Head-to-head AI inference benchmark comparison of GB300 NVL72 (NVIDIA Blackwell) and MI325X (AMD CDNA 3) on DeepSeek R1. 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.
Setting 28 tok/s/user as the target on DeepSeek R1, GB300 NVL72 produces 9105 tok/s/chip ($0.07 per million tokens) and MI325X produces 811 ($0.38). GB300 NVL72 is 435% cheaper per token; GB300 NVL72 delivers 1023% more tok/s/chip.
At 37 tok/s/user interactivity on DeepSeek R1, GB300 NVL72 delivers 8803 tok/s/chip at $0.07 per million tokens; MI325X delivers 646 tok/s/chip at $0.47. GB300 NVL72 is 549% cheaper per token; GB300 NVL72 delivers 1262% more tok/s/chip at this point.
GB300 NVL72 posts 8703 tok/s/chip for $0.07 per million tokens at 45 tok/s/user on DeepSeek R1; MI325X posts 447 tok/s/chip for $0.68. GB300 NVL72 is 826% cheaper per token; GB300 NVL72 delivers 1845% more tok/s/chip. (Numbers reflect the default 8k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | GB300 NVL72:9105.2MI325X:811.0 | GB300 NVL72:8802.7MI325X:646.3 | GB300 NVL72:8703.0MI325X:447.4 |
| Cost ($/M tok) | GB300 NVL72:$0.070MI325X:$0.377 | GB300 NVL72:$0.073MI325X:$0.473 | GB300 NVL72:$0.074MI325X:$0.683 |
| tok/s/MW | GB300 NVL72:4294889MI325X:479902 | GB300 NVL72:4152206MI325X:382442 | GB300 NVL72:4105187MI325X:264725 |
| Concurrency | GB300 NVL72:~1533MI325X:~27 | GB300 NVL72:~1229MI325X:~17 | GB300 NVL72:~1229MI325X:~10 |
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
DeepSeek R1 0528 671B 8K / 1K Total Tokens per $1 TCO vs. Interactivity