Server Processor (CPU) Buyer's Guide

Workload Classification — Frequency, Cores, or Both

The first question for any CPU purchase is whether your workload is frequency-bound, core-bound, or balanced. Frequency-bound workloads benefit from clock speed: single-threaded transactional databases, VDI host-side processing, latency-sensitive trading systems. Core-bound workloads benefit from core count: virtualization hosts (especially with many small VMs), Kubernetes worker nodes, video encoding, ML training. Balanced workloads benefit from both.

Modern Intel Xeon and AMD EPYC SKU tables let you pick the trade-off. The same generation typically offers a high-frequency 16-core part at 3.5 GHz, a balanced 32-core at 2.9 GHz, and a high-core 64-core at 2.4 GHz. Pick the spot that fits your workload, not the spot with the highest core count.

For VMware vSphere with per-core licensing, the math changes: each additional core costs $190 in VVF or $350 in VCF licensing. A high-frequency 16-core SKU may outperform a 32-core SKU on frequency-bound workloads while costing $3,000/year less per host in licensing.

Socket and Chipset — Match the Motherboard

CPU socket compatibility is the first hard filter. Intel Xeon Scalable 1st-4th generations use LGA 3647 / LGA 4189 / LGA 4677 sockets respectively; each socket family covers specific CPU generations. AMD EPYC uses SP3 (1st-3rd gen) and SP5 (4th-5th gen).

BIOS revision matters more than people realise. A motherboard with the right physical socket may not POST with a newer-generation CPU until the BIOS is flashed to the version that supports it. Best practice: flash the BIOS to current before swapping CPUs.

Within a socket family, microcode and firmware support determine which steppings boot. Older motherboards may need a BIOS update to recognise newer SKUs. Pro Disk Network ships CPUs labelled with the QVL-validated motherboard families and recommended BIOS minimum versions where known.

AMX, AVX-512, and Hardware Acceleration

Recent Intel Xeon generations (Sapphire Rapids, Emerald Rapids, Granite Rapids) include AMX (Advanced Matrix Extensions) silicon for INT8 inference acceleration. Workloads using ONNX Runtime, OpenVINO, or oneAPI inference frameworks see 4-6x throughput uplift on BERT and 2-3x on small Llama models.

AVX-512 is the broader vector instruction set that benefits scientific computing, encryption, and video processing. Most modern Xeon SKUs support AVX-512; consumer Intel parts disable it.

For AMD EPYC: AVX-512 is supported on Genoa (4th gen) and newer; Milan (3rd gen) supports AVX2 but not 512. AMD does not ship AMX; for AI inference workloads, the Intel decision is clearer.

For pure virtualisation hosts running general workloads, AMX and AVX-512 add minimal value — pick the SKU that fits your workload otherwise.

EVC Mode and Cluster Homogeneity

vSphere clusters with vMotion require CPU compatibility across all nodes. EVC (Enhanced vMotion Compatibility) mode sets a baseline of CPU instructions visible to VMs, ensuring a VM can migrate from any node to any other node without instruction-set surprises.

Adding a newer-generation CPU to an existing cluster requires setting EVC mode to the older baseline, which disables newer instructions on the new CPU. This is sometimes the right trade-off (compatibility wins over peak performance), but plan EVC at cluster creation, not afterward — changing EVC mode requires powering off all VMs.

For Hyper-V clusters with Live Migration, the compatibility constraint is similar but slightly more relaxed. The "Migrate to a physical computer with a different processor version" option allows cross-generation migration at the cost of some advanced instructions.

For Kubernetes / OpenShift, CPU homogeneity is less critical because pods are stateless. CPU heterogeneity matters mostly for resource-request planning when different nodes have different core counts.

New, Bulk, or Refurbished — Procurement Strategy

New (sealed OEM) CPUs are the only choice for tier-1 production databases and HA cluster build-outs. Premium runs 15-25% over bulk-OEM. Full OEM warranty (typically 5 years) applies.

New (bulk OEM) CPUs are factory-fresh tray units sold without retail packaging or heatsink. Pro Disk Network warranties bulk OEM for 36 months. Cost-optimal for hyperconverged clusters and general-purpose virtualisation.

Refurbished CPUs are pulled from decommissioned servers, burn-in tested at full TDP, and carry a 12-month advance-replacement warranty. Cost savings: 50-65% vs new. Appropriate for lab, dev/test, capacity expansion, and EOL refresh cycles where the underlying platform has 12-24 months of remaining service life.

For matched-pair dual-socket configurations, Pro Disk Network ships sequential-serial pairs from the same lot for free on orders of 4+ pairs. This eliminates stepping-mismatch issues that occasionally cause subtle stability problems in production.