Google deepens Intel tie

Artificial intelligence data centers are full of expensive chips that spend part of their day doing janitor work. On April 9, Intel and Google said they are expanding a multiyear deal so more of that housekeeping gets pushed off the main processors and more of the machine can be used for models. (intel.com) Google Cloud is committing to multiple future generations of Intel Xeon processors, which are the general-purpose central processing units that run the basic logic of a server. Intel said those chips will keep powering Google Cloud systems for artificial intelligence, inference, and ordinary cloud workloads. (intel.com) The second part is more unusual: the two companies are widening a joint effort to build custom application-specific integrated circuit infrastructure processing units. That is a mouthful for a chip built to handle network, storage, and security chores so the main processor does not have to. (intel.com) Intel has been making this argument for years with its infrastructure processing unit line. In Intel’s own examples, moving virtual switch work onto an infrastructure processing unit raised system throughput by 50% versus a standard network card in some microservices tests. (intel.com) Google already uses the same basic idea in its own cloud hardware. Google says its Titanium system offloads networking and storage processing to free up the central processing unit, and its newer Axion virtual machines pair that offload layer with Google-designed processors. (cloud.google.com) That makes this Intel deal less of a sudden pivot and more of a widening hardware stack. Google is still building its own Tensor Processing Units for artificial intelligence and Axion chips for general computing, while also locking in Intel central processors and co-designing more infrastructure silicon with Intel. (cloud.google.com 1) (cloud.google.com 2) (intel.com) Intel also gets something it badly needs: a marquee cloud customer saying central processing units still matter in artificial intelligence systems. Intel’s announcement explicitly framed modern artificial intelligence infrastructure as “heterogeneous,” meaning different chip types split the work instead of one processor doing everything. (intel.com) This is also not the first time the two have built one of these chips together. Reuters reported in September 2022 that Intel and Google Cloud launched a co-designed infrastructure processing unit aimed at making data centers more secure and efficient. (investing.com) The practical bet is simple: if an artificial intelligence server spends fewer cycles moving data, checking packets, and handling storage traffic, the same rack can devote more of its power budget to training or serving models. That is why a deal about “boring” infrastructure chips can matter just as much as the flashier race for graphics processors. (intel.com)