Takeaways
- We are working with ARM to create a new CPU for data centers and large-scale AI projects.
- With ARM, we will develop multiple generations of CPUs, expanding our silicon portfolio and enabling more AI experiences.
We are partnering with ARM to co-develop a CPU specifically architected for the computational intensity of AI training and inference in data centers. As AI workloads accelerate, data centers demand CPUs with increased processing power and efficiency to support large, distributed AI deployments. This collaboration will yield processors optimized for high-throughput, energy-efficient operation critical to Meta’s scalable AI infrastructure.
Delivering AI experiences at a global scale demands a robust, adaptable portfolio of custom silicon solutions purpose-built to accelerate AI workloads and optimize performance across Meta’s platforms. We worked alongside ARM to develop the AGICPU to deploy an efficient compute platform that significantly improves our data center performance density and supports a multi-generation roadmap for our evolving AI CPS systems.
Santosh Janardhan, Head of Infrastructure, Meta.
Arm has introduced the AGI CPU, a production-ready chip built on Arm’s Neoverse platform to support the next generation of AI infrastructure.
For the first time in over 35 years, ARM is releasing its own silicon chips actual physical chips. This expands ARM’s Neoverse platform beyond just IP (intellectual property) and technology designs for compute subsystems. It completes the building-block process, giving customers more ways to use ARM technology. Now they can build custom silicon, use platform-level solutions, or choose ARM-designed processors. This move responds to the fast-changing AI infrastructure and the growing demand for ready-to-use ARM platforms that can be quickly deployed at scale.
The Rise Of The Agentic AI Infrastructure
AI systems now run nonstop globally. Previously, human interaction limited computing speed. Agentic AI removes this bottleneck by enabling software agents to orchestrate tasks, leverage multiple models, and make decisions instantly.
As AI workloads become more multifaceted, CPUs are essential for maintaining reliable operation across distributed AI systems in data centers. They direct multiple tasks while coordinating accelerators, memory, storage, and software agents at scale.
These changes put new demands on CPUs, so processors must evolve.
ARM Neoverse powers the leading hyperscale and AI platforms, including AWS, Graviton, Google, Axon, Microsoft, Azure, Copart, and NVIDIA, Vera. As AI infrastructure grows, partners ask ARM for more. The ARM AGI CPU meets this need.
This Is ARM AGI CPU Built for Rack Scale Agentic Efficiency
Agentic AI workloads need steady, large-scale performance. The ARM AGI CPU is built to deliver strong performance for each task, even under heavy workloads and across thousands of cores, while staying within the power and cooling limits of today’s data centers.
Every aspect of the ARM AGI CPU from clock rates and memory bandwidth to input/output (I/O) paths is designed for highly parallel agentic AI workloads. The architecture ensures high performance in dense computing environments typical of modern data centers.
The ARMS reference server uses a 1-OU unit about 1.7 inches tall, a 2‑node setup with 2 chips and dedicated memory and I/O connections, and 272 cores per blade, a server component. These blades fill a standard air‑cooled 36 kW rack with 30 blades providing 8,160 cores in total. Arm also worked with Supermicro on a liquid‑cooled 200 kW design that can host 36 Arm AGI CPUs, totaling more than 45,000 cores. With this setup, the Arm AGI CPU can deliver over twice the performance per rack compared to the latest x86 systems thanks to the strengths of the Arm architecture and careful alignment of system resources with computing needs.
- The ARM EGR CPUs’ high memory bandwidth allows more effective threads per rack, while x86 CPUs slow down as cores compete under heavy workloads.
- The high-performance, efficient single-threaded ARM Neoverse V3 CPU cores (processing units that handle one instruction thread at a time) outperform older architectures. So, each ARM thread gets more done inherently more usable threads, and getting more work per thread leads to huge performance gains per rack.
The first product of ARM’s AGI CPU is its first data center CPU for the AI era. It offers faster performance per rack and is more efficient than older CPUs. Meta is the lead partner. The ARM AGI CPU is ARM’s first data center CPU for the AI era, offering faster performance per rack and more efficiently than previous CPUs. Meta is the lead partner and core developer focused on improving app infrastructure and integrating with our MTIA silicon. The ARM AGI CPU will be available through ARM, and we will open our board and rack designs under the Open Compute project this year, expanding to deliver production silicon CPUs optimized for large-scale agentic AI deployments. Together, we are combining ARM’s foundation of high-performance power-efficient computing with Meta’s infrastructure expertise to support the next generation of AI systems. Rene Haas, CEO, Arm.
Our ARM CPUs will enhance our silicon portfolio, helping us build a robust, flexible hardware stack to bring AI to billions worldwide.
Source: Meta Partners With Arm to Develop New Class of Data Center Silicon
