Meta Deploys Nvidia's Grace CPUs
Meta's multi-billion dollar partnership with Nvidia marks the first large-scale deployment of Nvidia's Grace CPUs in a hyperscaler environment. The move signals a strategic shift for major cloud providers, who have historically sourced general-purpose CPUs from other vendors, toward more vertically integrated GPU/CPU architectures for AI workloads.
- The Grace CPU features 72 Arm Neoverse V2 cores and is connected to a Hopper GPU via Nvidia's NVLink-C2C interconnect, which provides 900 GB/s of bidirectional bandwidth—7x faster than the PCIe Gen 5 standard. - This move is part of a broader multi-year partnership that includes Meta's future deployment of millions of Nvidia's Blackwell and Rubin GPUs, as well as the next-generation "Vera" CPUs. - While Meta has previously used Grace Hopper "Superchips" (integrated CPU+GPU), this is the first large-scale deployment of Grace as a standalone, CPU-only system, signaling Nvidia's direct challenge to Intel and AMD in the general-purpose server CPU market. - Nvidia claims the Grace CPU can deliver up to double the performance-per-watt for some of Meta's backend and AI workloads that do not require a GPU. - The decision to buy from Nvidia contrasts with the "build" strategy of other hyperscalers like Google (Axion CPUs, TPU ASICs) and Amazon (Graviton CPUs, Trainium ASICs), which are increasingly vertically integrating with their own custom silicon. - Meta is pursuing a dual strategy, continuing to invest in its own custom silicon program, the Meta Training and Inference Accelerator (MTIA), to run alongside vendor hardware for its recommendation and ranking models. - The collaboration includes deep co-design work on software and system libraries to optimize performance for Meta's specific workloads, including its large-scale recommendation systems and emerging "agentic AI" applications. - Meta will also integrate Nvidia's Spectrum-X Ethernet networking platform and utilize Nvidia's Confidential Computing for privacy-preserving AI features within services like WhatsApp.
