South Korea’s top web portal giant Naver Corp. has replaced the main chip supplier of its artificial intelligence server for its map service, Naver Place
Classification for fake data entries is a pretty lightweight task. My guess is Naver is experimenting with classical ML modeling （not deep neural nets) that is less dependent on GPU hardware acceleration and sometimes more performant on CPU.
Yea H100 is not price efficient for inference at all. This isn’t surprising.
Seemed to be more a problem with supply than anything else.
Classification for fake data entries is a pretty lightweight task. My guess is Naver is experimenting with classical ML modeling （not deep neural nets) that is less dependent on GPU hardware acceleration and sometimes more performant on CPU.