Nvidia’s InfiniBand Problem - Spectrum-X AI Fabric, Tomahawk-5, Jericho-3AI, Quantum-2
Competing teams and technologies, but can the pivot challenge Broadcom's dominance
In the gold rush to build out AI infrastructure, GPUs are the picks and shovels. With Nvidia’s historic earnings last week, the semiconductor market soared. The market incorrectly bid up some other companies with disproportionately less AI exposure to try and capture some of the gold rush. To be clear, the largest percentage of spend for AI infrastructure goes to the AI accelerators. The next largest shift in dollar share goes to networking. Mapping out which hyperscalers are using what networking gear is extremely important.
From the outside, Nvidia clearly has the most visionary CEO of the 21st century, Jensen Huang, with a cohesive grand strategy in accelerated computing and AI. A handful of years ago, Nvidia acquired a leading networking firm, Mellanox, to further that vision. Despite the clear goals, within the networking stack, there is an immense amount of complexity. As such, there is an internal battle raging at Nvidia. Nvidia has a fundamental problem with the InfiniBand network stack.
SemiAnalysis is an ad-free, reader-supported publication. To receive new posts and support, consider becoming a subscriber.