Head-to-head AI inference benchmark comparison of GB300 NVL72 (NVIDIA Blackwell) and MI325X (AMD CDNA 3) on DeepSeekv4 Pro 0813 1.6T. 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 →
At 72 tok/s/user on DeepSeekv4 Pro 0813 1.6T, GB300 NVL72 delivers 63263 tok/s/chip at $0.01 per million tokens; MI325X hasn't been benchmarked at this target.
GB300 NVL72 hits 6382 tok/s/chip for $0.10 per million tokens at 114 tok/s/user on DeepSeekv4 Pro 0813 1.6T. No MI325X data at this operating point.
GB300 NVL72: 3488 tok/s/chip, $0.18 per million tokens at 156 tok/s/user on DeepSeekv4 Pro 0813 1.6T. MI325X is unmeasured here. (Numbers reflect the default agentic-traces · fp4 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | GB300 NVL72:63263.2MI325X:— | GB300 NVL72:6382.4MI325X:— | GB300 NVL72:3488.1MI325X:— |
| Cost ($/M tok) | GB300 NVL72:$0.010MI325X:— | GB300 NVL72:$0.101MI325X:— | GB300 NVL72:$0.184MI325X:— |
| tok/s/MW | GB300 NVL72:29841134MI325X:— | GB300 NVL72:3010585MI325X:— | GB300 NVL72:1645309MI325X:— |
| Concurrency | GB300 NVL72:~619MI325X:— | GB300 NVL72:~4MI325X:— | GB300 NVL72:~1MI325X:— |
DeepSeek V4 Pro 0813 1.6T • FP4 • 8K / 1K • Source: SemiAnalysis InferenceX™ • Input $1/M tok · Output $1/M tok
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.
Gross token revenue a GPU could earn per hour at this operating point. The normalized source prices uncached input and output at $1 per million and cached input at $0.10 per million. The OpenRouter source uses the selected model’s current public prices, with a 10%-of-input fallback when no cache-read price is published. Revenue prices uncached input, cached input, and output separately. Input share comes from compatible measured input/output throughput. When disaggregated rates use different GPU denominators, fixed sequences use ISL:OSL and Agentic traces use measured prompt/generation tokens. Agentic cache hit combines GPU and external cache when external cache is reported, otherwise GPU and CPU cache. Historical Trends interpolates total throughput, input share, and cache hit separately before pricing. A partially measured cache frontier receives no cache discount. This turns the throughput/interactivity tradeoff into a business-facing SLA curve.
Formula: $/GPU/hr = (uncached input × input price + cached input × cache-read price + output × output price) per GPU-second, scaled to one hour
DeepSeek V4 Pro 0813 1.6T • FP4 • 8K / 1K • Source: SemiAnalysis InferenceX™ •