hardware by the InferenceX fleet, sweeping concurrency on a single-turn chat workload (8k input / 1k output) to trace the throughput-versus-interactivity frontier; the newest run landed on 2025-10-29. The same derivation powers the InferenceX overview leaderboard.