New 'Yellow Pages' Ranks AI Datacenter CPUs

- The "AI Datacenter CPU Yellow Pages" report evaluates processors on their suitability for two primary AI workload types: "reasoning" and "agentic" (or "action-oriented"). Reasoning workloads are memory-bandwidth intensive, while agentic workloads benefit from higher core counts and larger caches for managing a large number of parallel tasks. - Nvidia's Grace CPU, part of the Grace Hopper Superchip, combines 72 Arm Neoverse V2 cores with a high-bandwidth NVLink-C2C interconnect that provides 900 GB/s of bidirectional bandwidth to a Hopper GPU. This design allows the GPU to directly access up to 480GB of the CPU's LPDDR5X memory, beneficial for large-scale AI models. - AMD's EPYC "Turin" processors, based on the "Zen 5" architecture, scale up to 192 cores and 384 threads. These CPUs support 12-channel DDR5 memory and offer specialized models with high frequencies to balance core count and single-thread performance for diverse datacenter workloads. - Intel's upcoming "Diamond Rapids" Xeon 7 processors are expected in 2026 and will feature up to 192 "Panther Cove" P-cores. These processors will be built on Intel's 18A process and will support up to 16 channels of DDR5 memory, targeting high-performance computing and AI. - Amazon's AWS Graviton5 is a custom-designed Arm-based processor with 192 cores, manufactured using a 3nm process. It is part of a broader trend where hyperscalers like Google (with its Axion CPU) and Microsoft (with its Cobalt CPU) are developing their own chips to optimize performance and cost for their specific cloud and AI services. - The market for AI server processors is experiencing rapid growth, with a projected value of over $57 billion by 2034, driven by the intensive computational demands of AI training and inference. This has caused a market share shift, with Nvidia growing to 86% of the AI data center revenue by late 2025, while Intel's share has decreased. - The increasing power density of these new AI-focused CPUs and GPUs is driving a transformation in data center design, with thermal management, particularly direct liquid cooling, becoming a primary determinant of facility architecture. The demand for power is projected to more than triple in the United States by 2030 to support this infrastructure. - CPUs are considered a critical component for orchestrating AI tasks, affecting GPU utilization and the total cost of ownership. Recent trends show a surge in demand for high-performance CPUs for reinforcement learning (RL) and managing complex AI agent workflows, leading to a renewed focus on CPU capabilities alongside GPUs.