On Tuesday, Broadcom Inc released a new chip that aims to connect supercomputers for artificial intelligence (AI) work using commonly used networking technology. Broadcom is a major provider of chips for Ethernet switches, which are used to link computers within conventional data centers.
The growth of AI applications like OpenAI's ChatGPT and Alphabet Inc's Bard has posed new challenges for data center networks. To respond to inquiries with human-like answers, such systems require vast amounts of data to be trained. This task is too large for a single computer chip to handle. Instead, it must be split across thousands of chips called graphics processing units (GPUs), which must then function as one giant computer to complete the job over several weeks or even months. As a result, the speed at which the individual chips can communicate is critical.
Broadcom has announced the Jericho3-AI chip, which can connect up to 32,000 GPU chips. The Jericho3-AI chip will compete with another supercomputer networking technology called InfiniBand. Nvidia Corp, which acquired InfiniBand leader Mellanox for $6.9 billion in 2019, is now the largest InfiniBand equipment manufacturer.
While Nvidia dominates the GPU market and Nvidia-Mellanox systems are some of the world's fastest supercomputers, many firms are hesitant to give up Ethernet, which is supplied by various companies, to buy both their GPUs and networking equipment from the same supplier, according to Ram Velaga, senior vice president and general manager of Broadcom's core switching group.
Velaga stated that "Ethernet, you can get it from multiple vendors - there’s a lot of competition. If we don’t come out with the best Ethernet switch, somebody else will. InfiniBand is a proprietary, single-source, vertically integrated kind of a solution."