Head-to-head AI inference benchmark comparison of GB300 NVL72 (NVIDIA Blackwell) and MI355X (AMD CDNA 4) on Kimi K2.5/K2.6/K2.7-Code 1T. 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.
GB300 NVL72 posts 13066 tok/s/chip for $0.05 per million tokens at 49 tok/s/user on Kimi K2.5/K2.6/K2.7-Code 1T; MI355X posts 3198 tok/s/chip for $0.13. GB300 NVL72 is 165% cheaper per token; GB300 NVL72 delivers 309% more tok/s/chip.
Throughput at 73 tok/s/user on Kimi K2.5/K2.6/K2.7-Code 1T: GB300 NVL72 hits 7869 tok/s/chip, MI355X hits 2298. Per-million costs land at $0.08 and $0.18 respectively. GB300 NVL72 is 122% cheaper per token; GB300 NVL72 delivers 242% more tok/s/chip.
GB300 NVL72 / MI355X on Kimi K2.5/K2.6/K2.7-Code 1T at 98 tok/s/user: 1812 / 1648 tok/s/chip, $0.35 / $0.25 per million tokens. MI355X is 40% cheaper per token; GB300 NVL72 delivers 10% more tok/s/chip. (Numbers reflect the default 8k/1k · fp4 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | GB300 NVL72:13066.3MI355X:3197.5 | GB300 NVL72:7868.9MI355X:2298.5 | GB300 NVL72:1812.0MI355X:1647.5 |
| Cost ($/M tok) | GB300 NVL72:$0.049MI355X:$0.130 | GB300 NVL72:$0.082MI355X:$0.181 | GB300 NVL72:$0.354MI355X:$0.253 |
| tok/s/MW | GB300 NVL72:6163326MI355X:1529923 | GB300 NVL72:3711767MI355X:1099740 | GB300 NVL72:854718MI355X:788278 |
| Concurrency | GB300 NVL72:~2151MI355X:~30 | GB300 NVL72:~700MI355X:~15 | GB300 NVL72:~76MI355X:~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
1 The ATOM engine is promising, however it has yet to serve production tokens. It is still in its infant stage.
Kimi K2.5/2.6/2.7-Code 1T 8K / 1K Total Tokens per $1 TCO vs. Interactivity