NVIDIA unveils Vera CPU for AI agent data centre workloads
Wed, 8th Jul 2026 (Today)
NVIDIA has introduced its Vera data centre CPU, aimed at AI agent workloads.
Vera is built around Olympus, NVIDIA's custom CPU core, which it says delivers 50% higher instructions per cycle than Grace. The processor has 88 cores and up to 1.2TB/s of LPDDR5X memory bandwidth, alongside a monolithic compute die designed to keep performance steady under load.
The launch reflects NVIDIA's argument that AI agents place new demands on data centre processors. Rather than focusing mainly on core counts, it argues that the key requirement is stronger single-threaded performance for the sequence of tasks between model calls, such as tool use, code execution, data processing and result checking.
According to NVIDIA, conventional server CPUs have been shaped by cloud economics that favour more cores at a lower cost per rentable core. It says that trend has reduced per-core speed and introduced memory-access bottlenecks, particularly in chiplet-based designs where not every core can use the processor's full memory performance.
Performance claims
NVIDIA says Vera delivers 1.8x the sustained per-core performance of x86 processors in loaded CPU workloads intended to represent agentic execution. It also says the chip provides 3.4TB/s of core-to-core bandwidth, which it describes as three times higher than other data centre CPUs.
It links those claims to the economics of AI infrastructure, arguing that delays in CPU tasks can leave graphics processors idle while they wait for surrounding work to finish, reducing utilisation in what it calls AI factories.
One example came from Perplexity, which tested Vera on a coding workflow that involved cloning a repository and running its test suite in sandboxes. NVIDIA says Vera completed the job about 1.5x faster than x86 systems and started concurrent sandboxes up to 1.9x faster.
Perplexity is assessing Vera for use in a production system, according to NVIDIA. It also says partners measured 3x faster large-scale SQL analytics with Starburst and up to 6x lower latency in real-time streaming with Redpanda, compared with leading x86 server CPUs.
Broader strategy
Vera is part of NVIDIA's broader effort to expand its role in data centres beyond GPUs. The same CPU will host GPUs in Vera Rubin systems and will also be used in the BlueField-4 STX storage processor, giving customers a common architecture across different parts of an AI installation.
That approach matters because AI agent workloads are not confined to one type of computing task. NVIDIA says a single deployment may need to run tools and sandboxes, handle data processing, serve requests and support reinforcement learning for model training, all on the same underlying infrastructure.
The chipmaker presents Vera as a response to a shift from intermittent, user-driven software activity to persistent, parallel agent-based work. In that model, swarms of agents run continuously, with each step depending on the output of the last, making the speed of each individual core more important than aggregate throughput alone.
NVIDIA also outlined the next step in its CPU roadmap. Rosa, a future processor based on its Rigel Arm v9.2 core, will deliver higher per-core performance than Olympus while keeping the same silicon footprint, according to the company. Planned changes include improved instruction delivery, a larger L2 cache and more efficient memory handling.
"The world counts in seconds. Agents count in nanoseconds."