NVIDIA Details Vera CPU with 88 Olympus Cores for Agentic AI
NVIDIA has detailed its new Vera CPU, which features 88 custom-designed Olympus cores and 1.2 TB/s LPDDR5X memory. The processor is specifically optimized to handle the unique single-thread performance and memory bandwidth demands of Agentic AI workloads. As Agentic AI systems increasingly rely on CPUs to execute code, call tools, and manage complex workflows, this hardware addresses critical bottlenecks in single-thread performance and memory access. It marks a significant shift toward hardware tailored for autonomous AI agents rather than just raw LLM inference. The Vera CPU features 176 SMT threads, a 10-wide decode engine, a neural branch predictor, and a 164 MB unified L3 cache connected via the NVIDIA Scalable Coherency Fabric. It also supports dual-socket configurations via NVLink-C2C, providing 176 PCIe 6.4 lanes and CXL 3.1 support.
## BACKGROUND
Agentic AI refers to autonomous AI systems that can pursue goals, use tools, and interact with databases or sandboxed environments to complete complex tasks. Unlike traditional AI models that run primarily on GPUs, these agentic workflows shift significant execution logic to the CPU, requiring high single-thread performance and low-latency memory access.