Qwen 3.5 397B-A17B · Chip comparison

Qwen 3.5 397B-A17B — B300 vs GB300 NVL72

Head-to-head AI inference benchmark comparison of B300 (NVIDIA Blackwell) and GB300 NVL72 (NVIDIA Blackwell) on Qwen 3.5 397B-A17B. 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 / GB300 NVL72 on Qwen 3.5 397B-A17B at 113 tok/s/user: 63091 / 150385 tok/s/chip, $0.01 / $0.00 per million tokens. GB300 NVL72 is 133% cheaper per token; GB300 NVL72 delivers 138% more tok/s/chip.

Around the middle of the 24–378 tok/s/user interactivity band, at 201 tok/s/user on Qwen 3.5 397B-A17B: B300 runs 33398 tok/s/chip at $0.02/M tokens, GB300 NVL72 runs 115865 at $0.01/M. GB300 NVL72 is 239% cheaper per token; GB300 NVL72 delivers 247% more tok/s/chip.

Setting 290 tok/s/user as the target on Qwen 3.5 397B-A17B, B300 produces 12164 tok/s/chip ($0.05 per million tokens) and GB300 NVL72 produces 75559 ($0.01). GB300 NVL72 is 508% cheaper per token; GB300 NVL72 delivers 521% more tok/s/chip. (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.)

View performance-per-dollar view →

Interpolated from real benchmark data. Edit target interactivity values below to compare at different operating points.

Metric	Interactivity (tok/s/user)	Interactivity (tok/s/user)	Interactivity (tok/s/user)
Throughput (tok/s/chip)	B300:63091.0GB300 NVL72:150384.7	B300:33398.3GB300 NVL72:115864.9	B300:12164.5GB300 NVL72:75558.6
Cost ($/M tok)	B300:$0.010GB300 NVL72:$0.004	B300:$0.019GB300 NVL72:$0.006	B300:$0.052GB300 NVL72:$0.008
tok/s/MW	B300:33205805GB300 NVL72:70936182	B300:17578057GB300 NVL72:54653249	B300:6402346GB300 NVL72:35640866
Concurrency	B300:~21GB300 NVL72:~704	B300:~23GB300 NVL72:~633	B300:~4GB300 NVL72:~218

Inference Performance

Agentic inference metrics from the AgentX scenario and fixed-sequence inference metrics across models, hardware configurations, and serving parameters.

Model

Scenario

Y-Axis Metric

Chip Config

Total Tokens per $1 USD (Owning - Neocloud Giant) vs. Interactivity

Qwen3.5 397B • FP4 • 8K / 1K • Source: SemiAnalysis InferenceX™

TCO $/chip/hr: VR200: 3.61H100: 1.55H200: 1.59B200: 2.07B300: 2.52GB200: 2.26GB300: 2.79MI300X: 1.16MI325X: 1.32MI355X: 2.09RTX6000PRO: 0.75Jalapeño (Teacup): 1.56

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