AI&Chips
346 posts

AI&Chips
@ai_chip_expert
Semiconductor (15+ Years), AI Service (7+ Years)








👀 Is #NVIDIA facing another shake-up? After China’s #DeepSeek made waves with its efficient AI model, #Huawei’s Ascend 910C is now in the spotlight, showing impressive inference power! 💡More: buff.ly/4huwfTx 🔗










DeepSeek could spark the golden era for Chinese chips: - DeepSeek V3 supports inference on Huawei Ascend chips from Day 1 - The Huawei 910C (competitor to Nvidia's H100), can do both training and inference - Nvidia's key moat is CUDA (software + ecosystem), Huawei maintains its own pytorch repo, which allows one line import to port CUDA to CUNN (its own CUDA). - Inference performance on Huawei 910C achieves 60% of the H100's performance from developers experience. With hand-written CUNN kernels and optimizations, the performance is higher. My prediction: - As AI model architectures converge to the Transformer, the importance of CUDA and PyTorch compilers diminishes since the engineers can handwrite the kernels in CUNN to highly optimize the performance. - With DeepSeek's cracked team working on Huawei chips, they could significantly reduce dependency on Nvidia, reducing costs by a lot. It's a choice they have to make since they never know when the US will have more GPU export restrictions. - Training remains a more challenging area where Nvidia maintains a strong lead, as the stability of long-term training seems to be a major hurdle for Chinese chips.

추정하기론 중국계(대만인일 가능성 또는 중국계 미국인 가능성) 인물이 한국인을 사칭하며 만든 가짜. 김서연을 KIM-SEO-YUEN으로 적는 한국인은 없을 것이다. (일반적으로 YEON)

딥시크의 한계




