AMD released the MI455X accelerator on October 8, 2026, stating it significantly improves AI inference performance compared to its predecessor. It is the company's first major AI inference update since the MI355X launch in 2025.
AMD reported a 34x higher token throughput at high interactivity, measured on the DeepSeek-V4-Flash model. That compares with a 4x increase at low interactivity.
The MI455X is built on the Helios platform and targets high-concurrency AI inference workloads. Availability begins with software enablement, initially for developers and data center operators.
"The results fall into two groups," said Edward Tian, AMD's AI performance lead. "The first is an end-to-end online serving benchmark on DeepSeek-V4-Flash, which is the number most representative of a real deployment."
The announcement follows AMD's participation in Advancing AI 2026, where the company shared performance results comparing the new MI455X against the current-generation MI355X. AMD did not say how the MI455X will perform on other models, and noted the results are point-in-time, as software stacks continue to mature.
Source: amd
