Fastest GPU for Llama 3.3 70B
B200 leads at 6,127 tokens/s per GPU, 113% ahead of MI355X.
Measured on a single-turn chat workload (8k input / 1k output) at a matched interactivity target of 50 tokens/s per user. Hardware without a measurement at this operating point is not ranked.
| Rank | GPU | Tokens/s per GPU | $ / 1M tokens | Precision | Engine |
|---|---|---|---|---|---|
| 1 | B200 | 6,127 | $0.078 | fp4 | vLLM |
| 2 | MI355X | 2,878 | $0.14 | fp4 | vLLM |
Optimizing for cost instead of speed? See the cheapest GPU for Llama 3.3 70B
Methodology
Every number on this page is a measurement, not a spec-sheet estimate. The InferenceX fleet serves Llama 3.3 70B on real hardware with community serving engines, sweeping concurrency to trace each platform's throughput-versus-interactivity frontier.
Platforms are then read at the same operating point (50 tokens/s per user) so the comparison is iso-interactivity: a GPU cannot win by quoting throughput at an unusably slow per-user speed. Cost converts measured throughput to $ per million total tokens using $/GPU/hr rates from the SemiAnalysis AI Cloud TCO model.
The derivation is shared with the InferenceX overview leaderboard, and results re-run continuously, so this ranking updates as new engine releases and configs land.
Frequently asked questions
- What is the fastest GPU for Llama 3.3 70B inference?
- As of the latest benchmark runs, B200 leads at 6,127 tokens/s per GPU on a single-turn chat workload (8k input / 1k output), measured at a matched interactivity target of 50 tokens/s per user.
- How is "fastest" measured?
- Every platform is read at the same interactivity tier (50 tokens/s per user) on a single-turn chat workload (8k input / 1k output), then ranked by measured throughput per GPU. This is the same derivation the InferenceX overview leaderboard renders, so the ranking can never disagree with the dashboard.
- How often is this ranking updated?
- Benchmarks re-run continuously on the InferenceX cluster fleet; the newest result feeding this ranking landed on 2025-10-29. The page always renders the latest data.