The result, in brief
At AIGCode, I led our training optimization for a customer MoE pre-training workload on Ascend NPUs. Our team reached 65% Model FLOPs Utilization. 36Kr and QbitAI reported the result in their coverage of the Ascend ecosystem.
Behind the number
How we improved training
-
Less time waiting
Coordinate communication and computation to reduce waiting across distributed MoE training. Synchronization, scheduling, and uneven expert workloads all affect efficiency.
-
Faster kernel execution
Improve low-level execution through kernel tuning and fusion, drawing on the Ascend CANN software stack and close collaboration across the hardware and software ecosystem.
Further reading
In the press
These three reports cover AIGCode and the Ascend ecosystem, including comments from co-founder and CTO Qiuwu Chen at a developer roundtable. English titles accompany the Chinese originals below.
-
The geeks committed to Kunpeng and Ascend have big plans
-
From infancy to adolescence: China's computing ecosystem through four developers' eyes
-
Behind DeepSeek V4's joint chip and model design: China's computing ecosystem gains momentum