Prime Intellect reports moving 1.6 TB of GLM-5.2 weights in about 9 seconds
The company says the transfer spans hundreds of GPUs and includes engine pause and resume. It puts an experimental path at about 4 seconds per update.
TLDR
Prime Intellect says it used NIXL to cut GLM-5.2’s weight transfer time to about 9 seconds, moving the full 1.6 TB policy between hundreds of GPUs. That timing includes pausing and resuming the engine. The company says an experimental path reduces the time further to about 4 seconds per update, saturating network bandwidth.
Combined views
2.2K
3 Sources, first seen 27d ago