Huawei reportedly deploying a 256K 950 SuperCluster
A post calls amassing a quarter-million 950PRs in one cluster “pointless” and relays a claim attributed to Wenfeng about the hardware needed to train an 800-billion-active-parameter model.
TLDR
A post claims Huawei is deploying a 256K 950 SuperCluster while questioning the value of gathering a quarter-million 950PRs in one cluster. It also says Wenfeng previously described 200,000 950DTs as the minimum needed to train a model with 800 billion active parameters.
Combined views
4.4K
1 Source, first seen 6h ago
Huawei reportedly deploying a 256K 950 SuperCluster
A post calls amassing a quarter-million 950PRs in one cluster “pointless” and relays a claim attributed to Wenfeng about the hardware needed to train an 800-billion-active-parameter model.
TLDR
A post claims Huawei is deploying a 256K 950 SuperCluster while questioning the value of gathering a quarter-million 950PRs in one cluster. It also says Wenfeng previously described 200,000 950DTs as the minimum needed to train a model with 800 billion active parameters.