On March 24 local time, Arm officially launched its first in-house developed CPU for artificial intelligence (AI) data centers, the AGI CPU, with internet giant Meta confirmed as one of the first customers.
Built on TSMC’s 3nm process, the single AGI CPU supports up to 136 Neoverse V3 cores, operating at a base frequency of 3.2 GHz and a peak frequency of 3.7 GHz. Featuring independent per-thread cores, it delivers deterministic performance under sustained workloads, avoiding throttling and thread idling. The 136 cores are distributed across two small chiplets, with a TDP of approximately 300 watts.
In terms of memory and interface support, the AGI CPU offers 12 channels of DDR5 memory at up to 8800 MT/s, delivering total memory bandwidth of over 800 GB/s, or 6 GB/s per core, with latency below 100ns. For I/O, it includes 96 PCIe Gen6 lanes and native support for CXL 3.0 for memory expansion and pooling.
Arm stated that the AGI CPU achieves more than double the per-rack performance compared to x86 CPU platforms. Based on Arm’s AGI CPU, capital expenditure (CAPEX) per gigawatt of AI data center computing power can be reduced by up to $10 billion relative to x86 platforms.
As an early partner and co-developer of the Arm AGI CPU, Meta announced it will use the agentic AI CPU to optimize infrastructure across its full suite of applications, with coordinated deployment alongside its proprietary Meta Training and Inference Accelerator (MTIA), enabling more efficient orchestration and scheduling in large-scale AI systems. The two companies have committed to long-term, deep collaboration across multiple generations of Arm AGI CPU products.
Arm projects that by 2030, its in-house chip business will generate around $15 billion in annual revenue, driving total company sales to $25 billion — more than five times current levels.
Buoyed by the positive news, Arm’s U.S. stock price soared 16.38% on March 24, pushing its market capitalization above $166.8 billion.
(Reprinted from https://news.eccn.com/）