PyTorch's FBTriton Kernels Outperform CUDA on Embedding Workloads

Meta's PyTorch team said its FBTriton Triton-based kernels for Table Batched Embedding operations outperform legacy CUDA kernels across recommendation-system workloads, according to a PyTorch blog post¹. Across 307 shard configurations, the median forward speedup is 1.28x. On a B200 GPU configuration, combined forward-backward latency dropped from 79.537 ms to 66.183 ms. TBE kernels handle embedding lookups across thousands of sharded GPUs, combining lookup and pooling in a single GPU launch to reduce overhead.