Intel and Google deepen tie

Google just made a very unglamorous AI bet: not on one more giant graphics chip, but on the plumbing around it. On April 9, Intel and Google said Google Cloud will keep using multiple future generations of Intel Xeon processors and will expand joint work on custom infrastructure chips. (intel.com) A central processing unit is the general manager inside a server. It handles operating systems, memory movement, scheduling, and the thousands of small jobs that have to happen before an artificial intelligence model can answer a prompt. (intel.com) A graphics processing unit is the specialist that does huge batches of math fast. That makes it great for training and running large artificial intelligence models, but a data center still needs other chips to feed it data, connect it to storage, and keep network traffic moving. (reuters.com) The chip Google and Intel are building together is an application-specific integrated circuit, which means a chip made for one narrow job instead of many. In this case the job is infrastructure processing: offloading networking, storage, and security work so the main processor is not stuck doing warehouse logistics. (datacenterdynamics.com) Intel says those infrastructure processing units are meant to improve utilization and make performance more predictable at hyperscale, which is the industry term for data centers with hundreds of thousands of servers. Google is not replacing accelerators here; it is trying to keep the rest of the system from becoming the bottleneck. (intel.com) This is not a brand-new relationship. Google Cloud already uses Intel Xeon 6 processors in its C4 and N4 virtual machine families, which are the rentable server instances customers buy for general computing and artificial intelligence inference workloads. (intel.com) Inference is the moment a trained model actually does work for a user, like writing a reply or classifying an image. Reuters reported that as the market shifts from training models to deploying them at scale, demand rises for the kind of central processing units that keep those live services running. (reuters.com) That helps explain why Google is signing up for “multiple generations” of Xeon instead of treating the processor as a commodity part it can swap later. A cloud operator gets lower engineering risk when the central processor, the network offload chip, and the software stack are planned together years ahead. (intel.com) For Intel, this is one of the clearest signs yet that it still has leverage in artificial intelligence infrastructure even without owning the headline-grabbing accelerator market. For Google, it is a way to build a more balanced machine room where expensive accelerators spend less time waiting on data and more time working. (techcrunch.com) The financial terms were not disclosed, and neither company gave a launch date for the new infrastructure processing units. What they did spell out is the design philosophy: artificial intelligence does not run on accelerators alone, so the next fight in cloud computing is over the whole system around them. (datacenterdynamics.com)