All model and GPU pairings
DeepSeekv4 Pro 0813 1.6TAMD CDNA 4

Running DeepSeek V4 Pro on MI355X

Quick answer

DeepSeek V4 Pro sustains 23,751 tokens/s per GPU on MI355X at 50 tokens/s per user, which works out to $0.018 per million tokens at hyperscaler pricing, served by SGLang. Fastest measured TTFT: 0.1 ms; fastest TPOT: 0.0 ms (each the best across all configs, not one run).

Benchmarked configs

236

Serving engines

ATOM¹, MoRI SGLang, Mooncake ATOMesh¹, SGLang, vLLM

Precisions

fp4, fp8

Run dates

2026-04-252026-09-16

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 the AgentX agentic coding workload, using the best engine and precision at that point.

Per-user targetTokens/s per GPU$ / 1M tokensEnginePrecision
30 tok/s33,667$0.012SGLangfp4
50 tok/s23,751$0.018SGLangfp4
75 tok/s5,565$0.075vLLMfp4
100 tok/s10,635$0.039SGLangfp4
150 tok/s4,898$0.085SGLangfp4
200 tok/s2,020$0.21SGLangfp4

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
Owning at Large Hyperscaler Volume$1.50$0.018
Retail$2.90$0.034

Frequently asked questions

How fast is DeepSeek V4 Pro on MI355X?
At an interactivity target of 50 tokens/s per user on the AgentX agentic coding workload, MI355X sustains 23,751 tokens/s per GPU serving DeepSeek V4 Pro with SGLang in FP4. Peak measured throughput across all configs is 52,737 tokens/s per GPU.
How much does it cost to serve DeepSeek V4 Pro on MI355X?
$0.018 per million total tokens at large-hyperscaler-volume ownership $/GPU/hr pricing, at 50 tokens/s per user. The retail rental tier is tabulated above; slower interactivity targets lower the cost further.
Which serving engines run DeepSeek V4 Pro on MI355X?
The runs behind this page used ATOM¹, MoRI SGLang, Mooncake ATOMesh¹, SGLang, vLLM 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 DeepSeek V4 Pro numbers measured?
Every number is measured on real MI355X hardware by the InferenceX fleet, sweeping concurrency on the AgentX agentic coding workload to trace the throughput-versus-interactivity frontier; the newest run landed on 2026-09-16. The same derivation powers the InferenceX overview leaderboard.

Explore the data