Industry Expert Blogs

No single GPU, XPU, or other AI accelerator can support the computational demand of AI workloads. Tens of thousands — and, in the near future, hundreds of thousands — must work together to share the processing load.

Llama3, for example, will need more than 700TB of memory and 16,000 accelerators for pre-training alone. And, like other AI models, its processing parameters are expected to double every four to six months.

This mass-scale parallel processing and continuous growth put a tremendous strain on the network fabrics that bring together AI clusters, and, more specifically, the interconnects that transport data between all of the accelerators within them.

Emerging standards like Ultra Ethernet and Ultra Accelerator Link (UALink) are addressing the need for larger AI clusters with higher-bandwidth, lower-latency interconnects. And the industry’s first Ultra Ethernet and UALink IP solutions, which we recently announced, will enable massive AI clusters to be scaled both out and up.

Click here to read more ...