Nota reports 52% faster AI inference with GPU-NPU split on AMD chip
Nota says splitting prefill to the NPU and decode to the GPU also cut energy per token by 31% in tests on the Qwen3-8B model.
Nota presented technology that splits AI inference work between the GPU and NPU on AMD's Ryzen AI Max+ 395 processor at AMD Advancing AI World Tour Seoul 2026, the company said on November 9, 2026. In its own tests, the split Prefill-on-NPU and Decode-on-GPU setup cut total answer completion time by up to 52% compared with AMD's existing GPU-NPU hybrid execution, according to datanet.co.kr.
Nota said per-token energy consumption dropped 31% and per-token generation time fell 41-46% in those tests, which used the Qwen3-8B model.
The company also reported results from applying an AI agent based optimization process to its NetsPresso optimization platform.
On AMD Ryzen AI environments running multiple language models, Nota reported 11-32% lower memory use than AMD's hybrid execution and 2.5-3.8x faster answer generation than NPU-only execution.
Earlier partnerships
In July, Nota joined AMD's Robotics Partner Network as one of key initial partners in the Asia-Pacific and Japan region. In October, the company showed NetsPresso-based optimization technology on Arm's AI developer platform. It has also worked with semiconductor companies including Samsung Electronics, Arm, FuriosaAI and Mobilint.
Quick answers
What results did Nota report for the GPU-NPU split on AMD Ryzen AI Max+ 395?
Nota said the Prefill-on-NPU and Decode-on-GPU setup cut total answer completion time by up to 52% versus AMD's existing hybrid execution, with 31% lower energy per token and 41-46% faster per-token generation, in tests on Qwen3-8B.
Where did Nota present the results?
At AMD Advancing AI World Tour Seoul 2026.