Meta AI Updates: May 6, 2026

1. Meta Open-Sources In-Kernel Broadcast Optimization for RecSys Inference

Meta. Meta engineers published In-Kernel Broadcast Optimization (IKBO), a kernel-model-system co-design technique that eliminates redundant user-embedding broadcast in recommendation inference by pushing the broadcast logic inside GPU kernels rather than materializing replicated tensors. The team reports a roughly 4x speedup on its Linear Compression kernel on H100 SXM5, 2.4x and 6.4x throughput gains for Flash Attention (kernel-only and kernel-plus-broadcasting), and up to a 2/3 reduction in compute-intensive net latency on co-designed models. IKBO is deployed across Meta’s recommendation funnel on both GPU and MTIA — including the Meta Adaptive Ranking Model — and the code is available in FBGEMM. Source