ModelTC / NNLQP

Apache License 2.0
33 stars 3 forks source link

Some wrong hardware TOPs information #3

Open wwcai-intellif opened 2 years ago

wwcai-intellif commented 2 years ago

Hi, the README says the processing power of Huawei Atlas300 is "INT8: 22 TOPS, FP: 11 TFLOPS" and “atlas300 is faster than mlu270 under the same setting” .

I checked in Atlas product webside, and found that the processing power of Huawei Atlas300 is 88TOPs Int8, not 22TOPs Int8, which may be the information of Atlas200.

The reference links are as follow: https://e.huawei.com/en/products/cloud-computing-dc/atlas/atlas-300-ai https://e.huawei.com/en/products/cloud-computing-dc/atlas/atlas-200-ai

Forwil commented 2 years ago

Hi sidedish:

atlas300 = 4 x ascend310 core,which means every model can use one ascend310 core to run. so we use 22Tops for measure.

wwcai-intellif commented 2 years ago

Thank for your reply!Besides,I‘m a little confused about “atlas300(22T int8) is faster than mlu270(128T int8) under the same setting”. Is mlu270 so inefficient?

Forwil commented 2 years ago

Hi sidedish:

mlu270 test result have some compile-option issue, we have NOT TUNE ON op fusion option, so it seem very slow. we will update the data soon.