Qwen 3.5 397B-A17B — H100 vs MI300X Performance per Dollar
Cost per million tokens of H100 (NVIDIA Hopper) versus MI300X (AMD CDNA 3) on Qwen 3.5 397B-A17B. Owning-hyperscaler TCO normalized by output tokens — performance per dollar across LLM workloads. Pick the more cost-efficient SKU at every target interactivity level. Use the chart controls below to switch sequences, precisions, and metrics — same interactions as the main inference chart.
Push Qwen 3.5 397B-A17B to 43 tok/s/user and H100 lands at $0.31 per million tokens against MI300X's $0.76 — H100 pulls ahead by 143%.
H100: $0.42 per million tokens. MI300X: $1.39. Both at 53 tok/s/user on Qwen 3.5 397B-A17B, with H100 232% cheaper.
Toward the upper edge of the 34–71 tok/s/user interactivity band — at 62 tok/s/user — H100 runs $0.73 per million tokens on Qwen 3.5 397B-A17B while MI300X runs $2.15. H100 is the cheaper choice by 195%. (Numbers reflect the default 1k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
Chip pricing (owning hyperscaler): H100 $1.17/chip/hr · MI300X $0.95/chip/hr. Source: SemiAnalysis Market July 2026 Pricing Surveys & AI Cloud TCO Model.
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Dollar per Million Tokens | H100:$0.315MI300X:$0.763 | H100:$0.419MI300X:$1.392 | H100:$0.728MI300X:$2.150 |
| Concurrency | H100:~106MI300X:~34 | H100:~67MI300X:~15 | H100:~29MI300X:~8 |
Cost per Million Total Tokens (Owning - Hyperscaler) vs. Interactivity
Qwen3.5 397B • FP4 • 8K / 1K • Source: SemiAnalysis InferenceX™
Qwen3.5 397B • FP4 • 8K / 1K • Source: SemiAnalysis InferenceX™TCO $/chip/hr: VR200: 3.61H100: 1.17H200: 1.22B200: 1.73B300: 2.26GB200: 1.86GB300: 2.31MI300X: 0.95MI325X: 1.1MI355X: 1.5RTX6000PRO: 0.68Jalapeño (Teacup): 1.47
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.
No data available
Please change the model, sequence, precision, date range or chip selection.
Shift+Scroll to zoom • Drag to pan • Double-click to reset • Click a point to pin tooltip
Interactivity is the rate at which a single user receives generated tokens while the model streams its answer — how quickly new words appear on screen. Higher values feel snappier; operators trade it against batch throughput.
Estimated infrastructure cost of producing one million total tokens (input + output) at this operating point, priced with the all-in hourly ownership cost of a hyperscaler operator. Lower is cheaper.
Formula: $/Mtok = all-in cost per chip-hour ($) × 1,000,000 ÷ (3,600 × total tok/s/chip)