CPUs for AI infrastructure: AMD EPYC, Intel Xeon, and NVIDIA Grace

Updated December 11, 2025

December 2025 Update: AMD server CPU share reaching 27.8% in Q3 2025, projected 40%+ by year-end—Intel at record-low 72.2%. ARM processors now 13.2% of server sales with NVIDIA Grace Blackwell driving 50% growth. EPYC 9005 Turin shipping with up to 192 cores (Zen 5c) and 512MB L3 cache. Intel Xeon 6 Granite Rapids focusing on AI inference acceleration.

AMD's server CPU market share reached 27.8% in Q3 2025, projected to exceed 40% by year-end and potentially reach 50% in 2026.¹ Intel's share fell to 72.2%—the lowest server CPU share the company has ever recorded.² ARM processors now account for 13.2% of total server sales, with NVIDIA's Grace Blackwell combination driving 50% growth in ARM server CPU volume.³ The CPU landscape for AI infrastructure is shifting faster than at any point in the last two decades.

GPUs dominate AI training and inference compute, but CPUs remain essential for orchestration, data preprocessing, serving inference requests, and managing the control plane. The choice of CPU affects memory bandwidth, I/O connectivity, power efficiency, and total system cost. Organizations deploying AI at scale must evaluate CPU selection as carefully as GPU procurement.

AMD EPYC 9005 Turin: the density leader

AMD launched the EPYC 9005 series (codenamed Turin) on October 10, 2024, representing the fifth generation of EPYC server processors built on Zen 5 architecture.⁴ AMD positions these as the world's best server CPU for enterprise, AI, and cloud workloads.⁵

Turin offers two core configurations. Standard Zen 5 models reach 128 cores per socket built on TSMC's 4nm process.⁶ The dense Zen 5c variants deliver 192 cores per socket using TSMC's 3nm process.⁷ The flagship EPYC 9965 provides 192 cores and 384 threads with a 500W TDP.⁸

Key specifications span the full product stack:

The architecture employs up to 16 CCDs (Core Compute Dies) in standard models and up to 12 CCDs in dense configurations, paired with a central I/O die.⁹ Standard CCDs contain eight Zen 5 cores each, while dense CCDs pack 16 Zen 5c cores.¹⁰

Memory support reaches DDR5-6000 with ECC across 12 channels, up from DDR5-4800 in the previous generation.¹¹ PCIe connectivity provides up to 128 Gen5 lanes with CXL 1.0 support for Type 1, 2, and 3 devices.¹² The SP5 socket maintains compatibility with Genoa and Bergamo systems.¹³

Turin processors deliver a 17% IPC uplift over the previous generation with full 512-bit AVX-512 data paths.¹⁴ The improvements translate directly to data preprocessing and inference serving workloads common in AI infrastructure.

AMD's data center business generated $3.7 billion in quarterly revenue, growing 57% year-over-year.¹⁵ The growth reflects both EPYC CPU and Instinct GPU sales as AMD gains share in both categories.

Intel Xeon 6 Granite Rapids: the AI inference advantage

Intel announced the Xeon 6 Granite Rapids 6900P series with models spanning 72 to 128 cores—finally exceeding AMD's core counts for the first time since 2017.¹⁶ The processors use Intel's Redwood Cove P-core architecture on the Intel 3 process node.¹⁷

Granite Rapids-AP (Advanced Performance) models use the Avenue City platform with the larger LGA 7529 socket.¹⁸ The larger socket enables 128 cores with support for 12-channel DDR5 memory and up to 192 lanes of PCIe 5.0 in two-socket configurations.¹⁹ L3 cache reaches an impressive 504 megabytes on top-tier SKUs.²⁰

Memory bandwidth represents a key differentiator. Granite Rapids UCC supports DDR5 at 6.4 GHz standard and multiplexed rank (MRDIMM) memory at 8.8 GHz.²¹ The higher memory speeds benefit memory-bound AI inference workloads.

The I/O architecture provides 136 PCIe 5.0 lanes—up from 128 in Emerald Rapids—with CXL 2.0 Type 3 support and up to 6 UPI links for multi-socket scaling.²²

Intel's distinct advantage lies in Advanced Matrix Extensions (AMX), which accelerate AI inference workloads.²³ The matrix engine supports AMX FP16 acceleration on the Xeon 6500P and 6700P processors.²⁴ Intel claims 5.5 times the AI inferencing performance in ResNet50 compared to AMD's competing 96-core Genoa flagship.²⁵

Benchmark testing confirms the AMX advantage. Phoronix documented massive AI performance benefits with AMX on Granite Rapids, showing significant inference throughput improvements over non-accelerated x86 execution.²⁶

Granite Rapids-D processors target edge computing and networking applications with availability in 2025.²⁷ Initial models reach 42 cores with 72-core variants expected later in the year.²⁸ The SoC variants integrate Intel Ethernet with 3.2 times improved RAN AI performance per core.²⁹

Intel's market share challenges persist despite competitive hardware. The company holds 72.2% of server CPU units but continues losing share quarter over quarter.³⁰ Revenue share tells a different story—AMD captures 37.2% when measuring by dollars rather than units, reflecting AMD's success in higher-priced segments.³¹

NVIDIA Grace: ARM enters the data center

NVIDIA Grace represents the company's first data center CPU, built on ARM Neoverse V2 cores specifically for AI and high-performance computing workloads.³² The architecture pairs with NVIDIA GPUs to create tightly coupled systems that eliminate traditional CPU-GPU communication bottlenecks.

The Grace CPU features 72 high-performance ARM Neoverse V2 cores with 4×128-bit SVE2 vector units per core.³³ Cache hierarchy includes 64KB L1 instruction and data caches, 1MB L2 per core, and 117MB shared L3.³⁴ Memory reaches 480GB of usable LPDDR5X with 546GB/s bandwidth at 250W TDP.³⁵

The Grace CPU Superchip combines two Grace CPUs connected via NVLink-C2C, delivering 144 ARM cores with up to 1TB/s memory bandwidth.³⁶ The interconnect achieves 900GB/s bandwidth between the two CPUs on a single module with up to 960GB LPDDR5X memory.³⁷

NVIDIA claims Grace delivers 2x performance per watt, 2x packaging density, and the highest memory bandwidth compared to contemporary x86 servers.³⁸ The efficiency advantages compound in power-constrained AI deployments.

The GB200 NVL72 configuration connects 36 Grace CPUs with 72 Blackwell GPUs in a rack-scale liquid-cooled design.³⁹ The system delivers 30x faster real-time inference for trillion-parameter large language models compared to previous generations.⁴⁰

Grace integrates fully with the ARM software ecosystem. The NVIDIA HPC SDK and all CUDA components provide ARM-native installers and containers.⁴¹ NVIDIA NIM microservices and NGC containers are optimized for ARM.⁴² All major Linux distributions run without modification.⁴³

ARM's data center momentum extends beyond NVIDIA. ARM Holdings projects its data center CPU market share to increase from approximately 15% in 2024 to 50% by the end of 2025.⁴⁴ The projection reflects aggressive growth from cloud-native ARM instances and NVIDIA's Grace adoption.

The GB10 Grace Blackwell superchip brings the architecture to desktop form factors for AI developers, researchers, and edge computing.⁴⁵ The system-in-package merges an ARM CPU with Blackwell GPU capabilities, enabling local AI development that previously required data center access.

Workload-specific CPU selection

CPU selection for AI infrastructure depends on the specific role within the deployment architecture. Different workloads favor different processor characteristics.

Control plane and orchestration workloads benefit from high core counts and memory capacity. Kubernetes control planes, job schedulers, and monitoring systems scale with available cores. AMD EPYC's 192-core density provides headroom for consolidation. Memory channels and capacity matter more than per-core performance for these workloads.

Data preprocessing pipelines transform raw data into training-ready formats. These workloads often scale with memory bandwidth rather than compute. Intel's MRDIMM support at 8.8 GHz provides bandwidth advantages. The preprocessing stage often runs on dedicated CPU-only systems feeding GPU clusters.

Inference serving workloads present the strongest case for CPU evaluation. While GPUs handle model execution, CPUs manage request routing, tokenization, and response assembly. Intel's AMX acceleration enables CPU-based inference for smaller models, potentially eliminating GPU requirements for appropriate workloads. The 5.5x ResNet50 performance advantage demonstrates the value proposition.

GPU host systems require CPUs that avoid becoming bottlenecks. PCIe lane counts determine how many GPUs connect to each CPU socket. EPYC's 128 Gen5 lanes and Granite Rapids' 136 lanes both support eight-GPU configurations. Memory bandwidth affects how quickly data moves to GPU memory for training batches.

Edge inference deployments favor power efficiency and integrated I/O. Granite Rapids-D integrates Ethernet connectivity for network inference appliances. Grace's ARM architecture provides the efficiency profile edge deployments require.

Infrastructure planning considerations

The CPU market dynamics favor multi-vendor evaluation. AMD's steady share gains create competitive pressure that benefits buyers. Intel's response with Granite Rapids demonstrates continued innovation despite market challenges. NVIDIA's Grace offers differentiation for GPU-centric architectures.

Memory architecture increasingly differentiates platforms. CXL support enables memory expansion beyond socket capacity. DDR5 speeds continue increasing with each generation. Organizations planning multi-year infrastructure should evaluate memory roadmaps alongside CPU specifications.

Power efficiency determines deployment density in constrained facilities. Grace's 2x performance per watt claim merits validation for specific workloads. The power advantage compounds across large deployments where facility limits constrain growth.

Software ecosystem requirements narrow choices for some organizations. x86 compatibility remains essential for legacy workloads. ARM adoption requires application validation and potentially recompilation. CUDA integration with Grace simplifies the transition for GPU-centric deployments.

Total cost of ownership calculations should include not just processor pricing but system costs, power consumption, and licensing. AMD's favorable pricing at the high end—the 192-core 9965 at $14,813—undercuts Intel's comparable offerings. However, Intel's AMX acceleration may reduce GPU requirements for inference, affecting the broader cost equation.

The CPU remains the foundation of AI infrastructure even as GPUs capture attention. Organizations that evaluate CPU selection with the same rigor applied to GPU procurement build more balanced, efficient systems. The competitive dynamics across AMD, Intel, and NVIDIA ensure that careful evaluation yields meaningful infrastructure advantages.

Key takeaways

For server procurement: - AMD EPYC Turin 9965: 192 cores (Zen 5c), 384MB L3, 500W TDP at $14,813; favorable pricing vs Intel at high end - Intel Xeon 6900P: 128 cores (Granite Rapids), 504MB L3, AMX FP16 acceleration, MRDIMM support at 8.8GHz - NVIDIA Grace: 72 ARM cores, 480GB LPDDR5X at 546GB/s bandwidth, 250W TDP; 2x perf/watt claimed

For workload planning: - Control plane/orchestration: favors high core counts (AMD 192-core density); memory capacity matters more than per-core speed - Data preprocessing: scales with memory bandwidth; Intel MRDIMM 8.8GHz provides advantage - Inference serving: Intel AMX delivers 5.5x ResNet50 performance vs AMD; may eliminate GPU requirement for smaller models - GPU host systems: PCIe lane counts (AMD 128, Intel 136 Gen5) determine max GPU count; memory bandwidth affects training batch data movement

For market dynamics: - AMD server CPU share: 27.8% (Q3 2025), projected 40%+ by year-end, potentially 50% in 2026 - Intel: 72.2% unit share but declining; 37.2% revenue share (AMD wins premium segment) - ARM: 13.2% of server sales; Grace Blackwell driving 50% growth in ARM server CPU volume

For infrastructure planning: - CXL support enables memory expansion beyond socket capacity; evaluate memory roadmaps alongside CPU specs - Power efficiency determines deployment density; Grace 2x perf/watt advantage compounds across large facilities - Software ecosystem: x86 compatibility essential for legacy; ARM requires validation but CUDA integration simplifies GPU-centric deployments

For TCO analysis: - AMD: favorable high-end pricing; 192-core 9965 at $14,813 undercuts Intel - Intel: AMX acceleration may reduce GPU requirements for inference, affecting broader cost equation - Grace: LPDDR5X integration reduces memory component costs but requires ARM application validation

References

1. TweakTown. "AMD has close to 40% of the server market, should match (and possibly surpass) Intel by 2026." 2025. https://www.tweaktown.com/news/105716/amd-has-close-to-40-of-the-server-market-should-match-and-possibly-surpass-intel-by-2026/index.html

2. HotHardware. "AMD Data Center Server Market Share Hits New High, Desktop Chips Away At Intel Too." 2025. https://hothardware.com/news/amd-server-revenue-market-share-hits-new-high

3. Semi Engineering. "Data Center CPU Dominance Is Shifting To AMD And Arm." 2025. https://semiengineering.com/data-center-cpu-dominance-is-shifting-to-amd-and-arm/

4. AMD Newsroom. "AMD Launches 5th Gen AMD EPYC CPUs, Maintaining Leadership Performance and Features for the Modern Data Center." October 2024. https://www.amd.com/en/newsroom/press-releases/2024-10-10-amd-launches-5th-gen-amd-epyc-cpus-maintaining-le.html

5. AMD Newsroom. "AMD Launches 5th Gen AMD EPYC CPUs."

6. Phoronix. "AMD Launches EPYC 9005 'Turin' Server Processors Review." October 2024. https://www.phoronix.com/review/amd-epyc-9005

7. Phoronix. "AMD Launches EPYC 9005 'Turin' Server Processors Review."

8. Tom's Hardware. "AMD EPYC 'Turin' 9005 Series - we benchmark 192-core Zen 5 chip with 500W TDP." October 2024. https://www.tomshardware.com/pc-components/cpus/amd-launches-epyc-turin-9005-series-our-benchmarks-of-fifth-gen-zen-5-chips-with-up-to-192-cores-500w-tdp

9. Phoronix. "AMD Launches EPYC 9005 'Turin' Server Processors Review."

10. Exxact. "AMD EPYC 9005 Announcement | 128 Zen 5, 192 Zen 5c Cores." 2024. https://www.exxactcorp.com/blog/news/AMD-EPYC-9005-announcement

11. AMD. "5th Generation AMD EPYC Processors." 2024. https://www.amd.com/en/products/processors/server/epyc/9005-series.html

12. AMD. "5th Generation AMD EPYC Processors."

13. Lenovo Press. "Lenovo ThinkSystem Servers with 5th Gen AMD EPYC 'Turin' Processors." 2024. https://lenovopress.lenovo.com/lp2038-lenovo-thinksystem-servers-with-5th-gen-amd-epyc

14. Phoronix. "AMD Launches EPYC 9005 'Turin' Server Processors Review."

15. WCCFTech. "AMD Manages to Shift The Dynamics of Data Center Markets Into its Favour." 2025. https://wccftech.com/amd-manages-to-shift-the-dynamics-of-data-center-markets-into-its-favour/

16. Tom's Hardware. "Intel Launches Granite Rapids Xeon 6900P series with 128 cores — matches AMD EPYC's core counts for the first time since 2017." September 2024. https://www.tomshardware.com/pc-components/cpus/intel-launches-granite-rapids-xeon-6900p-series-with-120-cores-matches-amd-epycs-core-counts-for-the-first-time-since-2017

17. Wikipedia. "Granite Rapids." 2025. https://en.wikipedia.org/wiki/Granite_Rapids

18. The Next Platform. "Intel Rounds Out 'Granite Rapids' Xeon 6 With A Slew Of Chips." February 2025. https://www.nextplatform.com/2025/02/24/intel-rounds-out-granite-rapids-xeon-6-with-a-slew-of-chips/

19. The Next Platform. "Intel Rounds Out 'Granite Rapids' Xeon 6."

20. Tom's Hardware. "Intel Launches Granite Rapids Xeon 6900P series."

21. The Next Platform. "Intel Shoots 'Granite Rapids' Xeon 6 Into The Datacenter." September 2024. https://www.nextplatform.com/2024/09/24/intel-shoots-granite-rapids-xeon-6-into-the-datacenter/

22. The Next Platform. "Intel Shoots 'Granite Rapids' Xeon 6."

23. Phoronix. "The Massive AI Performance Benefit With AMX On Intel Xeon 6 'Granite Rapids.'" 2024. https://www.phoronix.com/review/intel-xeon-6-granite-rapids-amx

24. The Next Platform. "Intel Shoots 'Granite Rapids' Xeon 6."

25. Tom's Hardware. "Intel Launches Granite Rapids Xeon 6900P series."

26. Phoronix. "The Massive AI Performance Benefit With AMX."

27. ServeTheHome. "Intel Xeon 6 SoC is Here Granite Rapids-D is HUGE." 2025. https://www.servethehome.com/intel-xeon-6-soc-is-here-granite-rapids-d-is-huge/

28. ServeTheHome. "Intel Xeon 6 SoC is Here."

29. Lanner Electronics. "Unlocking the Future: Intel Xeon 6 SoC Powers Next-Gen Network AI Solutions." 2025. https://www.lannerinc.com/news-and-events/eagle-lanner-tech-blog/unlocking-the-future-intel-xeon-6-soc-powers-next-gen-network-ai-solutions

30. Network World. "Despite the hubbub, Intel is holding onto server market share." 2025. https://www.networkworld.com/article/4040551/despite-the-hubbub-intel-is-holding-onto-server-market-share.html

31. TweakTown. "AMD has close to 40% of the server market."

32. NVIDIA. "NVIDIA Grace CPU and Arm Architecture." 2025. https://www.nvidia.com/en-us/data-center/grace-cpu/

33. Glenn Klockwood. "NVIDIA Grace." 2024. https://www.glennklockwood.com/garden/processors/Grace

34. Glenn Klockwood. "NVIDIA Grace."

35. Glenn Klockwood. "NVIDIA Grace."

36. NVIDIA. "Introducing the NVIDIA Grace CPU Superchip." 2024. https://www.nvidia.com/en-us/data-center/grace-cpu-superchip/

37. NVIDIA Developer Blog. "NVIDIA Grace CPU Superchip Architecture In Depth." 2023. https://developer.nvidia.com/blog/nvidia-grace-cpu-superchip-architecture-in-depth/

38. NVIDIA Developer Blog. "Revolutionizing Data Center Efficiency with the NVIDIA Grace Family." 2024. https://developer.nvidia.com/blog/revolutionizing-data-center-efficiency-with-the-nvidia-grace-family/

39. NVIDIA. "NVIDIA Grace CPU and Arm Architecture."

40. NVIDIA. "NVIDIA Grace CPU and Arm Architecture."

41. NVIDIA Developer Blog. "NVIDIA Grace CPU Integrates with the Arm Software Ecosystem." 2023. https://developer.nvidia.com/blog/nvidia-grace-cpu-integrates-with-the-arm-software-ecosystem/

42. NVIDIA Developer Blog. "NVIDIA Grace CPU Integrates with the Arm Software Ecosystem."

43. NVIDIA Developer Blog. "NVIDIA Grace CPU Integrates with the Arm Software Ecosystem."

44. Semi Engineering. "Data Center CPU Dominance Is Shifting To AMD And Arm."

45. SKR TEK. "Nvidia GB10 'Grace Blackwell' Superchip: Desktop AI Redefined." July 2025. https://skrmicrotek.com/2025/07/07/nvidia-gb10-grace-blackwell-superchip-desktop-ai-redefined/

SEO Elements

Squarespace Excerpt (159 characters): AMD EPYC Turin: 192 cores at $14,813. Intel Xeon 6: 5.5x AI inference boost. NVIDIA Grace: 2x perf/watt. Server CPU landscape for AI infrastructure in 2025.

SEO Title (55 characters): CPUs for AI Infrastructure: EPYC vs Xeon vs Grace 2025

SEO Description (155 characters): AMD reaches 27.8% server share. Intel counters with AMX AI acceleration. NVIDIA Grace brings ARM to data centers. Analysis of CPUs for AI infrastructure.

URL Slugs: - Primary: cpus-for-ai-infrastructure-epyc-xeon-grace-server-processors - Alt 1: amd-epyc-9005-turin-192-core-server-cpu-2025 - Alt 2: intel-xeon-6-granite-rapids-amx-ai-inference - Alt 3: nvidia-grace-arm-cpu-data-center-ai