GPU pooling: from time sharing to elastic scaling

2026-05-13 · Su Yanqing · General Manager, YF Cloud

The resource pool schedules by GPU memory, resident models and concurrency slots, tracking queue time as a separate request metric.

Implementation and exception handling

Small batches first verify available GPU memory. Scaling reviews monitor cold starts and model cache hits.

Operations records

Design record YF-CLOUD-RFC-202604. The regular maintenance window is Wednesday, 22:00–23:00. Changes involving service interruptions are announced separately on the status page.

Related resources:Integration and operations documentation · Customer workspace · Service status