NVIDIA Vera CPU: 1.8x Faster Cores Boost AI Factory Throughput

NVIDIA Vera CPU delivers 1.8x faster sustained per-core performance under full socket load, the company says, and targets the CPU bottlenecks that limit reinforcement learning training and agentic inference. The design pairs an 88-core Olympus implementation, a monolithic compute die and LPDDR5x memory to cut peak loaded latency and raise per-core bandwidth.

How does Vera speed agentic workflows?

Vera accelerates agentic workflows by raising sustained per-core throughput under full socket load to 1.8x the baseline and completing up to 85% of environment evaluations in the same training window where a baseline CPU might only complete 45%. That higher per-core performance comes from the Olympus core with a neural branch predictor, a 10-wide decode front end, deep out-of-order execution and NVIDIA Spatial Multithreading, all aimed at reducing stalls in sequential, branch-heavy workloads such as simulators, Python logic and tool calls.

Vera’s design prioritizes sequential performance at scale. The architecture focuses on sustained single-threaded work while sockets remain densely loaded by sandboxes, tool invocations and environment rollouts that feed reinforcement learning. More completed evaluations per cycle produce a richer training signal for GPU-based gradient computation.

What latency and memory improvements does Vera bring?

Vera reduces peak loaded latency by 40% compared with x86 CPUs and supplies up to 1.2 TB/s total memory bandwidth, with up to 14 GB/s per core, the company states. These figures aim to keep tool responses and sandboxed tasks predictable when sockets host thousands of concurrent environments.

Those memory and latency characteristics stem from a monolithic compute die with 88 Olympus cores, a large unified cache and NVIDIA Scalable Coherency Fabric. The fabric avoids cross-chiplet core-to-core hops and the tail latency cliffs seen in multi-chiplet designs. Using power-efficient LPDDR5x, NVIDIA says Vera delivers more than three times the per-core memory bandwidth at less than half the power of traditional data center CPUs.

How does Vera affect GPU efficiency and inference sessions?

Vera reduces the CPU-side gaps between GPU steps so that KV caches are less likely to be evicted, keeping more active context resident in HBM and cutting expensive recompute. Slower CPUs increase the chance that a session’s KV cache will be evicted during tool calls, forcing the GPU to rebuild prior context. Faster tool execution, higher sustained per-core throughput, high memory bandwidth and predictable latency compress that CPU-side portion of the trace and increase the share of GPU time spent on token generation rather than recompute.

NVIDIA positions this as a direct lever on fleet productivity: more timely RL feedback improves gradient quality and time-to-convergence, and lower loaded latency preserves service-level agreement responsiveness in interactive agentic deployments.

Why it matters

Agentic AI depends on tightly coupled CPU and GPU work: the CPU executes sandboxed evaluations, orchestration, retrieval and tool calls that feed and pause model steps. Vera reframes the CPU from background infrastructure to a performance vector that affects RL training throughput, inference latency and GPU utilization. Concrete numbers matter here: improving per-core sustained performance from a baseline that completes 45% of evaluations to up to 85% alters how much useful feedback GPUs receive, while 40% lower peak loaded latency and up to 1.2 TB/s memory bandwidth reduce the operational friction that forces recompute and idle GPU cycles.

What to watch

Look for independent benchmarking and system-level results to confirm these improvements in real workloads: NVIDIA points readers to Vera CPU benchmarking by Phoronix and the NVIDIA Vera Rubin NVL72 as next steps for real-world performance data.