Tom Forsyth on Nostr: nprofile1q…a90j6 Aside from all the LRB-specific stuff, having done both CPU and ...
nprofile1qy2hwumn8ghj7un9d3shjtnddaehgu3wwp6kyqpqdymfj9a0sjmnh7e93fs7mxklyj0v7g7mlgj8tc7ef9jllrkpzypqwa90j6 (nprofile…90j6) Aside from all the LRB-specific stuff, having done both CPU and GPU sides, I completely agree that the hardest part of all of this is managing queues.
This seems to be universal. It was the hardest part of Cell, it was the hardest part of LRB, it was the hardest part of working on Gen, and your experiences match this.
Yes, if you blow your queues out on a CPU, the code still runs (unlike a GPU). But you thrash the caches and perf is still a disaster.
This seems to be universal. It was the hardest part of Cell, it was the hardest part of LRB, it was the hardest part of working on Gen, and your experiences match this.
Yes, if you blow your queues out on a CPU, the code still runs (unlike a GPU). But you thrash the caches and perf is still a disaster.