Nvidia Vera Architecture Signals Shift to Custom Arm Cores for AI Infrastructure
Nvidia’s Vera CPU marks a shift to fully-custom Armv9.2 silicon, utilizing a monolithic 3nm compute die and simultaneous multithreading to optimize host processing for agentic AI workloads.
Nvidia is shifting its datacenter strategy with the launch of the Vera CPU, marking the company’s first fully-custom processor design. According to The Register, this move represents a direct challenge to the x86 dominance of Intel and AMD, with Vera positioned as a standalone platform available to hyperscalers and cloud providers, including Alibaba, ByteDance, Meta, Oracle, CoreWeave, Lambda, Nebius, and NScale. Unlike previous iterations that relied on off-the-shelf Arm Neoverse or Cortex IP, Vera utilizes a custom Armv9.2-compatible core architecture dubbed Olympus. This silicon is designed to function both as an AI head node for GPU management and as a host for agentic AI workloads that operate outside of GPU acceleration.
The Vera architecture utilizes a monolithic compute die fabricated on TSMC’s 3nm process, a design choice Nvidia argues improves core-to-core bandwidth and latency compared to multi-die compute approaches. The Register reports that the compute die is surrounded by disaggregated chiplets, including eight LPDDR5x memory controllers and dedicated I/O dies for PCIe 6.4, CXL 3.1, and NVLink Chip-to-Chip (C2C) interfaces. This packaging strategy bears structural similarities to Amazon’s Graviton 4, separating memory and I/O functionality from the compute complex. The platform is available in single-socket configurations or as a Superchip, which links two Vera CPUs via NVLink-C2C to provide 176 cores and 352 threads.
The Olympus core itself features a 10-wide decoder and dispatch, eight integer ALUs, and six vector/FP pipelines supporting SVE2 and FP8. While the architecture shares some lineage with the Cortex X925, Nvidia has implemented significant custom modifications to maximize instructions per clock (IPC). Central to this is a custom neural branch predictor capable of exploring two branches simultaneously, which The Register notes is designed to reduce pipeline stalls in branch-heavy workloads like Python scripts. The core also incorporates memory renaming to accelerate store-to-load dependency chains and a value prediction scheme that executes dependent instructions speculatively, targeting performance gains in pointer-heavy software and graph traversal tasks.
Memory and interconnect bandwidth are defining features of the Vera platform. The system supports up to 1.5 TB of LPDDR5X memory, which, in the Superchip configuration, utilizes 16 SOCAMM2 modules to deliver an aggregate bandwidth of 2.4 TB/s. This represents a significant throughput increase over traditional server DDR5 implementations. Furthermore, the Superchip utilizes a 1.8 TB/s bidirectional NVLink-C2C interconnect to join the two CPUs. This high-speed fabric is critical for Nvidia’s reference designs, which aim to pack up to 128 Superchips—totaling 22,528 cores and 384 TB of memory—into a single liquid-cooled rack for large-scale agentic AI deployments.
A notable departure from typical Arm-based server designs is the inclusion of simultaneous multithreading (SMT). While Arm historically favored single-threaded cores for power efficiency, Nvidia has implemented two-way SMT in the Olympus core to better utilize execution units during idle cycles. This implementation, combined with the chip’s focus on mitigating pipeline bubbles, suggests a design philosophy centered on high-throughput orchestration rather than raw single-threaded performance. By integrating these specific architectural optimizations, Nvidia is attempting to create a CPU that functions as a high-performance host for AI agents, effectively blurring the lines between traditional host processing and the unified memory fabric of the accelerator cluster.
Sources
- 01 A deep dive into Nvidia's Vera CPU and the Olympus cores that power it — The Register