ACCL+: FPGA-Based Collective Engine for Distributed Applications

Explore a 15-minute conference talk from USENIX OSDI '24 that introduces ACCL+, an innovative FPGA-based collective communication library. Discover how this open-source solution enhances the development of distributed applications using FPGAs in cloud deployments. Learn about ACCL+'s portability across platforms, support for various networking protocols, and its ability to enable direct FPGA-to-FPGA collective communication. Understand its dual functionality as a collective offload engine for CPU applications and a component in FPGA-based distributed systems. Examine the performance evaluation results comparing ACCL+ to software MPI over RDMA on an FPGA cluster with 100 Gb/s networking. Gain insights into practical use cases, including CPU-based vector-matrix multiplication and FPGA-based distributed deep-learning recommendation inference.

OSDI '24 - ACCL+: an FPGA-Based Collective Engine for Distributed Applications