Broadcom launches Jericho3-AI chip for linking supercomputers: How it works
US-based Broadcom has announced a new chip for linking supercomputers. It is called Jericho3-AI and will be used for joining the systems for artificial intelligence (AI) work. The company claims that Jericho3-AI will revolutionize AI networking, thanks to its best-in-class facilities like zero-impact failover, perfect load balancing, ultra-high radix, and congestion-free operation. So, how does Jericho work? Let us find out.
Why is it necessary to link supercomputers?
For AI applications like Google Bard and ChatGPT, supercomputers have to be trained on lots of data. For this work, thousands of chips called graphics processing units (GPUs) are required, which work for weeks or even months at a time. The faster the connection speed between chips, the quicker the training is completed. Jericho3-AI claims to accelerate the process.
Why should Jericho be used?
The Jericho3-AI can connect up to 32,000 GPUs at the same time. This makes it a direct rival of NVIDIA's InfiniBand gear. Broadcom's chip claims to ensure optimal network utilization even when the network load is "highest." It also promises end-to-end traffic scheduling so that there is no jitter in the network during operations, say when an AI computation cycle is underway.
Jericho is 10% faster than its rivals
Worried about job completion on time? No more. Jericho's zero-impact failover facility makes sure there is auto path convergence in less than 10ns. Also, as mentioned earlier, Jericho3-AI can be connected with up to 32,000 GPUs at once, at a speed of up to 800Gbps in an individual cluster. Compared to alternative networking solutions, Jericho claims to finish jobs 10% faster.
A few more facts about Jericho3-AI
Jericho3-AI has an ethernet bandwidth of 26 petabits per second, which is four times that of Jericho2. It also consumes 40% less power per gigabit. Jericho promises both high bandwidth as well as low-latency.
'Jericho3-AI will reduce training time of large-scale AI models'
Describing the reason behind introducing Jericho3-AI, Ram Velaga, senior vice president and general manager at Broadcom said, "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 a significant reduction in job completion time compared to any other alternative in the market," he added.
