British chip design company Arm recently announced the launch of a new Arm AGI CPU, entering the data center market for the first time as its own mass-produced chip, targeting emerging "agent" artificial intelligence workloads. This marks a major shift in the business model of this company that has long been focused on licensing IP. For more than three decades, Arm has only provided architecture and core design, which are manufactured by partners. Now it directly participates in the frontline competition of high-performance data center processors with its AGI CPU series.
According to reports, Arm AGI CPU is built based on 3 nm process technology. A single chip integrates up to 136 Neoverse V3 cores. The thermal design power consumption of the entire card is 300 W. Each core can obtain 6 GB/s memory bandwidth and control the access delay within 100 ns. It can support a single chip with a maximum memory capacity of 6 TB and is compatible with DDR5-8800 specifications. In terms of I/O, the processor provides 96 PCIe Gen 6 lanes and integrates CXL 3.0 and AMBA CHI (Coherent Hub Interface) interconnections to support high-bandwidth, low-latency collaborative computing requirements. Each core can independently handle a program thread. Arm said that this design helps avoid threads being crowded or idling for a long time in sustained high-load scenarios and improves overall resource utilization.
In terms of computing power density, which is of great concern to data centers, figures released by Arm show that under an air-cooling system, a single rack can deploy up to 8,160 cores, while the use of liquid cooling solutions can increase this number to more than 45,000. Arm predicts that in order to meet future AI development needs, the CPU computing power required per gigawatt of power in data centers will be more than four times higher than the current level. It believes that the traditional x86 architecture has too much overhead and complexity in this new workload, which is not conducive to further improvements in performance and energy efficiency. Official comparative data shows that the Arm solution can achieve more than twice the performance of x86 in terms of per-rack performance, and has the potential to bring up to $10 billion in cost savings per gigawatt of AI data center computing power.
At the ecological and implementation level, Meta is the core partner and main deployer of this AGI CPU. It will combine the processor with its self-developed MTIA accelerator for large-scale AI scheduling and orchestration on its platform. In addition to Meta, many companies including OpenAI, Cloudflare, Cerebras, Rebellions, SAP and SK Telecom have confirmed their adoption of this platform. In terms of hardware, manufacturers such as ASRock Rack, Lenovo, Quanta, and Supermicro are launching server products based on this platform. The first batch of systems have been provided to some customers, and it is expected to be available on a wider scale in the second half of this year.
Arm said that there are more than 50 ecological partners behind this product, including cloud services, chip manufacturing and storage manufacturers such as AWS, Google, Microsoft, NVIDIA, Samsung, TSMC and Micron, forming a complete chain covering design, manufacturing, hardware systems and cloud deployment. For Arm, the launch of AGI CPU is not only a technical layout, but also means that its role extends from "architecture provider" to "complete solution provider", competing for a higher voice in the new round of AI data center infrastructure competition.