NVIDIA Vera CPU Sets Benchmark for AI Agent Speed and Efficiency
Home
Latest reporting
News
NVIDIA’s Vera CPU delivers 1.6x more agents per watt and 2x faster agent speed compared to leading x86 CPUs, redefining AI infrastructure performance.
NVIDIA’s Vera CPU is redefining AI infrastructure by delivering industry-leading performance and energy efficiency for agentic AI workloads. Officially launched on March 16, 2026, Vera is designed to handle the complex demands of AI agents, which alternate between GPU-based reasoning and CPU-based execution tasks. The CPU’s architecture addresses the challenge of maximizing task speed and concurrency within power-constrained data centers.
Breakthrough Performance Metrics
In a 100-megawatt data center comparison, NVIDIA reported that Vera supports 1.6 times more agents per watt than the fastest x86 CPU and delivers twice the agent speed of the most energy-efficient x86 CPU. This performance is achieved through Vera’s 88 Olympus cores, which support 176 hardware threads, and its monolithic compute die with 164 MB of unified L3 cache. The CPU also features a coherent on-die fabric capable of 3.4 TB/s bisectional bandwidth, reducing contention and latency.
Additional specifications include up to 1.5 TB of LPDDR5X memory, 1.2 TB/s of memory bandwidth, and a configurable thermal design power of 250–450 watts. The CPU integrates seamlessly with GPUs via 1.8 TB/s of NVLink-C2C bandwidth, enabling efficient CPU-GPU communication.
Partner Results Highlight Real-World Impact
Early adopters have demonstrated Vera’s capabilities in production-style tests:
Perplexity: Achieved 1.5 times faster agentic coding tasks compared to traditional CPUs, with plans to deploy Vera in an upcoming production system.DeepInfra: Reported up to 2.2 times faster agent orchestration compared to the best x86 processor, making it the fastest among four tested architectures.ClickHouse: Scaled near-linearly across 176 cores without Vera-specific tuning, topping the ClickBench hardware leaderboard.DuckLabs: Delivered 1.5 times the median TPC-H SF1000 performance of an Intel Xeon 6 baseline.Daytona: Reduced sandbox initialization time from 88.4 seconds on x86 to 10.8 seconds with Vera, achieving 12.7 completed jobs per second at 2,000 concurrent sandboxes.Modal: Achieved up to twice the sandbox throughput per core versus comparable x86 hardware, with performance gains increasing under heavy load.
Implications for AI Infrastructure
Vera’s ability to combine high speed and energy efficiency within a single CPU simplifies infrastructure choices for AI factories. By eliminating the need to balance between fast and dense x86 parts, operators can maximize work done per watt while reducing latency for individual tasks. As AI workloads continue to grow, Vera’s performance metrics position it as a critical component for next-generation data centers.
