ForkUnion v3 Cuts Latency 3–6× vs OpenMP with NUMA-Aware Thread Pool

viglovikov · x · 2026-08-21

Developer released ForkUnion v3, a low-latency NUMA-aware thread-pool for C/C++/Rust/Zig.

• Performance Gains: Benchmarked on Nebius fat NUMA machines, it achieves 3–6× lower latency than OpenMP and 12–16× lower latency than Taskflow and Rayon.
• Architecture: Eliminates heap allocations, mutexes, and CAS atomics from the hot path to resolve bottlenecks when scaling across 100+ cores (e.g., NVIDIA DGX servers and Arm Vera racks).
• Use Cases: Currently powers USearch vector search, NumKong LLM KV-cache/attention kernels, and StringZilla bioinformatics workloads.

Original post →

More from Infra

• Moore's Law Fading: Non-Silicon Computing and Novel Architectures to See Capital Influx — MikePFrank · 2026-08-21
• "Why do we need more datacenters? Just write faster kernels" — basedjensen · 2026-08-21
• Cornell Nested Architecture Cuts Training Compute by 36% — burkov · 2026-08-21
• Tencent releases FlashPrefill V2 for efficient long-context LLM serving — tencent · 2026-08-21
• SGLang author asks community for pain points, vows to fix them — BanghuaZ · 2026-08-21
• Local LLM upgrade dilemma: Dual NVIDIA or switch to AMD? — thatObstinateGuy · 2026-08-21

Infra channel →