NVIDIA has unveiled its Vera CPU, a processor specifically engineered for the agentic AI era, prioritizing high-performance single-threaded execution over traditional core-count scaling. Designed to minimize latency in AI agent loops, the Vera CPU features the custom Olympus core, which NVIDIA reports provides a 50% increase in instructions per cycle compared to its predecessor, the Grace CPU.
The Shift Toward Single-Threaded AI Performance
Modern data center CPUs have historically prioritized high core counts to maximize cost-efficiency for cloud multi-tenancy. However, this architectural focus creates bottlenecks for agentic AI, which requires rapid, sequential processing. According to NVIDIA, agentic workflows—where a model must reason, execute a tool call, process data, and verify a result before moving to the next step—depend heavily on the speed of individual cores rather than the total number of cores available.
When a CPU is designed primarily for core density, it often sacrifices the memory bandwidth and instruction processing speed necessary for these sequential loops. NVIDIA’s approach with the Vera CPU seeks to reverse this trend by emphasizing sustained per-core performance, ensuring that each step of an agent’s task completes without waiting for resource-heavy contention.
Architecture and Performance Metrics
The Vera CPU is built on a monolithic compute die, a design choice intended to reduce latency and improve data movement efficiency. Key technical specifications include:
- Custom Olympus Cores: These cores are designed to deliver a 50% improvement in instructions per cycle (IPC) over the NVIDIA Grace CPU.
- Memory Bandwidth: The system supports up to 1.2TB/s of LPDDR5X memory bandwidth while maintaining a power profile of less than 40 watts for memory operations.
- Interconnect Speed: The architecture provides 3.4TB/s of core-to-core bandwidth, which NVIDIA claims is three times higher than current industry-standard data center CPUs.
By maintaining high bandwidth across all 88 cores, the Vera CPU prevents the bottlenecks that typically occur in chiplet-based architectures. Performance testing in agentic environments indicates that Vera delivers 1.8 times the sustained per-core performance of x86-based processors.
Practical Applications in AI Factories
For organizations building AI factories, CPU performance directly impacts GPU utilization. If a CPU cannot process code execution or data retrieval fast enough, the expensive GPUs in the cluster remain idle.
Perplexity, an AI search company, tested the Vera CPU against standard x86 hardware. In coding workflows—such as cloning repositories and running test suites—Perplexity found that Vera completed tasks approximately 1.5 times faster. Furthermore, the ability to initiate concurrent sandboxes was recorded at up to 1.9 times faster than the x86 baseline. Data-heavy tasks also showed significant gains; partners reported 3 times faster large-scale SQL analytics using Starburst and 6 times lower latency for real-time streaming via Redpanda compared to leading x86 server CPUs.
Future Roadmap
NVIDIA has already outlined its path beyond the Vera CPU with the upcoming "Rosa" CPU. This next-generation processor will utilize the "Rigel" core, an Arm v9.2-based architecture. According to the company, Rigel will maintain the same silicon footprint as Olympus while offering further enhancements to instruction delivery, a larger L2 cache, and more efficient memory management.
By standardizing these CPUs across the NVIDIA Vera Rubin platform and BlueField-4 STX storage processors, the company intends to provide a unified architecture for the entire AI factory stack, from data retrieval to model training and agentic execution.