Host compute
CPU Catalog
A starter comparison of CPUs that matter for AI systems, from conventional server sockets to custom Arm and unified-memory designs.
How to read this table
CPU comparison is tricky because sockets, instances, boost behavior, memory population, and compiler/runtime maturity all matter.
This table tracks platform orientation and public headline specs; detailed benchmarking should live in separate posts.
| Metric | Xeon 6 | EPYC 9005 | Grace CPU Superchip | Axion | Graviton4 | M4 Max |
|---|---|---|---|---|---|---|
| Vendor | Intel | AMD | NVIDIA | Amazon | Apple | |
| Family | Xeon 6 P-core and E-core families | 5th Gen EPYC Turin | Grace | Google Cloud Axion | AWS Graviton | Apple Silicon M-series |
| ISA / core | x86-64; Granite Rapids P-cores and Sierra Forest E-cores | x86-64; Zen 5 and Zen 5c | Arm Neoverse V2 | Arm Neoverse V2 | Arm Neoverse V2 | Arm-based Apple CPU cores |
| Core count | Up to 128 P-cores or up to 288 E-cores by platform class | Up to 192 cores / 384 threads | 144 Arm cores in Grace CPU Superchip | Instance-visible core counts vary by Google Cloud machine type | 96 cores per processor class | Up to 16 CPU cores |
| Memory system | DDR5 and MRDIMM platform options | 12 channels DDR5; large socket memory capacity | Up to 960 GB LPDDR5X ECC | DDR5 cloud server platform | DDR5 cloud server platform | Unified LPDDR memory, up to 128 GB |
| Bandwidth notes | Intel positions Xeon 6 around higher memory bandwidth and I/O versus prior Xeon generations | DDR5-6400 class platform bandwidth with 12 memory channels | Up to 1 TB/s class aggregate memory bandwidth depending on module configuration | Google emphasizes performance and efficiency versus comparable x86 and Arm cloud instances | AWS positions Graviton4 with higher memory bandwidth than Graviton3 | Up to 546 GB/s unified memory bandwidth |
| Target role | General server, cloud, HPC, networking, and edge | General server, cloud, HPC, database, and AI host CPU | HPC and AI host CPU paired with NVIDIA accelerator platforms | General-purpose cloud workloads and CPU-side AI infrastructure | EC2 general-purpose, memory-optimized, and scale-out cloud workloads | Workstation-class client compute and local model experimentation |
| Notes | Track P-core and E-core families separately for benchmark work. | Strong baseline for high core-count x86 host compute. | Interesting because memory bandwidth and energy efficiency are first-class design points. | Public die-level details are sparse, so track by cloud instance family. | Most useful to compare through EC2 instance families rather than bare chip SKUs. | Not a server CPU, but useful for unified-memory local LLM and media workloads. |
| Sources |