CPU Platforms: AMD EPYC vs. Intel Xeon
CPU Platform Comparison (2026)
| Dimension | AMD EPYC (Genoa/Turin) | Intel Xeon (Sapphire Rapids/Granite Rapids) |
|---|---|---|
| Core count (per socket) | Up to 128 cores | Up to 60 cores |
| Memory channels | 12 channels DDR5 | 8 channels DDR5 |
| PCIe lanes | 128 PCIe 5.0 lanes | 80 PCIe 5.0 lanes |
| Memory bandwidth | 460 GB/s | 307 GB/s |
| Best for | High core count, memory bandwidth, cloud workloads | Single-threaded performance, AI acceleration (AMX), specific ISV certifications |
| Typical use case | Virtualization hosts, databases, HPC | ERP, financial applications, ISV-certified workloads |
Platform selection should be driven by workload requirements and ISV certifications: not by general preference. Some enterprise applications (SAP, Oracle) have specific certification requirements that may constrain platform choice.
GPU Infrastructure for AI Workloads
NVIDIA H100 and H200 GPUs are the current standard for enterprise AI training and inference. Each GPU has 80–141 GB of HBM3/HBM3e memory and 3.35–4.8 TB/s of memory bandwidth, specifications that enable the large model training and inference workloads that CPU infrastructure cannot support.
Single GPU Server (1x H100)
4-GPU Server (4x H100)
8-GPU Server (8x H100, DGX H100)
GPU Cluster (8+ DGX nodes)
Facility readiness before GPU deployment
Memory Architecture
DDR5 is the current standard for enterprise server memory. It provides 1.5–2x the bandwidth of DDR4 at the same capacity: which directly benefits memory-bandwidth-bound workloads like databases, analytics, and AI inference. The transition from DDR4 to DDR5 is one of the most significant performance improvements available in a compute refresh.
For in-memory database workloads (SAP HANA, Oracle TimesTen), memory capacity is the primary constraint. 4-socket and 8-socket servers with 3–12 TB of RAM are required for large in-memory databases. These configurations require specialized server platforms and are not available from all OEMs.
Storage Interfaces
Server Storage Interface Comparison
| Interface | Max Throughput | Latency | Best For |
|---|---|---|---|
| NVMe (PCIe 5.0) | 14 GB/s per drive | ~70 microseconds | Databases, AI checkpointing, high-IOPS workloads |
| NVMe (PCIe 4.0) | 7 GB/s per drive | ~100 microseconds | General high-performance storage |
| SAS 24G | 2.4 GB/s per drive | ~1 millisecond | High-capacity spinning disk, tape |
| SATA 6G | 600 MB/s per drive | ~1 millisecond | Cold storage, backup targets |
Virtualization Considerations
Virtualization adds overhead: typically 5–15% for CPU and memory, and higher for storage and network I/O. For most enterprise workloads, this overhead is acceptable. For latency-sensitive workloads (financial transaction processing, real-time analytics) and compute-intensive workloads (AI training, scientific computing), bare-metal deployment may be required to achieve performance targets.
GPU passthrough: assigning a physical GPU directly to a VM: reduces virtualization overhead for GPU workloads but limits the flexibility of the virtualization platform. GPU partitioning (NVIDIA MIG) allows a single GPU to be divided into multiple isolated instances, enabling multi-tenant GPU deployments without the overhead of full virtualization.
Server Form Factors
Tower
Small offices, edge locations, development environments
No rack required. Limited scalability. Not appropriate for data center deployments.
1U/2U Rack
General enterprise workloads, virtualization hosts, web servers
Standard data center form factor. High density. Wide OEM selection.
4U Rack (GPU)
AI training, GPU-accelerated workloads
Required for 4–8 GPU configurations. High power density (10–30 kW/rack). Requires liquid cooling at full density.
Blade
High-density virtualization, HPC
Shared chassis reduces cabling. Vendor lock-in. Less common in new deployments.
Modular/Sled
Hyperscale, cloud-native
Open Compute Project (OCP) designs. Lower cost at scale. Limited OEM support.