“These accelerators have gotten more and more necessary in cloud infrastructure as a consequence of their superior price-performance and price-efficiency ratios, which result in higher return on investments,” he stated.
Microsoft was a bit late in becoming a member of the customized chip revolution. Its rivals had launched customized chips for AI workloads years earlier, AWS within the type of Trainium and Inferentia, and Google within the type of Tensor Processing Models (TPUs), nevertheless it wasn’t till final 12 months’s Ignite convention that Microsoft unveiled its first customized chips, Maia and Cobalt, to sort out inner AI workloads and make its knowledge facilities extra vitality environment friendly.
Dashing AI dataflows with DPUs, not GPUs
At this 12 months’s occasion, Microsoft launched two extra chips: the Azure Increase DPU to speed up knowledge processing, and the Azure Built-in HSM module to enhance safety.
The Azure Increase DPU is a hardware-software co-design, particular to Azure infrastructure, that runs a customized, light-weight data-flow working system that Microsoft claims allows greater efficiency, decrease energy consumption, and enhanced effectivity in comparison with conventional implementations.
Microsoft can be introducing a brand new model of its liquid-cooling sidekick rack to assist servers working AI workloads, and a brand new disaggregated energy rack co-designed with Meta that it claims will allow 35% extra AI accelerators to slot in every server rack.
“We count on future DPU-equipped servers to run cloud storage workloads at 3 times much less energy and 4 instances the efficiency in comparison with current servers,” the corporate stated in a weblog put up.
