Passive Oculink setup unintentionally creates CUDA task scheduler
TinFoilHat_69 · reddit · 2026-08-29
While optimizing a passive Oculink x4 setup on an AM4 platform for Qwen inference, the author created a task scheduler using NCCL to SHM hooks and shared memory locking. It resolves PCIe Gen 3 bottlenecks by keeping memory addresses stable.
More from coding & agent
- Databricks Launches GLM 5.3 with Agent Routing via Unity Gateway — pwendell · 2026-08-29
- Single vs Multi-Agent Debate: Is Simplicity Better for Personal Assistants? — manosaie · 2026-08-29
- MIT Professor Replaces Self with Codex for Website Maintenance — soumitrashukla9 · 2026-08-29
- Three architectures for agentic search: where to put the intelligence — hugobowne · 2026-08-29
- Healthy Aging Atlas MCP server released — modelcontextprotocol · 2026-08-29
- MCP server for Atlassian and Bitbucket released — modelcontextprotocol · 2026-08-29