Research on AI Chip Large Model Train Testing Technology
Hongfei Qiu, Pengcheng Jiang, Zhiqin Huang · 2024
Theoretically the performance of AI chips large model training performance mainly depends on the specification parameters of the chip and quantity, in practice, the performance of inference cards is also related to the inference optimization technology of the chip. In response to this issue of AI chips large model training performance, we conducted researches on the specification information of current mainstream chip manufacturers. On the basis of studying existing evaluation indicators, a performance measurement scheme was proposed by using MFU as an important performance indicator of AI chips large model training performance, and five products were tested. The experimental results indicate that this scheme can effectively quantitatively evaluate the training performance of AI chips.