Meta will deploy standalone Nvidia Grace CPUs in production, with Vera to follow — company sees perf-per-watt improvements of up to 2X in some CPU workloads
Source: Tom’s Hardware
Image credit: Nvidia
Meta expands Nvidia partnership
Meta announced a multi‑year strategic partnership with Nvidia to broaden its use of Nvidia technology in hyperscale data centers built for AI training and inference. The plan includes deploying millions of Blackwell and Rubin GPUs as part of Meta’s AI spending program, which could total up to $135 billion by 2026.
Performance gains with Grace CPUs
Nvidia’s data‑center chief Ian Buck told The Register that Meta is observing up to 2× performance‑per‑watt on certain workloads when using the Grace platform. Meta is also test‑driving Nvidia’s next‑gen Vera CPU, reporting “very promising” results, with large‑scale Vera‑only deployments potentially starting as early as 2027.
Grace CPU specifications
- Cores: 72 Arm Neoverse V2 cores
- Memory: Up to 480 GB of LPDDR5X in the standalone C1 configuration
- Superchip option: Two dies linked via NVLink‑C2C, providing 144 cores, up to 960 GB of LPDDR5X, and up to 1 024 GB/s aggregate memory bandwidth (depending on memory capacity).
Vera CPU specifications
- Cores: 88 custom Arm cores (up to 176 threads)
- Memory: Supports up to 1.5 TB of LPDDR5X with up to 1.2 TB/s memory bandwidth
- Connectivity: PCIe Gen 6 and Compute Express Link 3.1
- Security: First Nvidia CPU to offer a confidential/trusted computing environment across its rack‑scale systems.
Spectrum‑X Ethernet switches
Meta will also deploy Nvidia Spectrum‑X Ethernet switches throughout its data centers. These switches feature co‑packaged optics, eliminating active cabling and reducing the power consumed by optical transceivers. Combined with the switch’s own power draw, this can account for up to 10 % of a rack’s power consumption, delivering significant performance‑per‑watt improvements for scale‑out applications.
All information is based on publicly available sources and statements from the companies involved.