Running Qwen3.5 on MI300X
Quick answer
Qwen3.5 sustains 583 tokens/s per GPU on MI300X at 50 tokens/s per user, which works out to $0.45 per million tokens at hyperscaler pricing, served by SGLang. Fastest measured TTFT: 0.2 ms; fastest TPOT: 0.0 ms (each the best across all configs, not one run).
Benchmarked configs
40
Serving engines
SGLang
Precisions
bf16, fp8
Run dates
2026-03-02 → 2026-08-12
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 | 1,208 | $0.22 | SGLang | fp8 |
| 50 tok/s | 583 | $0.45 | SGLang | fp8 |
| 75 tok/s | - | - | SGLang | fp8 |
| 100 tok/s | - | - | SGLang | fp8 |
| 150 tok/s | - | - | SGLang | fp8 |
| 200 tok/s | - | - | SGLang | fp8 |
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 | $0.95 | $0.45 |
| Neocloud | $1.16 | $0.55 |
| Retail | $1.30 | $0.62 |
Frequently asked questions
- How fast is Qwen3.5 on MI300X?
- At an interactivity target of 50 tokens/s per user on a single-turn chat workload (8k input / 1k output), MI300X sustains 583 tokens/s per GPU serving Qwen3.5 with SGLang in FP8. Peak measured throughput across all configs is 3,331 tokens/s per GPU.
- How much does it cost to serve Qwen3.5 on MI300X?
- $0.45 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 Qwen3.5 on MI300X?
- The runs behind this page used SGLang in BF16, FP8. Engines are rebuilt and re-benchmarked continuously, so the best config can change between visits.
- How are these Qwen3.5 numbers measured?
- Every number is measured on real MI300X 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 2026-08-12. The same derivation powers the InferenceX overview leaderboard.