In a first for the longtime British chip design firm ARM, it’s set to launch its very own silicon. The 136-core CPU is designed to power AI—because what isn’t in 2026—and has already inked Meta as a major customer. This chip isn’t designed for training or inference, but to handle AI agents. The GPUs and TPUs/LPUs will handle the inferencing, while the CPU does the managing. It’s the middle manager of AI hardware.
ARM is convinced this kind of hardware is necessary, though; indeed, it claims it’s what we need to reach the next level of AI capability.
“We think that the CPU is going to be fundamental to ultimately achieving AGI,” Mohamed Awad, ARM’s EVP of cloud AI, told The Register.
This 136-core CPU is based on ARM’s Neoverse V3 architecture and is being built on a TSMC 3nm process. Each core runs at 3.7GHz boost with a 3.2GHz base and features 2MB of L2 cache, with 128MB of shared system-level cache (SLC). The cores are spread across two dies, with 68 cores apiece. The total power draw for each chip is 300W, so those data centers better have some hefty cooling to chill down more than just the 600W+ GPUs.
Credit: ARM
Each CPU supports 12 channels of DDR5 memory—six per die—with official support for memory speeds up to 8,800 MT/s. That gives each chip a total bandwidth of around 825 GB/s. There’s also support for up to 96 lanes of PCIe 6, and to minimize latency, all I/O and memory controller functions are built into the same die rather than handled separately, as you’d see on a consumer Ryzen CPU, for example.
Although what ARM has built is what it calls a CPU, it’s adopting a design philosophy similar to many ASICs: ditching everything it doesn’t need. That, ARM claims, will give it an advantage in the new hardware race ahead.
“The way that legacy CPUs had been built worried about things like support for legacy applications,” Awad said. “We specifically didn’t want to add things that weren’t going to…be 100% utilized in the mission of this device.”
These CPUs will come in various rack designs. The base model will have a 36kW air-cooled rack with 60 CPUs in the stack spread across 30 compute blades. There’s a dense 200kW liquid-cooled version with 336 of the CPUs inside, delivering a bonkers 455,696 cores.
Meta is one of the first customers for these new CPUs, but OpenAI, SAP, Cerebras, Cloudflare, F5, SK Telecom, and Rebellions are also listed as early customers.
Expect the new CPUs to start showing up in data centers by the end of the year.
