Computer Hardware3 min
AMD says new software makes its MI355X AI chip write text up to 3.3x faster
AMD is open-sourcing tuned ways to run three AI models, Qwen3.8-27B, Qwen3.8-Flash-Next and GPT-OSS-120B, on its Instinct MI355X data-centre GPU. It combines several techniques: running a model's whole step as one long-lived GPU program, letting a small model guess tokens ahead, lower-precision numbers, and keeping data next to each of the chip's two dies. AMD reports about 1,100, 750 and 1,470 tokens per second for one request at a time, 1.9 to 3.3 times faster than its comparison setups.
Hype check: These are AMD's own numbers for one request at a time, and the comparison setups differ in precision and drafter models. With the same drafter on both sides, the GPT-OSS-120B gain falls to 1.6x.