Intel + Google chip pact

Intel and Google said on April 9 that they signed a multiyear chip deal built around Intel Xeon central processing units and custom infrastructure processing units inside Google’s cloud systems. Google is not replacing graphics processing units for artificial intelligence, but it is putting more of the non-math work back onto other chips. (intel.com) A central processing unit is the general manager chip in a server: it schedules tasks, moves data, runs operating systems, and handles the parts of artificial intelligence that are not giant matrix multiplications. Intel said Xeon will keep powering Google Cloud across inference, general-purpose computing, and parts of the artificial intelligence stack. (intel.com) An infrastructure processing unit is the traffic-cop chip in a data center: it offloads networking, storage, and security jobs that would otherwise eat into central processing unit time. Google has already used custom Intel infrastructure processing units in its C3 and Z3 machine families, where those chips sit next to Intel Xeon processors. (cloud.google.com) Google’s earlier C3 launch spelled out the pitch in concrete terms: pair a 4th Gen Intel Xeon Scalable processor with a custom Intel infrastructure processing unit so the main processor can spend more time on customer workloads. That is the same design logic now being extended into the artificial intelligence buildout. (cloud.google.com) The reason this is coming up now is that artificial intelligence data centers turned into graphics processing unit factories in 2023 and 2024, with buyers chasing the fastest training chips first. That solved one bottleneck, but it also made power, networking, memory movement, and server utilization more expensive constraints. (cnbc.com) Intel is trying to win the next round by arguing that a server is not one chip but a stack of jobs, and each job should land on the cheapest chip that can do it well. In its April 9 announcement, Intel said the Google deal will span multiple future Xeon generations and expanded co-development of application-specific integrated circuit infrastructure processing units. (intel.com) Google already has proof points for that mixed approach in public products. Its C4 virtual machines based on Intel Xeon 6 reached general availability with up to 4.2 gigahertz peak frequency, and Google said those instances can deliver up to 30% better general compute performance and up to 60% better machine learning recommendation performance than the prior generation. (cloud.google.com) That does not mean Google is betting against its own tensor processing units or against graphics processors from Nvidia and Advanced Micro Devices. It means Google is treating artificial intelligence infrastructure more like an airport, where one machine scans bags, another handles tickets, and another flies the plane. (intel.com) For Intel, the deal is also a credibility play after years of watching the artificial intelligence boom center on accelerator chips it did not dominate. For Google, the upside is lower waste inside giant clusters, because every storage packet or security task moved off the main processor can free more expensive compute for customer jobs. (cnbc.com) So the news is not “Intel found the new artificial intelligence chip.” The news is that one of the biggest cloud builders is doubling down on the older idea that balanced systems win at scale, and Intel still has a seat at that table if the central processing unit and the infrastructure processing unit can make the expensive accelerators work harder. (intel.com)