American semiconductor giant Broadcom has launched a new network switch chip for connecting supercomputers engaged in operations related to artificial intelligence (AI).

Called Jericho3-AI, the chip is designed to reduce the time taken to complete any AI task.

With the rise of AI apps like Alphabet’s Bard and OpenAI’s ChatGPT, new challenges have emerged for the networks inside data centres.

Enormous amounts of data are used to train these systems to enable them to give human-like responses to the prompts.

The task is too big for one computer chip to handle; therefore, it is divided across tens of thousands of graphics processing units (GPUs), which then operate as one massive computer to complete it.

Broadcom claims that Jericho3-AI can connect up to 32,000 GPUs.

The chipmaker added that the Jericho3-AI fabric delivers 40% less power per gigabit while providing 26 petabits per second of Ethernet capacity, nearly four times as much as the previous generation.

In comparison to other networking solutions, the Jericho3-AI fabric is claimed to speed up project completion times for AI benchmarks by at least 10%.

Broadcom general manager of Core Switching Group Ram Velaga 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.”