MI355X (ROCm, SGLang, FP8) cost per million tokens vs B200, GLM-5 on one node, best point, May 2026
~29% lower (B200 1.41x)observed
| Value kind | observed — Reported measurements, counts, and specifications keep the precision and scope stated by their source; an exact specification is not treated as a range. |
|---|---|
| Scope | GLM-5 FP8, ISL 8192/OSL 1024, single node, SGLang v0.5.12 images on both chips. Best point is 18 tok/s/user with MTP, where SemiAnalysis puts B200 at 1.41x MI355X's cost per million tokens, so MI355X is about 29% lower ($0.30 vs $0.22 per million tokens, rounded). The advantage holds from about 10 to 77 tok/s/user; B200 is ahead at about 90. The article headline's '40% cheaper' is the 1.41x ratio, not a 40% saving. |
| Caveat | One model and one workload, not a general ROCm-vs-CUDA result. On SemiAnalysis's October 2026 dashboard MI355X costs 51% more on DeepSeek V4 Pro and 58% less on Qwen3.5 397B (best published run per chip; configurations differ). |
| As of | 2026-05 |
| Source | SemiAnalysis InferenceX · GLM-5 FP8, ISL 8192/OSL 1024, single node: best point 18 tok/s/user with MTP, B200 $0.30/M vs MI355X $0.22/M (1.41x) |
| Review | checking…review by 2027-01-31 · fast cadence |
| Recorded changes | last 2026-10-19 · 3 revisions tracked |
| Claim id | mi355x-rocm-sglang-fp8-cost-per-mtoken |
Where the guide uses it
← Full numbers register — every date-stamped figure in the guide, with revision history.