Running GLM-5 on GB200 NVL72
Quick answer
GLM-5 sustains 1,553 tokens/s per GPU on GB200 NVL72 at 50 tokens/s per user, which works out to $0.33 per million tokens at hyperscaler pricing, served by Dynamo SGLang. Fastest measured TTFT: 0.4 ms; fastest TPOT: 0.0 ms (each the best across all configs, not one run).
Benchmarked configs
102
Serving engines
Dynamo SGLang, Dynamo TRTLLM
Precisions
fp4, fp8
Run dates
2026-07-08 → 2026-07-16
Throughput at every interactivity target
Serving is a trade-off: push more concurrent users through a GPU and each user's tokens arrive slower. The ladder below reads the measured frontier at each per-user speed target on a single-turn chat workload (8k input / 1k output), using the best engine and precision at that point.
| Per-user target | Tokens/s per GPU | $ / 1M tokens | Engine | Precision |
|---|---|---|---|---|
| 30 tok/s | - | - | Dynamo SGLang | fp4 |
| 50 tok/s | 1,553 | $0.33 | Dynamo SGLang | fp4 |
| 75 tok/s | 2,865 | $0.18 | Dynamo SGLang | fp4 |
| 100 tok/s | 1,411 | $0.37 | Dynamo SGLang | fp4 |
| 150 tok/s | 339 | $1.52 | Dynamo SGLang | fp4 |
| 200 tok/s | - | - | Dynamo SGLang | fp4 |
What serving actually costs
Converting the 50 tokens/s per user operating point to $ per million total tokens across rental pricing tiers from the SemiAnalysis AI Cloud TCO model.
| Pricing tier | $ / GPU / hr | $ / 1M tokens |
|---|---|---|
| Hyperscaler | $1.86 | $0.33 |
| Neocloud | $2.26 | $0.40 |
| Retail | $2.60 | $0.47 |
Frequently asked questions
- How fast is GLM-5 on GB200 NVL72?
- At an interactivity target of 50 tokens/s per user on a single-turn chat workload (8k input / 1k output), GB200 NVL72 sustains 1,553 tokens/s per GPU serving GLM-5 with Dynamo SGLang in FP4. Peak measured throughput across all configs is 10,949 tokens/s per GPU.
- How much does it cost to serve GLM-5 on GB200 NVL72?
- $0.33 per million total tokens at hyperscaler $/GPU/hr pricing, at 50 tokens/s per user. Neocloud and retail rental tiers are tabulated above; slower interactivity targets lower the cost further.
- Which serving engines run GLM-5 on GB200 NVL72?
- The runs behind this page used Dynamo SGLang, Dynamo TRTLLM in FP4, FP8, including disaggregated prefill and multi-node serving. Engines are rebuilt and re-benchmarked continuously, so the best config can change between visits.
- How are these GLM-5 numbers measured?
- Every number is measured on real GB200 NVL72 hardware by the InferenceX fleet, sweeping concurrency on a single-turn chat workload (8k input / 1k output) to trace the throughput-versus-interactivity frontier; the newest run landed on 2026-07-16. The same derivation powers the InferenceX overview leaderboard.