All model and GPU pairings
gpt-oss 120BAMD CDNA 3

Running gpt-oss-120b on MI300X

Quick answer

gpt-oss-120b sustains 8,067 tokens/s per GPU on MI300X at 50 tokens/s per user, which works out to $0.033 per million tokens at hyperscaler pricing, served by vLLM. Fastest measured TTFT: 0.0 ms; fastest TPOT: 0.0 ms (each the best across all configs, not one run).

Benchmarked configs

52

Serving engines

vLLM

Precisions

fp4

Run dates

2026-03-092026-04-21

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 targetTokens/s per GPU$ / 1M tokensEnginePrecision
30 tok/s10,570$0.025vLLMfp4
50 tok/s8,067$0.033vLLMfp4
75 tok/s5,938$0.044vLLMfp4
100 tok/s4,659$0.057vLLMfp4
150 tok/s3,190$0.083vLLMfp4
200 tok/s1,022$0.26vLLMfp4

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$0.033
Neocloud$1.16$0.040
Retail$1.30$0.045

Frequently asked questions

How fast is gpt-oss-120b on MI300X?
At an interactivity target of 50 tokens/s per user on a single-turn chat workload (8k input / 1k output), MI300X sustains 8,067 tokens/s per GPU serving gpt-oss-120b with vLLM in FP4. Peak measured throughput across all configs is 11,449 tokens/s per GPU.
How much does it cost to serve gpt-oss-120b on MI300X?
$0.033 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 gpt-oss-120b on MI300X?
The runs behind this page used vLLM in FP4. Engines are rebuilt and re-benchmarked continuously, so the best config can change between visits.
How are these gpt-oss-120b 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 2026-04-21. The same derivation powers the InferenceX overview leaderboard.

Explore the data