GLM 5.3 744B — B300 vs H200
Head-to-head AI inference benchmark comparison of B300 (NVIDIA Blackwell) and H200 (NVIDIA Hopper) on GLM 5.3 744B. 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.
AgentX replays real coding-agent sessions rather than fixed-length prompts, so context grows turn over turn and most of each request is served from cache instead of being recomputed. That turns the comparison into a systems question: KV transfer between nodes, prefix-aware routing, and cache capacity all move the curve alongside raw chip throughput. Fixed-sequence workloads stay the clean baseline for kernel and silicon performance, so the two scenarios answer different questions about the same hardware. Learn more about AgentX →
B300 hits 7311 tok/s/chip for $0.09 per million tokens at 171 tok/s/user on GLM 5.3 744B. No H200 data at this operating point.
B300: 5171 tok/s/chip, $0.12 per million tokens at 217 tok/s/user on GLM 5.3 744B. H200 is unmeasured here.
At 262 tok/s/user on GLM 5.3 744B, B300 delivers 4002 tok/s/chip at $0.16 per million tokens; H200 hasn't been benchmarked at this target. (Numbers reflect the default agentic-traces · fp4 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | B300:7310.5H200:— | B300:5170.9H200:— | B300:4002.3H200:— |
| Cost ($/M tok) | B300:$0.086H200:— | B300:$0.121H200:— | B300:$0.157H200:— |
| tok/s/MW | B300:3847647H200:— | B300:2721500H200:— | B300:2106449H200:— |
| Concurrency | B300:~9H200:— | B300:~6H200:— | B300:~4H200:— |