AI2 revamps GPU scheduling to boost research impact
The Allen Institute for AI has overhauled how it assigns GPU resources. Previously, researchers used a priority system prone to users gaming the rules and excessive negotiation by IT. Now, GPU time is distributed using explicit time budgets, a tiered fair-share approach, and time-slicing, making the process more transparent and less manual. The aim: high-value research gets better access while keeping GPUs working at full capacity.
- AI2 manages thousands of NVIDIA H100, B200, and B300 GPUs
- Former system led to users monopolizing or mislabeling workloads
- IT staff spent significant time managing scheduling disputes
- New system uses time budgets and fair-share allocation
- Goal is to prioritize impactful research and maximize GPU use
Sources covering this
More in Enterprise
California launches behavioral health data dashboard with Databricks
California has rolled out a public dashboard that centralizes behavioral health data from all 58 counties, built on Databricks.
Alteryx integrates Live Query with Google BigQuery
Alteryx has rolled out its Live Query feature for Google Cloud’s BigQuery, letting companies process and analyze unstructured data like…
Kubernetes containers drop cgroup v1 as edge AI grows
Kubernetes is phasing out support for cgroup v1, an older Linux kernel interface, as part of a shift toward handling more AI workloads…
TextExpander now offers a free plan with core features
TextExpander, the text shortcut tool, just introduced a free plan.