New chip on the block: Broadcom's Jericho3-AI can connect up to 32,000 GPU chips
American semiconductor manufacturing company Broadcom Inc. has released a new chip Jericho3-AI, which is being touted by the company as the highest-performance fabric for artificial intelligence (AI) networks. The new chip will wire together supercomputers.
Jericho3-AI is packed with features like improved load balancing - which ensures maximum network utilization under the highest network loads, congestion-free operation which implies no flow collisions and no jitter, high radix which allows Jericho3-AI to connect to 32,000 GPUs collectively, and Zero-Impact Failover - ensuring sub-10ns automatic path convergence. All of this would lead to cutting down on the job completion times for AI workload.
Designed to lower time spent networking during AI training
The performance improvement that Jericho3-AI brings with it is also cost-effective. It decreases the cost of running AI workloads. Broadcom Inc. claims that the network, in effect, pays for itself. The new chip could be particularly useful for its clients. One of the biggest clients that Broadband currently has on is Apple, with which it currently has a $15 billion deal.
Ram Velaga, senior vice president and general manager, Core Switching Group, Broadcom, said in the statement: “The benchmark for AI networking is reducing the time and effort it takes to complete the training and inference of large-scale AI models. Jericho3-AI delivers significant reduction in job completion time compared to any other alternative in the market.”
Jericho3-AI has been launched at a time when AI is omnipresent. According to International Data Corporation (IDC), it is estimated that worldwide spending on AI, including software, hardware, and services for AI-centric systems, will reach $154 billion in 2023, an increase of 26.9 percent in 2022. And over $300 billion by 2026.
AI-powered platforms need immense computing and processing power for data analytics workloads.
Broadcom’s Jericho3-AI chip will now compete with market leader NVIDIA's InfiniBand, which is another supercomputer networking technology. Reuters spoke to Velaga, who said that while NVIDIA systems are the world’s fastest supercomputers, many companies are reluctant to give up Ethernet to buy both GPUs and networking gear from the same supplier.
"Ethernet, you can get it from multiple vendors - there's a lot of competition," Velaga told Reuters. "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."