Broadcom Takes on NVIDIA's InfiniBand and NVSwitch with a Single Chip
Broadcom challenges NVIDIA with Tomahawk Ultra, a low-latency Ethernet chip rivaling InfiniBand & NVSwitch for AI/HPC clusters.
"ChipPub" Publication: 20% Discount Offer Link.
InfiniBand was always considered the mainstream fabric for cross-PC, server, storage, and network use, but this effort failed. The remaining efforts of InfiniBand gained a second life at the turn of the millennium, becoming a high-performance, low-latency interconnect for supercomputers running simulations and models.
For decades, InfiniBand found a niche market thanks to the low latency brought by Remote Direct Memory Access (RDMA). RDMA allows CPUs, GPUs, and other types of XPUs to directly access each other's main memory without going through the entire network software stack.
This was also one of the reasons why Nvidia spent $6.9 billion to acquire Mellanox Technologies more than five years ago. Nvidia undoubtedly foresaw the booming development of GenAI and knew it needed InfiniBand as the backend network connecting GPU server nodes so they could collaborate to train AI models.
GPU Memory Clustering Innovation
Meanwhile, inside GPU servers, Nvidia needed a way to cluster GPU memory so they could share work and run AI training and inference routines as well as HPC code, just like CPUs have Non-Uniform Memory Access (NUMA) technology, which makes multiple compute engines appear to system software as one huge single device.