Nvidia Vera CPU Launch Redefines AI Computing Through Custom Olympus Cores and High Bandwidth Interconnects for Global Enterprise Deployment and Autonomous Agent Success
Nvidia today launched its Vera CPU, which it said is the first processor design tailored for running autonomous AI agents. The processor has entered production and it promises to complete tasks 1.8x faster than x86 processors. This is the first expansion in processors beyond the Grace CPU line, which has had 2.5M units shipped so far. This piece of hardware is designed to fulfill the workloads that agentic systems generate when software models need to execute code, perform a step, and leverage a tool on their own within modern data centers.
The Vera processor's system design relies on 88 custom Olympus cores and utilizes Spatial Multithreading and a LPDDR5X memory subsystem that can output 1.2 TB/s of memory bandwidth. According to open source testing suite Phoronix benchmarks, it performed with the fastest throughput among CPUs in common agentic tasks like database operations, running python, and code compilation, thereby addressing the latency that occurs with AI accelerators that are limited by the CPU.
Major finance and aerospace firms, along with AI labs, have started testing out the hardware for an immediate production rollout. Lynn Martin, president of NYSE Group, pointed out the scale at which its platform operates (1.1T messages a day) and confirmed that the firm is integrating the new chips with Redpanda and HPE, saying:
"At the NYSE, our goal is to maximize the performance, efficiency and reliability of the underlying infrastructure. With the NVIDIA Vera CPU, we will be expanding our computational capabilities and lowering latency to deliver the resilient and AI ready financial market infrastructure we operate on a daily basis."
Advanced labs are keen to use it to take on intensive reasoning pipelines. James Bradbury, the director of compute at Anthropic, stated:
"Scaling computing power is a key requirement to build next generation AI models" and the firm is "excited for Vera to emerge as a valuable platform for agentic workloads."
Oracle Cloud Infrastructure and other hyperscalers will be shipping supercomputers powered by Vera chips to support demanding reasoning pipelines soon.
The CPU is designed to serve as a host to Nvidia Vera Rubin platforms by utilizing 2nd generation NVLink C2C interconnect for 1.8 TB/s coherent bandwidth between the CPU and GPU, while rack scale confidential computing protocols will be applied to secure sensitive data during execution. This is also bundled in the BlueField 4 STX platform, which packages a network interface card, processor, and a range of security features into an integrated, secure by design solution.
Nvidia is rolling out hardware in multiple different server configurations to suit various data center environments: densely packed, liquid cooled racks, and flexible 2 socket, air cooled designs. It will be carried to market by major hardware manufacturers (Dell, HPE, Lenovo, and Supermicro) and represents the first standard non x86 processor option for enterprise servers. Cloud providers like CoreWeave, Lambda, Nebius and Together AI are lining up to ship supercomputers powered by Vera chips, with full partner availability beginning this autumn.
