NVIDIA describes Topograph, an approach to topology-aware GPU workload placement in AI factories. Poor placement fragments topology domains and pushes traffic over shared links, cutting throughput and raising job costs while GPUs idle on data. The tool targets more efficient use of power-limited AI factory infrastructure.
NVIDIA describes DSX MaxLPS, a policy-governed power-sharing approach that allocates power dynamically across AI factory resources. AI factories are typically provisioned for the rare case when every GPU peaks simultaneously, leaving capacity underused. The post claims customers can deploy more compute within the same power envelope.
NVIDIA describes DSX MaxLPS, a policy-governed power-sharing approach that allocates power dynamically across AI factory resources. AI factories are typically provisioned for the rare case when every GPU peaks simultaneously, leaving capacity underused. The post claims customers can deploy more compute within the same power envelope.
A Medium post lays out how to plan GPU memory for AI inference in DePIN contexts. It frames memory budgeting as the sum of model weights, the key-value cache, runtime allocations, and operating headroom.
A Medium post lays out how to plan GPU memory for AI inference in DePIN contexts. It frames memory budgeting as the sum of model weights, the key-value cache, runtime allocations, and operating headroom.
A team built a constraint-aware GPU allocator and benchmarked it against a FIFO scheduler across seven scenarios. On identical hardware and workloads, GPU utilization rose by up to 33 percentage points and priority-weighted output by up to 105%, with no hardware changes—only the order of allocation decisions.
A team built a constraint-aware GPU allocator and benchmarked it against a FIFO scheduler across seven scenarios. On identical hardware and workloads, GPU utilization rose by up to 33 percentage points and priority-weighted output by up to 105%, with no hardware changes—only the order of allocation decisions.
Google Cloud's Dynamic Workload Scheduler (DWS) is introduced as a solution to reduce GPU wait times on GKE. The guide explains how DWS helps schedule GPU workloads more efficiently, addressing the common problem of GPU scarcity.
Google Cloud's Dynamic Workload Scheduler (DWS) is introduced as a solution to reduce GPU wait times on GKE. The guide explains how DWS helps schedule GPU workloads more efficiently, addressing the common problem of GPU scarcity.
Ai2's AI Infrastructure team replaced its priority-based GPU cluster scheduler with a system combining GPU time budgets, hierarchical fair-share allocation, and a time-slicing contract. The change targets thousands of NVIDIA H100, B200, and B300 GPUs serving ~150 researchers with 2-3x oversubscribed demand, and shifted allocation debates from ad-hoc operations to administrative budgeting.
Ai2's AI Infrastructure team replaced its priority-based GPU cluster scheduler with a system combining GPU time budgets, hierarchical fair-share allocation, and a time-slicing contract. The change targets thousands of NVIDIA H100, B200, and B300 GPUs serving ~150 researchers with 2-3x oversubscribed demand, and shifted allocation debates from ad-hoc operations to administrative budgeting.