Loading activity…
@stasbekman
Toolmaker. Software creator, optimizer and harmonizer. Makes ML systems work and fly @ Snowflake.
AI
A user says unidirectional NVLink 6 bandwidth per GPU fell from 1.8 to 1.5 TB/s for upcoming Rubin products.
AI
A user recommends Tim Dettmers’ blog as useful advice for undergrads and grad students.
AI
A user argues Intel is no longer trying to compete head-on with AMD and Nvidia as it did with Gaudi.
AI
A user says the coding agents they tried kept dropping tasks while testing dozens of configurations with slow tests.
AI
A user also says NVFP4 offers higher accuracy and recommends it over MXFP4 for Blackwell and later hardware.
AI
Its creator says an agent can assess single- or multi-node clusters and deliver a detailed report in a few hours.
AI
A user says CPUs, once used only for DataLoader work, are needed again with RAG and RL tool calling. They ask what other unexpected CPU uses people have seen in ML training.
AI
Research engineer Stas Bekman flags NCCL2 and fault tolerance updates in the new release.
AI
Feature reduces GPU memory use for long sequences or larger models.
