GPUsIntel · OAM, Gaudi 2 or Gaudi 3

Gaudi 2 and Gaudi 3Have or need Intel Gaudi accelerators?

Intel Gaudi 2 and Gaudi 3 are purpose-built AI accelerators, used in OAM or mezzanine platforms rather than as casual PCIe training cards. Tell us the exact generation you have, or the exact generation you need. We verify the hardware, match both sides, and run the deal through payment and delivery. If you're selling, you pick which offer to take.

Intel Gaudi server node, a chassis holding accelerator modules marked Intel Gaudi
Image: Intel

What the Gaudi 2 / 3 is

Gaudi is Intel's (formerly Habana) training and inference accelerator family, with on-board Ethernet-style scale-out rather than NVIDIA NVLink. The software stack is Intel's Gaudi software, not CUDA.

Published figures:

  • Gaudi 2 — 96 GB HBM2e, about 2.45 TB/s, 600 W card-level TDP (HL-225 class).
  • Gaudi 3 — 128 GB HBM2e, 3.7 TB/s. OAM TDP is published at 900 W air-cooled or 1,200 W liquid-cooled.

These usually move as complete 8-accelerator nodes, not as loose experiment cards. If you have NVIDIA or AMD systems, use those family pages.

Typical configs

Gaudi 2

Memory
96 GB HBM2e
Power
600 W class
Platform
Mezzanine or OAM servers, often eight accelerators per node, with the Gaudi networking built in

Gaudi 3

Memory
128 GB HBM2e, 3.7 TB/s
Power
900 W air or 1,200 W liquid, per OAM
Platform
Same idea, higher power, often liquid

What affects condition and lead time

  • Gaudi 2 versus Gaudi 3, and air versus liquid on Gaudi 3.
  • Complete node versus modules, including the scale-out cables and switches that belong with the platform.
  • Software stack revision.
  • Test and wipe records.
  • Location. Every buyer and destination is screened against export-control and sanctions lists before the hardware moves.

Colo and power

  • Gaudi 2 at 600 W is already a dense node at eight accelerators.
  • Gaudi 3 at 900 W air or 1,200 W liquid per OAM needs a site that was designed for it.
  • The hardware does not include a colo site. If you need a place to land it, tell us the power, the cooling, and the site you have in mind.

The deal

How a deal works

How a used GPU deal works
  1. 01

    Verify

    Agreements first. We confirm the serials and condition, and qualify both sides.

  2. 02

    Match

    Vetted buyers make offers. The seller chooses which offer to take.

  3. 03

    Pay and deliver

    Every destination is screened. Payment happens with delivery.

Questions people ask

What is Intel Gaudi?

A family of Intel AI accelerators, with Gaudi 2 at 96 GB and Gaudi 3 at 128 GB, used in dedicated OAM or mezzanine servers and Intel's software stack.

How is Gaudi 3 different from Gaudi 2?

Gaudi 3 has 128 GB of HBM2e at 3.7 TB/s, and a higher power envelope (900 W air or 1,200 W liquid per OAM). Gaudi 2 has 96 GB and a 600 W-class card TDP.

Does Gaudi use CUDA?

No. It runs on Intel's Gaudi software. Tell us that on both sides of a match so the stack is not a surprise.

How much power does Gaudi 3 use?

Official material puts OAM TDP at 900 W with air cooling or 1,200 W with liquid cooling, per accelerator.

How do I sell Gaudi accelerators I have?

Tell us Gaudi 2 or Gaudi 3, air or liquid, how many, the condition, and where they sit. We verify the hardware, match your Have with a Need, and handle the deal. You pick which offer to take.

How do I get Gaudi accelerators I need?

Tell us the generation, how many, and what power and cooling you have ready. We match your Need with a Have and run the deal through payment and delivery. We don't publish prices.

How do you check buyers and export rules?

Every buyer and every seller passes a KYC identity and company check, and a Replenysh person reviews every buyer before they see a deal. Every buyer and destination is screened against export-control and sanctions lists before the hardware moves.

Have or need Intel Gaudi accelerators?

Tell us what you have, or what you need. We verify, match, and handle the deal.