InferenceX Ranks GB200 First for MiniMax M3, While MI355X Has Lower Token Cost

Infrastructure teams serving MiniMax M3 now face a clean split in InferenceX’s latest ranking: GB200 NVL72 delivers more throughput, while MI355X carries the lower reported price per million tokens. The comparison holds both systems to the same 50-tokens-per-second-per-user target on an agentic coding workload, making the choice depend on whether capacity or token cost is the immediate constraint.

InferenceX ranks GB200 NVL72 first at 33,210 tokens per second per GPU. MI355X is second at 29,616 tokens per second per GPU, a 12% gap. The newest benchmark result feeding the live page landed on September 4.

The comparison

The headline result is not a maximum-throughput contest at arbitrary latency. InferenceX says it sweeps concurrency on real hardware to map each platform’s throughput-versus-interactivity frontier, then reads every platform at the same user-speed point. Systems without a measurement at that point are excluded from the ranking.

That makes the $0.016-per-million-token figure for GB200 and $0.014 for MI355X consequential rather than incidental. InferenceX derives those figures from measured throughput and GPU-hour rates in SemiAnalysis’s AI Cloud total-cost-of-ownership model; they are reported costs, not a list price for either system.

Software is part of the result

The table should not be read as a hardware-only verdict. Both entries use FP4 precision, but GB200 NVL72 runs Dynamo vLLM while MI355X runs vLLM. The live ranking also reruns as engine releases and configurations change, so its ordering is a current operating result rather than a permanent specification-sheet comparison.

An earlier InferenceX analysis of MiniMax M3 illustrates why that distinction matters. It found B200 and B300 ahead of their rack-scale GB200 NVL72 and GB300 NVL72 counterparts on total-cost-of-ownership-normalized throughput. It attributed the rack-scale disadvantage in part to a Dynamo router whose work grows with the number and length of live prefixes, rather than tokens generated, alongside a lack of well-tuned kernels for wide expert, context and tensor parallelism on this model.

The limits of the snapshot

MiniMax M3’s long contexts expose limitations that a fixed prompt test can miss. The earlier analysis said no submission used context parallelism despite a p90 input sequence length of 317,000 tokens. It said the model’s four KV heads cap distributed context parallelism at two even with tensor parallelism of eight, while separate handling for the model’s MSA indexer remained under development in vLLM.

The same analysis identified a separate AMD constraint: none of its Pareto-optimal points used KV-cache offload to DRAM, while every NVIDIA Pareto-optimal point above concurrency 20 did. It traced that difference to inefficient GPU-to-CPU transfers in AMD vLLM before ROCm 7.14 added a batched-copy API. Those implementation details help explain why a speed-and-cost decision on agent traffic cannot be reduced to a chip name alone.

For now, the ranking answers a narrower procurement question: which measured configuration supplies the most MiniMax M3 output at a fixed interactive rate, and which does so at the lower reported token cost. New engine releases, configurations and measurements can alter that answer as the live page updates.