NVIDIA preempted AMD’s event by publishing technical details on the Vera CPU, which it bills as the first processor designed at the per‑core level for “Agentic AI” workloads. The announcement covers both the Olympus software and hardware architecture, performance figures, novel memory support, and progress toward mass production.
A core architecture built for Agentic AI
According to NVIDIA, Agentic AI shifts critical execution paths onto the CPU: agents run sandboxed code, invoke tools, access databases, and orchestrate complex control loops. To meet these demands, Vera is based on the Olympus architecture and so‑called Olympus cores, engineered to preserve high single‑thread performance even under full load.
Key innovations include a Neural Branch Predictor intended to better handle the irregular control flow and long‑range dependencies typical of agent workloads. Each Olympus core pairs with a 10‑wide decoder, and the design favors a “wide‑issue, deep‑pipeline” microarchitecture. A Vera socket houses 88 cores for 176 threads.
Vera is also presented as the first CPU to support FP8 precision natively in its cores, bringing numeric capabilities on the CPU closer to those traditionally reserved for GPUs.
Memory and bandwidth: SOCAMM2 LPDDR5X
NVIDIA highlights memory as the bottleneck for running many agents concurrently. Vera adopts SOCAMM2 LPDDR5X, which is claimed to deliver 1.2 TB/s of bandwidth—roughly double that of a conventional DDR5 server—while consuming less than half the power. In reported tests, Vera would sustain 90% of its peak bandwidth, ensuring that execution contexts remain fed even at large scale.
Measured performance
[Text cut off in the source]
