CPU for a 2-, 4- or 8-GPU server: AMD EPYC or Intel Xeon, cores, lanes, memory channels

The CPU a GPU server needs, by card count

A server with 2, 4 or 8 GPUs such as the NVIDIA RTX PRO 6000 Blackwell Server Edition needs two processors with at least six physical cores per GPU, a base clock of 2.0 GHz or more and PCIe Gen5, by NVIDIA's configuration guide for NVIDIA-Certified Systems. AMD EPYC 9005 and Intel Xeon 6 both meet that, and for most inference servers the core count is easy to reach. The choice between them comes down to memory channels, PCIe lanes per socket for GPUs, network cards and NVMe drives, and often the server model itself.

NVIDIA-Certified Systems Configuration Guide, Table 1, last updated 30 September 2026; NVIDIA RTX PRO AI Factory Enterprise Reference Architecture, components, 18 May 2026; both read on 24 October 2026. Totals are our arithmetic for RTX PRO 6000 Server Edition cards (96 GB, up to four MIG instances each).

AMD product pages for the EPYC 9135, 9355 and 9555 and AMD EPYC 9005 datasheet (PID 242937604-G, November 2025); Intel product specifications for the Xeon 6747P and 6960P; all read on 24 October 2026. All five support two sockets. The Xeon 6960P also takes MRDIMMs at 8800 MT/s.

The rest of this page goes through the three NVIDIA documents, the cores they add up to, PCIe lanes, memory channels and the choice between the two platforms.

What NVIDIA's three documents ask for

NVIDIA publishes three sets of host rules for PCIe GPU servers. They agree on the base clock and on Gen5 x16 per GPU, and differ on sockets and cores.

NVIDIA-Certified Systems Configuration Guide (30 September 2026); RTX PRO AI Factory RA, components (18 May 2026); NVIDIA AI Enterprise reference architecture, compute node hardware (undated); read on 24 October 2026. All three ask for a 2.0 GHz minimum clock.

The certified-systems guide names "NVIDIA Grace, Intel Xeon scalable processor (Emerald Rapids) or AMD Turin" as examples of "PCIe Gen5 (or later when available) capable CPUs". The two reference architectures list Intel Emerald Rapids, Sapphire Rapids, Granite Rapids and Sierra Forest and AMD Genoa and Turin. Both lists are examples; the processors compared on this page are PCIe 5.0 parts by AMD's and Intel's specifications.

The guide asks for GPUs "evenly distributed across CPU sockets and PCIe root ports" and for "NICs and NVMe drives" placed "within the same PCIe switch or root complex as the GPUs". The H200 NVL architecture, which NVIDIA calls "PCIe Optimized 2-8-5", adds that NVLink-bridged pairs should sit under the same socket; pairing across sockets is "acceptable but not recommended". NVIDIA presents all these values as starting points to adapt to the workload.

Cores per GPU for 2, 4 and 8 cards

The serving engine itself needs few cores. vLLM counts one API server process, one engine process and one worker per GPU, so N GPUs need at least 2 + N physical cores, as our guide to CPU, RAM and storage for a 4-GPU server works out from vLLM's tuning guide. NVIDIA's six or seven cores per GPU cover the rest of the host: tokenisation, document parsing for RAG, a vector database, monitoring agents and, in fine-tuning, the data loader.

The figure grows with virtualisation and MIG. Each RTX PRO 6000 Server Edition splits into up to four MIG instances, and the reference architecture asks for two cores per instance on top of the seven per GPU and the two per GPU for the operating system or hypervisor. By our arithmetic, four cards fully split into 16 instances need 28 + 32 + 8 = 68 cores, and eight cards with 32 instances 136. For the H200 NVL, which splits into up to seven instances, four fully split cards need 28 + 56 + 8 = 92 cores.

NVIDIA's own eight-GPU systems sit above both rules. Its DGX H200 has 112 cores for eight GPUs, 14 per GPU, and its DGX B300 two Xeon 6776P with 64 cores each, 16 per GPU, as our 4-GPU guide lists from the DGX user guides. These are SXM training systems; an inference server with one model copy per card needs fewer.

NVIDIA documents above; vLLM tuning guide and ZeRO-Offload as cited in our CPU, RAM and storage guide. Assignment of workloads to rules is our recommendation.

PCIe lanes: GPUs, network cards and NVMe

Each GPU needs its own Gen5 x16 link, by all three documents. Two cards take 32 lanes, four 64 and eight 128. Network cards take x16 each and NVMe drives x4.

AMD's datasheet gives each EPYC 9005 "128 PCIe® Gen 5 I/O lanes (up to 160 in 2-socket servers)", because in a two-socket server part of each processor's lanes links the two sockets. Intel gives the Xeon 6747P 88 lanes and the Xeon 6960P 96; Intel's processors link their sockets over separate UPI links, so two of them give 176 or 192 lanes by our arithmetic.

Four cards fit either platform with room to spare. Four GPUs, two network cards and four NVMe drives take 64 + 32 + 16 = 112 lanes, below one EPYC 9005's 128, so even one processor has the lanes. The certified-systems guide still starts from two sockets, and two give each processor two GPUs, one network card and its own drives, as the placement rule asks.

Eight cards do not fit directly on most servers. The 2-8-5-200 layout of NVIDIA's RTX PRO reference architecture adds five network adapters to the eight GPUs, which would be 128 + 80 = 208 lanes if every device takes x16, above what two processors of either platform provide. Eight-card servers therefore put PCIe switches between the processors and the cards, or ConnectX-8 SuperNICs with a built-in switch, as our article on the NVIDIA RTX PRO Server explains. Behind a switch, the processor's lane count matters for the uplinks, and the switch layout matters for traffic between the cards. Our page on how many GPUs fit in one server covers the lane budget for larger builds.

We build servers with two, four or eight cards on either platform and lay out the slots per socket before we quote; send us the cards, network speed and drives you need, and our team replies with a configuration and the PCIe layout.

Memory channels: EPYC 9005 and Xeon 6

Memory channels decide how system memory reaches the certified guide's twice-the-GPU-memory value with one module per channel, which runs at the highest speed. The reference architecture asks for "at least one DIMM per channel".

AMD EPYC 9005 datasheet and product pages (12 channels per socket); Intel specifications for the Xeon 6960P (12 channels) and 6747P (8 channels); capacities are our arithmetic.

On two EPYC 9005, the three capacities that equal twice the memory of 2, 4 and 8 RTX PRO 6000 Server Edition, 384, 768 and 1,536 GB, each fill all 24 channels with one module. On two Xeon 6700P, 1,536 GB for eight cards also fills every channel once, with 96 GB modules, while 768 GB for four cards does not come out evenly with 32 or 64 GB modules; the choice is then 1,024 GB, or a second module per channel at a lower speed. Dell and HPE give 6400 MT/s at one module per channel and 5200 MT/s at two for their Xeon 6 servers, as our guide to how much RAM a GPU server needs lists.

The board matters as much as the processor. Cisco's UCS C845A M8 with EPYC 9005 wires eight channels per processor with two slots each, per the same guide, so check the slot count and speed table for the exact server.

Clock speed, core count and power

All five processors in the table clear the 2.0 GHz base clock. The limit matters at the top of the core range: Intel lists its 86-core Xeon 6788P at 2.00 GHz, exactly at the minimum. AMD's 64-core EPYC 9555 has a base clock of 3.2 GHz and a boost of up to 4.4 GHz, its frequency-optimised sibling EPYC 9575F 3.3 GHz and up to 5 GHz; the 16-core EPYC 9135 has 3.65 GHz base and 4.3 GHz boost.

More cores raise the processor's power, and the server's power with it. The EPYC 9135 has a default TDP of 200 W, the 9355 280 W, the 9555 360 W and the 9575F 400 W; Intel gives 330 W for the Xeon 6747P and 500 W for the 6960P. Our wall-power estimates add the cards, the processors and 15 per cent for the rest of the server.

Our estimate: board power from NVIDIA's product brief, default TDP from AMD's product pages, plus 15 per cent for memory, drives, network cards, fans and conversion losses.

These are the figures of our eight-card guide and our other server pages, which use the same processors. On a four-card server the processors can decide the feed. At the cards' 450 W mode, four cards with two EPYC 9355 come to about 2.7 kW, below the 2.94 kW we plan per 230 V 16 A feed; with two 400 W EPYC 9575F the same server reaches about 3.0 kW, above it, by the same estimate; our article on GPU server power on 16 A, 32 A and three-phase does the feed arithmetic.

What servers and test systems ship with

Published systems show what makers and test labs pair with these cards.

StorageReview HPE ProLiant DL380a Gen12 review (6 November 2025, modified 14 October 2026); MLPerf® Inference: Datacenter v6.0, closed division, Llama 2 70B (llama2-70b-99), entry 6.0-0021, retrieved from mlcommons.org on 24 September 2026, result verified by MLCommons Association; Helix engineering blog, 14 August 2026; NVIDIA DGX B300 user guide as cited in our 4-GPU guide. The MLPerf name and logo are registered and unregistered trademarks of MLCommons Association in the United States and other countries. All rights reserved. Unauthorized use strictly prohibited. See www.mlcommons.org for more information.

Helix's server with 256 GB is far below twice the GPU memory and runs one model on bare metal; it shows that the rule is a starting point, not a requirement of the software. The server models also fix the platform: Dell's PowerEdge XE7740 takes Xeon 6 and the XE7745 EPYC 9005, HPE's ProLiant Compute DL380a Gen12 takes two Xeon 6 and the DL385 Gen11 EPYC, Cisco's UCS C845A M8 AMD Turin.

EPYC or Xeon: a recommendation per use case

Our recommendation from the rules, specifications and shipped systems above.

For two and four cards we prefer two EPYC 9005, because their 24 channels reach the 2× value with one module per channel and each socket has the lanes for its GPUs, network card and drives. On a Xeon 6700P platform, plan the memory population first; the cores and lanes are sufficient for four cards. For eight cards both platforms work behind PCIe switches, and the server model, its inlet limits and its power supplies weigh more than the processor brand.

Our examples cover the base case and leave headroom: two EPYC 9135 give 32 cores for two cards (12 by the guide, 18 by the RA with an operating system), two EPYC 9355 give 64 for four (24 and 36) and two EPYC 9555 give 128 for eight (48 and 72). Fully MIG-split servers need 34, 68 and 136 cores, so for those we quote the next larger processor. Tell us whether the server will run VMs, MIG or fine-tuning, and we size the processors for it.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We build AI servers to order with two, four or eight RTX PRO 6000 Server Edition, RTX PRO 4500 Server Edition or H200 NVL cards on AMD EPYC or Intel Xeon processors, with DDR5 ECC memory populated evenly across the channels. Our reference configurations follow NVIDIA's certified-systems guide: two AMD EPYC processors, at least six cores per GPU and at least twice the GPU memory. Each server is assembled and tested before delivery, and every component carries the manufacturer warranty on one EU contract and invoice.

FAQ