has been fully migrated to Huawei's Ascend 950PR, with its underlying code rewritten from CUDA to Huawei's self-develope...

has been fully migrated to Huawei's Ascend 950PR, with its underlying code rewritten from CUDA to Huawei's self-developed CANN framework, scheduled for release in late April. The operating logic of Nvidia's ecosystem is very simple: as long as a new model runs smoothly on CUDA first, even if Huawei, Google TPUs, and AMD MI300 have decent single-card performance on paper, they will only be "alternative solutions" for customers because migration means rewriting operators, redoing #AI

Read Original

Related