Aura Digital
Sofia · 42.69°N 23.32°E
Get a quote
All articles
Guides · November 16, 2026 · 6 min read

AMD EPYC vs Threadripper for AI: lanes, memory, sockets

EPYC and Threadripper compared for AI and multi-GPU machines: PCIe lanes, memory channels, ECC, sockets, cores and power, from AMD specs, November 2026.

EPYC and Threadripper use the same "Zen" core family but target different machines, rack servers and desktop workstations. This comparison covers what that changes for an AI or multi-GPU build, using AMD's and NVIDIA's published specifications as of November 2026.

Which one for whom

  • One to four GPUs at a desk, with up to 1 TB of RAM: Threadripper 9000 on TRX50. It gives up to 80 PCIe 5.0 lanes and 64 cores. The 9980X lists at $4,999.
  • Up to eight GPU links in a workstation, with up to 2 TB: Threadripper PRO 9000 WX on WRX90. It gives 128 PCIe 5.0 lanes and 8 memory channels. The 96-core 9995WX lists at $11,699.
  • A rack server with eight GPUs or more than 2 TB of RAM: EPYC 9005 on SP5, in a high-frequency part. It adds a second-socket option, 12 channels and server reliability features. The 64-core 9575F lists at $9,238 at 1,000-unit pricing.
  • A new rack build that needs PCIe 6.0 or the most memory bandwidth: EPYC 9006 on SP7. It has 16 channels and up to 1,638 GB/s per socket. The 96-core 9686F lists at $11,434 at 1,000-unit pricing.

The trade-off: Threadripper stops at one socket and 2 TB. EPYC parts are built for server enclosures and rack airflow. Prices are AMD's CPU list figures, not street prices.

AMD EPYC vs Threadripper at a glance

Threadripper (TRX50) Threadripper PRO (WRX90) EPYC 9005 (SP5) EPYC 9006 (SP7)
Max cores per CPU 64 96 192 256
Sockets 1 1 1 or 2 1 or 2
PCIe Up to 80 Gen 5 Up to 128 Gen 5 128 Gen 5; up to 160 in 2P Gen 6 x96
Memory channels 4 8 12 16
Memory speed DDR5-6400 RDIMM DDR5-6400 RDIMM Up to DDR5-6400 RDIMM 8000, MRDIMM 12800
Max memory 1 TB 2 TB 6 TB per socket 4 TB per socket (16 x 256 GB)
Top-part power 350 W 350 W 500 W 600 W

Sources: AMD's Threadripper platform table, the EPYC 9005 data sheet, the EPYC 9996 page and the EPYC 9006 white paper. AMD notes that 9006 specifications are subject to change.

The PCIe and memory rows usually decide a multi-GPU build. Core count rarely does.

Threadripper vs EPYC on PCIe lanes: count x16 links

NVIDIA's configuration guide, updated in October 2026, asks for a Gen 5 x16 link per GPU for RTX PRO 6000 and H200 NVL. Dividing lanes by 16 gives a ceiling: 5 links on TRX50, 8 on WRX90 or a single EPYC 9005, and 10 on a two-socket EPYC 9005.

Real builds get fewer. NICs and NVMe drives use lanes too, and the motherboard decides how many x16 slots exist. Physical width also matters: a workstation RTX PRO 6000 is dual slot. NVIDIA says a PCIe switch may be optional for cost-effective inference servers, and it wants NICs under the same switch or root complex as the GPUs.

The EPYC 9006 product pages list PCIe 6.0 x96, which is 6 x16 links per socket. AMD puts Gen 6 at 64 GT/s per lane, twice the Gen 5 rate. Read the lane map of the specific board before planning around it.

Memory channels: EPYC leads on bandwidth and capacity

AMD lists per-socket bandwidth of 614 GB/s for the EPYC 9965 and 1,024 or 1,638 GB/s for the EPYC 9996, depending on the DIMM type. The same arithmetic of channels times 8 bytes times MT/s gives about 410 GB/s for WRX90 and 205 GB/s for TRX50.

Capacity comes first for most GPU hosts. NVIDIA's guide asks for at least twice the total GPU memory, spread evenly across sockets and channels, while its RTX PRO reference design asks for at least 128 GB per GPU. With 96 GB cards, the twofold rule means 768 GB for four cards, within TRX50's 1 TB, and 1,536 GB for eight, within WRX90's 2 TB.

Bandwidth matters most when work reaches the CPU side. That includes hybrid CPU and GPU inference in llama.cpp and KV cache offload to host memory. In one SemiAnalysis B200 run, nearly 20% of cache hits came from offload, on top of a 73% HBM hit rate.

ECC and RDIMM: both platforms have it, EPYC goes further

Both Threadripper lines use RDIMM, and AMD lists ECC as enabled by default on the 9995WX and the 9980X. The EPYC 9996 page lists RDIMM and MRDIMM.

EPYC adds Advanced Memory Device Correction, a form of ECC that AMD compares to Chipkill and that can keep a DIMM with errors in service. AMD also says EPYC platforms can hot swap PCIe and NVMe devices. Threadripper offers something EPYC pages do not list: overclocking support. AMD's product pages add that its warranty does not cover overclocking damage.

Multi-socket: only EPYC has it, and the links use lanes

An earlier AMD Threadripper PRO brief treats the line as a single-socket workstation CPU, and the platform table has no two-socket option.

On EPYC 9005, the sockets talk over three or four 16-lane links. Four links give up to 512 GB/s between CPUs, and three free 16 lanes per CPU for the 160-lane total. NVIDIA's guides ask for GPUs spread evenly across sockets and root ports. AMD says most applications run well with each socket as one NUMA domain (NPS=1).

NVIDIA's general guide sets a minimum of two sockets for its certified inference and training servers. The RTX PRO reference design accepts one socket, and its CPU list names AMD Genoa and Turin, which are EPYC generations.

Cores: both platforms cover GPU hosting

The RTX PRO design asks for at least 7 physical cores per GPU, plus 2 per GPU for the OS or virtualization. That comes to 36 cores for four GPUs and 72 for eight, at a 2.0 GHz clock or better. A 64-core 9980X covers the first case and a 96-core 9995WX covers the second. A single 64-core 9575F would fall short of the second.

AMD positions high-frequency EPYC parts as AI host CPUs and aims the 256-core parts at running many agents. The 9575F has a 3.3 GHz base and boosts to 5 GHz. The 9965 has a 2.25 GHz base with up to 3.7 GHz. The 9995WX has a 2.5 GHz base and boosts to 5.4 GHz.

Form factor and power: the GPUs dominate

The CPU is the smaller share of the power budget. Threadripper parts default to 350 W with no cooler in the box. The EPYC 9965 defaults to 500 W, and the EPYC 9996 to 600 W.

NVIDIA lists three RTX PRO 6000 variants: a 600 W workstation card, a 300 W Max-Q card and a passively cooled server card at 400-600 W. Passive cards rely on chassis airflow, which points to a server enclosure.

Four 600 W cards plus a 350 W CPU reach 2,750 W before drives, fans and PSU losses. A 15 A, 120 V circuit carries 1,800 W, so that build needs a dedicated higher-power circuit. Four Max-Q cards with the same CPU come to 1,550 W.

FAQ

Can a Threadripper PRO CPU go on a TRX50 board?

Yes. AMD's platform table lists Threadripper PRO 9000 WX for both TRX50 and WRX90. On TRX50, the platform limits apply: 4 memory channels and up to 80 PCIe 5.0 lanes. Non-PRO Threadripper 9000 parts are listed only for TRX50.

Does a second EPYC CPU double the PCIe lanes?

No. A two-socket EPYC 9005 server has up to 160 PCIe 5.0 lanes, against 128 with one CPU, because some of each CPU's lanes become links between the sockets. By the same arithmetic, four links instead of three leave 128 lanes in total.

What does the "P" suffix on an EPYC part mean?

AMD says "P" parts are optimized for single-socket servers, with all SERDES links dedicated to PCIe. Parts without the suffix work in one-socket and two-socket designs. The EPYC 9005 data sheet marks 9655P as 1P only.

Is DDR5-6400 guaranteed on EPYC 9005?

No. AMD's product page limits DDR5-6400 to targeted customers and configurations, and it requires OEM enablement and a BIOS update. The data sheet ties 6400 MT/s to one DIMM per channel, while the 6 TB maximum assumes two.

#amd epyc#threadripper#ai server#pcie lanes#gpu server

We design custom AI servers

We configure each server for the models you will run, and our build partner assembles and tests it. 3-year warranty, or 5 years at extra cost.

Get in touch
Call Us