Partner POV | Advancing Agentic Workflows With AMD EPYC 9006 Series Server CPUs

This article was written by Madhu Rangarajan and Zainab Al-Shamma and provided by AMD.

Purpose-built compute across the agentic workflow 

An agentic workflow is not a single workload running on a single type of system. For infrastructure leaders, that means processor selection can no longer be treated as a single decision across the data center. The workflow spans agent execution, accelerator hosting, and the enterprise services agents use to complete work, with each placing different demands on the CPU. The AMD EPYC 9006 Server CPUs portfolio addresses those demands with four purpose-built processor families: AMD EPYC 9006 SP7, AMD EPYC 9006 SP8, AMD EPYC 9006X SP7 and AMD EPYC 9006 LP Server CPUs. Each family includes multiple SKUs optimized for different workloads and infrastructure constraints. Each purpose-built CPU family is matched to the work it serves best across the agentic workflow and the cloud, enterprise and HPC infrastructure running alongside it.

AMD EPYC 9006 SP7 Server CPUs: Leadership per-Core performance, built to scale  

Scale agents, feed accelerators, and power the workloads agents rely on

For agent sandbox execution. Large-scale agent execution means running thousands of concurrent agents, executing the code and tasks they spawn and holding system state and memory at scale, all of which require maximum thread density. Meeting that demand takes a CPU built for throughput and concurrency above all.  

• Responsive agent execution: Leadership performance per core with simultaneous multithreading (SMT) for concurrency and utilization
• Maximum agent capacity: Highest thread density1, with up to 256 cores and 512 threads
• Efficient agent scale: The most agents per watt, per dollar, and per rack2,3,4

With these capabilities, AMD EPYC 9006 SP7 Server CPUs give infrastructure leaders both responsive per-agent performance and the aggregate capacity to run massive agent fleets efficiently. 

For the AI host node. On a GPU server, keeping accelerators fully utilized is what turns their raw compute power into real throughput. That requires strength from each CPU core as well as the capacity to scale overall throughput. In high-frequency configurations designed for AI host nodes, AMD EPYC 9006 SP7 Server CPUs deliver: 

• Strong host-node performance: Leadership performance per core and frequencies up to 5GHz
• Faster CPU-to-GPU data movement: PCIe Gen 6 for 2X the bandwidth per lane versus the prior generation
• Sustained accelerator utilization: high memory bandwidth to keep accelerators fed

For general-purpose servers. Modernizing an AI-first data center means increasing workload density, creating room for new agentic services, and supporting a diverse set of workloads with different constraints through a broad range of SKUs. That takes flagship general-purpose performance and high thread count. 

• Higher VM and agent density per rack: highest thread density
• Lower TCO through consolidation: More workloads on fewer servers5

For hyperscalers and large enterprises, that consolidation helps reclaim floor space, power, and cooling capacity for growing agentic workloads. 

AMD EPYC 9006 SP8 Server CPUs: Right-sized enterprise compute for ROI 

Balance compute, memory and I/O to scale agents and power enterprise workloads

For agent sandbox execution. Not every agentic deployment needs 256 cores; edge sites, smaller clusters, and power-constrained racks prize efficiency over maximum density. What they need is a right-sized, low-power option that keeps cost per agent low. 

• Right-sized agent execution: Flexible 8-to-128-core options
• Efficient deployment: Low-power configurations matched to workload needs

The AMD EPYC 9006 SP8 Server CPUs family complements the AMD EPYC 9006 SP7 Server CPUs family, giving customers a practical way to scale concurrent agents at the core count and power envelope that fits each deployment, while helping improve power efficiency and cost per agent across diverse infrastructure constraints. 

For general-purpose servers. Enterprises run a wide mix of workloads — databases, storage, application services, and AI-support tasks — each with different demands. The challenge is to fit the right amount of compute, memory, and I/O to each one, so every server is sized to the job it does.  

AMD EPYC 9006 SP8 Server CPUs provide a balanced design across compute, memory, and I/O, with flexible core counts that help customers right-size each server and improve their return on infrastructure investment.  

• Balanced enterprise compute: A balanced design across compute, memory, and I/O
• Enterprise ROI: Leadership performance per system dollar, right-sized cores, and multithreading can help reduce software licensing costs

AMD EPYC 9006X SP7 Server CPUs: Fast time-to-insight for HPC and technical computing 

Accelerate memory-sensitive workloads with high cache and bandwidth

For HPC & technical computing. Memory-sensitive and data-intensive workloads — simulation, modeling, large-scale analytics, and large-context reasoning — stall when the processor waits on memory instead of computing. The fix is to keep large working sets close to the cores with far more cache and bandwidth.  

AMD EPYC 9006X SP7 Server CPUs are built to accelerate memory-sensitive workloads with:

• Larger working sets closer to the cores: AMD 3D V-Cache technology delivers up to 3X the L3 cache per core versus comparable AMD EPYC 9006 SP7 SKUs6
• Faster latency-sensitive processing: Frequencies above 5GHz
• Higher data throughput: Up to 1.6 TB/s of memory bandwidth

Together, these capabilities keep data close to the cores to accelerate time-to-insight across simulation, modeling, large-scale analytics, retrieval, and large-context reasoning. It is the portfolio's CPU family for the most bandwidth- and cache-bound work, agentic or otherwise. 

AMD EPYC 9006 LP Server CPUs: Powering the next generation of rack-scale AI 

Drive high AI throughput across the rack

For the AI host node. The next wave of AI infrastructure is being designed at the scale of the rack, not the server. These dense, accelerator-rich systems need a host node that sustains high AI throughput across the whole rack while keeping every GPU fed.  

 AMD EPYC 9006 LP Server CPUs are purpose-built for rack-scale AI host nodes with: 

• Faster CPU to GPU Connectivity: with 112 Gb/s xGMI
• High-frequency host processing: Frequencies up to 5GHz
• Power-efficient memory: LPDDR5X support

For infrastructure leaders designing rack-scale AI, AMD EPYC 9006 LP Server CPUs bring these capabilities together to help sustain accelerator utilization and AI throughput as deployments grow from a single node to an entire AI factory. 

Match the CPU to the work  

The portfolio is deliberately built around four processor families, each offering multiple SKUs optimized for different workloads and infrastructure requirements: 

• AMD EPYC 9006 SP7 Server CPUs: Leadership per-core performance and scale for agent sandbox execution, AI host nodes, and high-thread density general-purpose computing
• AMD EPYC 9006 SP8 Server CPUs: Right-sized efficiency for agent sandboxes and mixed enterprise workloads
• AMD EPYC 9006X SP7 Server CPUs: High-cache performance for memory-sensitive HPC and technical computing
• AMD EPYC 9006 LP Server CPUs: Optimized AI host node for next-generation rack-scale AI systems

AMD partners are already building systems, software and memory around this portfolio. Here is what leading memory partners see. 

Our collaboration with AMD showcases how tightly integrated CPU, memory, and storage innovations unlock next-generation performance. By pairing the latest high-capacity Micron server DRAM and industry-first PCIe Gen6 SSDs with the AMD EPYC 9006 Series Server CPUs, we are enabling agentic AI infrastructure that scales seamlessly, delivers faster insights, and accelerates the most demanding data center workloads. - Alvaro Toledo, VP and GM of the Americas Core Data Center Business, Micron

As agentic AI scales, the whole memory and storage path has to keep up with the CPU. Working with the AMD EPYC 9006 Series Server CPUs, we can help customers move data to compute faster and sustain the throughput these workloads demand — from real-time services to the largest data-intensive jobs. - Sang Joon Hwang EVP, Leader of Memory Product & Technology, Samsung

As a key partner to AMD, SK Hynix affirms that memory bandwidth is the decisive factor for agentic and data-intensive workloads. Deploying our memory alongside the AMD EPYC™ 9006 Series ensures cores and accelerators are continuously supplied with data, enabling customers to maximize the productive output of every server. - Sang Kwon Lee, VP DRAM Product Planning, SK Hynix

Together, we advance Agentic AI 

Agentic AI is an orchestrated workflow, spanning agent execution, accelerator inference and the enterprise workloads underneath. Each places different demands on the CPU, making the right processor choice essential at each stage. The AMD EPYC 9006 Server CPUs portfolio provides four purpose-built processor families, each with multiple SKUs optimized for those distinct roles, supported by a broad ecosystem. This gives customers the flexibility to match the right CPU to the workload across the agentic AI workflow, without adding integration complexity. With AMD EPYC 9006 Series Server CPUs, the agentic foundation is ready when you are.