bibibi.ccLoading…
Connect scattered updates into a story you can understand.
A report citing SemiAnalysis measurements said that Ascend 950PR in 1055 groups of commonly used large-model matrix calculations had an actual speed of approximately H100 of 43%. The report said that DeepSeek is improving efficiency through software adaptation, while large-scale training still faces coordination and stability challenges.
The report states that Perplexity and NVIDIA proposed JOLT, incorporating the teacher branch into training as well, and achieved Qwen3-0.6B accuracy in the MATH test. 72.14%AppWorld results show higher performance after adding the student's final task reward.
According to the report,DeepSeek has made adapting to domestic training chips such as Huawei's an important strategy. A new batch of Huawei chips is expected as early as the 4th quarter, while Nvidia chips are still currently used for training.