Radar ·
How Blackwell moves matmul accumulation into Tensor Memory
Radar: technical post describing Tensor Memory in Blackwell as a separate 256 KB per-SM space for Tensor Core results, and stating, according to the author, that this reduces matmul issue from 128 threads to one.
According to the post, Tensor Memory in Blackwell is a separate 256 KB per-SM space for Tensor Core results. The author states that this organization reduces matmul issue from 128 threads to one.
The material shows how Tensor Memory changes accumulator storage and the thread model used for matmul on the GPU. As a technical bookmark, it is worth checking the original to confirm the architecture description, context and the author's methodology before citing any figures.
Original source: Como Blackwell move a acumulação de matmul para Tensor Memory