Hacker News new | past | comments | ask | show | jobs | submit login

As people stated in the announcement post, a high memory bandwidth doesn't really benefit CPUs much. That's also why AMD and Intel don't have high memory bandwidth on their cores, because it doesn't really help performance.

Where it is beneficial is the GPU; for comparison AMD and Nvidia cards often exceed 400GB/s. A desktop RTX 3090 has 900GB/s.




Intel is definitely looking at HBM for Sapphire Rapids though.

https://www.anandtech.com/show/16795/intel-to-launch-next-ge...


I don’t think you could even top 150GB/s sustained if you ran memcpy on all ten threads at once. (Though that would be 300 total.)


This turned out to be correct, look at https://www.anandtech.com/show/17024/apple-m1-max-performanc... -- the "sustained" measurement is on the right hand side of the graph. It hits 243GB/s bandwidth with 8+2 threads, 224 with just the 8 performance cores.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: