Head-to-head AI inference benchmark comparison of B200 (NVIDIA Blackwell) and H200 (NVIDIA Hopper) on Llama 3.3 70B. Latency, throughput, and cost across LLM workloads. Use the chart controls below to switch sequences, precisions, and metrics — same interactions as the main inference chart.
B200 / H200 on Llama 3.3 70B at 45 tok/s/user: 5994 / 2978 tok/s/chip, $0.08 / $0.11 per million tokens. B200 is 42% cheaper per token; B200 delivers 101% more tok/s/chip.
Around the middle of the 14–138 tok/s/user interactivity band, at 76 tok/s/user on Llama 3.3 70B: B200 runs 3981 tok/s/chip at $0.12/M tokens, H200 runs 1930 at $0.18/M. B200 is 45% cheaper per token; B200 delivers 106% more tok/s/chip.
Setting 108 tok/s/user as the target on Llama 3.3 70B, B200 produces 2352 tok/s/chip ($0.20 per million tokens) and H200 produces 1152 ($0.29). B200 is 44% cheaper per token; B200 delivers 104% more tok/s/chip. (Numbers reflect the default 8k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | B200:5993.5H200:2977.6 | B200:3981.0H200:1930.3 | B200:2351.5H200:1152.0 |
| Cost ($/M tok) | B200:$0.080H200:$0.114 | B200:$0.121H200:$0.176 | B200:$0.204H200:$0.294 |
| tok/s/MW | B200:3505000H200:2173427 | B200:2328088H200:1408963 | B200:1375171H200:840884 |
| Concurrency | B200:~32H200:~32 | B200:~28H200:~11 | B200:~9H200:~8 |
Agentic inference metrics from the AgentX scenario and fixed-sequence inference metrics across models, hardware configurations, and serving parameters.
TCO $/chip/hr:
Source: SemiAnalysis Market July 2026 Pricing Surveys & AI Cloud TCO Model
Note: Disaggregated inference configurations (e.g., MoRI SGLang, Dynamo TRTLLM) calculate cost per million tokens per decode chip or per prefill chip, rather than per total chip count. This makes direct token cost comparison with aggregated configs not an apples-to-apples comparison.
Note: Disaggregated inference configurations (e.g., MoRI SGLang, Dynamo TRTLLM) calculate input throughput per decode chip or per prefill chip, rather than per total chip count. This makes direct input throughput comparison with aggregated configs not an apples-to-apples comparison.
Note: Disaggregated inference configurations (e.g., MoRI SGLang, Dynamo TRTLLM) calculate output throughput per decode chip or per prefill chip, rather than per total chip count. This makes direct output throughput comparison with aggregated configs not an apples-to-apples comparison.
Note: Disaggregated inference configurations (e.g., MoRI SGLang, Dynamo TRTLLM) calculate power per decode chip or per prefill chip, rather than per total chip count. This makes direct power comparison with aggregated configs not an apples-to-apples comparison.
Note: Disaggregated inference configurations (e.g., MoRI SGLang, Dynamo TRTLLM) calculate Joules per decode chip or per prefill chip, rather than per total chip count. This makes direct Joules per token comparison with aggregated configs not an apples-to-apples comparison.
Matching measurements exist, but their chip series are hidden.
Shift+Scroll to zoom • Drag to pan • Double-click to reset • Click a point to pin tooltip
Llama 3.3 70B Instruct 8K / 1K Total Tokens per $1 TCO vs. Interactivity