Broadcom launches Tomahawk Ultra AI chip to take on NVIDIA
What's the story
Broadcom has unveiled a new networking processor, the Tomahawk Ultra, to accelerate artificial intelligence (AI) data processing. The processor is designed to connect hundreds of chips working together in data centers. It will compete directly with NVIDIA's NVLink Switch but can connect four times as many chips. The Tomahawk Ultra uses an enhanced version of Ethernet instead of a proprietary protocol for data transfer.
Technological leap
Tomahawk Ultra connects 4 times as many chips as NVLink
The Tomahawk Ultra chip is a major technological leap, acting as a traffic controller for data between multiple closely-packed chips in a data center. Broadcom's Senior Vice President Ram Velaga said that the new chip can connect four times as many chips as NVIDIA's NVLink Switch. This makes it an ideal solution for scale-up computing, where developers need to harness the computing power of nearby chips quickly.
Manufacturing details
Broadcom took 3 years to develop the chip
The Ultra family of chips will be produced by Taiwan Semiconductor Manufacturing Company (TSMC) using its 5nm process. The processor is already being shipped. Broadcom's engineers took almost three years to develop the design, which was initially intended for high-performance computing. However, with the rise of generative AI, Broadcom modified the processor for use by AI companies due to its scaling capabilities.