DeepSeek-Huawei Partnership: DeepSeek and Huawei are creating programming infrastructure tailored specifically for Huawei’s Ascend AI accelerators.
Open-Source AI Tools: To foster broader Ascend development, DeepSeek has made software covering computation, programming, and communication open-source.
TileLang Programming Language: TileLang acts as a high-level programming layer designed to streamline development and assist in optimising workloads for AI chips.
Challenge to Nvidia CUDA: This effort takes direct aim at a primary Nvidia strength: its entrenched CUDA software ecosystem that powers AI development and optimisation.
128-Chip Supernode System: The two companies collaborated to develop a supernode leveraging 128 Ascend 950 chips, thereby enhancing chip-to-chip communication and computation.
Six Software Components: DeepSeek has rolled out six Ascend-tailored components encompassing communication, computation, programming, attention, data-processing workloads, and kernels.
Building Developer Adoption: According to Huawei, the Ascend platform currently supports upwards of 90 prominent open-source projects, such as PyTorch, vLLM, Triton, and veRL.
Limits Still Remain: Huawei continues to encounter component shortages, and DeepSeek relies on Nvidia hardware for certain intensive model-training tasks.
Read More Stories
Join our WhatsApp Channel to get the latest news, exclusives and videos on WhatsApp
