Powered by NVIDIA Vera CPU
AI agents constantly execute code, query databases, and run tool chains. Traditional server CPUs choke on these irregular workloads, leaving expensive GPUs waiting. The NVIDIA Vera CPU is custom-built to eliminate this exact bottleneck and drive next-generation AI factories.
Vera delivers fast single-threaded performance for complex, unpredictable workflows. Featuring 88 custom Olympus cores on a monolithic die, it accelerates code compilation, tool calling, and Python execution by delivering up to 1.8x higher sustained per-core performance under full socket load compared to leading x86 CPUs. It keeps agent sandboxes responsive and AI factory data pipelines moving smoothly.
Running thousands of concurrent agent sandboxes demands massive data throughput. Vera pairs SOCAMM2 LPDDR5X memory to deliver 1.2 TB/s of aggregate memory bandwidth, or 14 GB/s per core. This delivers twice the memory bandwidth of traditional DDR5 platforms while using significantly less power across your data center.
Built on a single compute die with a 164MB L3 cache, Vera eliminates die-hop latency. Its 1.8 TB/s NVLink™-C2C connection links directly to NVIDIA Rubin GPUs for unified memory sharing, while PCIe® Gen 6 and CXL™ 3.1 provide high-speed I/O. When scaled up to a fully integrated, liquid-cooled rack featuring 256 interconnected Vera CPUs, the architecture supports over 22,500 concurrent CPU environments.
