DeepSight Releases Ascend 950-Optimized Compute and Communication Libraries

DeepSight released its Ascend 950-optimized computation and communication libraries on September 30. The libraries include DeepGEMM-Ascend for matrix operations and DeepEP-Ascend for distributed communication in machine learning training and inference.
According to DeepSight's official GitHub repository, DeepGEMM-Ascend supports the Ascend 950 while maintaining API compatibility with the existing DeepGEMM. DeepEP-Ascend aligns its public buffer API with NVIDIA's DeepEP version for the Huawei Ascend NPU.
The DeepEP-Ascend repository reports bandwidth measurements of 373–375GB/s for dispatch and 345–347GB/s for combine, measured on the Ascend 950DT using CANN 9.2.0 and test hardware configuration.
Korean Source
This article is an English localization of a Korean-language crypto news report. Original headline: 딥시크, 화웨이 어센드 950용 연산·통신 라이브러리 공개