AMD's MI355X recorded a lower reported inference cost than Nvidia's GB200 NVL72 in SemiAnalysis's InferenceX comparison on the Kimi K3 2.8T model.
At the highest interactivity level in the comparison, the MI355X was listed at $0.15 per million tokens, compared with $0.22 for the GB200 NVL72. The InferenceX page describes the AMD system as “46% cheaper.”
The comparison measures cost per million tokens at different operating points, giving readers a way to assess how the two accelerator platforms perform under varying interaction demands.
Cost advantage changes by operating point
The reported difference was smaller at 71 tok/s/user. At that operating point, the MI355X was listed at $0.05 per million tokens, while the GB200 NVL72 was listed at $0.07.
InferenceX describes this result as the “MI355X edges GB200 NVL72” and reports “a 35% cost-per-token gap.” These figures indicate that the relative cost advantage changes depending on the selected performance and interactivity level.
At another listed operating point of 118 tok/s/user, the MI355X was reported at $0.14 per million tokens. The supplied comparison does not provide the corresponding GB200 NVL72 figure for that point, so no direct percentage gap is stated here.
Why the comparison matters
Inference costs are important for operators serving large language models because token pricing is tied to the expense of generating responses. A lower cost per million tokens can be relevant when evaluating hardware for workloads with different response-speed and concurrency requirements.
However, the reported figures apply specifically to the Kimi K3 2.8T comparison and its stated operating points. They should not be treated as a universal ranking across every model, workload or deployment configuration.
The available information identifies SemiAnalysis's InferenceX comparison as the basis for the figures. It does not provide additional details on system configuration, power use, hardware pricing or the testing methodology in the supplied material.
Conclusion
SemiAnalysis's reported InferenceX results show the AMD MI355X costing less per million tokens than Nvidia's GB200 NVL72 on Kimi K3 2.8T at the listed operating points, with the stated advantage varying by interactivity level.
Frequently Asked Questions
Q. Which AMD accelerator was compared?
The comparison evaluated AMD's MI355X.
Q. Which Nvidia system was used for comparison?
The Nvidia system was the GB200 NVL72.
Q. Which model was tested?
The comparison used Kimi K3 2.8T.
Q. What was the reported cost at the highest interactivity level?
The MI355X was listed at $0.15 per million tokens, compared with $0.22 for the GB200 NVL72.
Q. What advantage did the InferenceX page report at the highest interactivity level?
The page described the MI355X as “46% cheaper.”
Q. What were the figures at 71 tok/s/user?
The MI355X was listed at $0.05 per million tokens, compared with $0.07 for the GB200 NVL72.
Q. What was reported at 118 tok/s/user?
The MI355X was reported at $0.14 per million tokens. the corresponding GB200 NVL72 figure is not yet available.
