A100 vs H100

A100 vs H100: who should pick which.

Have an A100 or an H100, or need one? Tell us which. We verify the hardware, match a Have with a Need, and handle the deal. If you're selling, you pick which offer to take.

NVIDIA

A100

Architecture
Ampere
GPU memory
40 GB HBM2 or 80 GB HBM2e
Form factors
SXM4, PCIe

NVIDIA

H100

Architecture
Hopper
GPU memory
80 GB (SXM, PCIe), 94 GB (NVL)
Form factors
SXM, PCIe, NVL

A100

Pick the A100 for Ampere, 40 GB or 80 GB, and a lighter rack. SXM is 400 W, and a custom thermal 80 GB SKU can go to 500 W.

NVIDIA A100

H100

Pick the H100 for Hopper and more memory bandwidth. It comes as 80 GB SXM, 80 GB PCIe, or 94 GB NVL, with SXM configurable up to 700 W.

NVIDIA H100

Where each generation sits

  1. Ampere

    A100

  2. Hopper

    H100

  3. Blackwell

  4. Blackwell Ultra

Side by side

GPU memory

A10080 GB

40 GB HBM2 or 80 GB HBM2e

H10094 GB

80 GB (SXM, PCIe), 94 GB (NVL)

Memory bandwidth

A1002,039 GB/s

2,039 GB/s (80 GB SXM), 1,935 GB/s (80 GB PCIe), 1,555 GB/s (40 GB)

H1003.9 TB/s

3.35 TB/s (SXM), 2 TB/s (PCIe 80 GB), 3.9 TB/s (NVL)

Per-GPU power

A100500 W

250 W (40 GB PCIe), 300 W (80 GB PCIe), 400 W SXM, up to 500 W on a custom thermal 80 GB SKU

H100700 W

Up to 700 W (SXM), 350 W (PCIe 80 GB), 350–400 W (NVL)

Architecture

A100. Ampere

H100. Hopper

Form factor

A100. SXM4, PCIe

H100. SXM, PCIe, NVL

NVLink

A100. 600 GB/s per SXM GPU

H100. 900 GB/s per SXM GPU; 600 GB/s on NVL

Common systems

A100. HGX A100 (4-, 8-, or 16-GPU), DGX A100, PCIe servers

H100. HGX H100 (4-GPU or 8-GPU), DGX H100

Published system power

A100. Dense 4U–8U class air-cooled nodes are common, and lighter than an H100 node

H100. Official DGX H100 is 8U at 10.2 kW. Partner HGX nodes vary.

Where each one fits

The A100 is the previous-generation NVIDIA data-center GPU. People still use it for training, fine-tuning, and inference when they want Ampere software and a known platform. It moves as 4-GPU and 8-GPU HGX servers, DGX A100 systems, and PCIe cards.

The H100 is Hopper, the volume workhorse for training and inference. SXM is the GPU in HGX H100 and DGX H100 8-GPU servers. The 80 GB PCIe card fits standard servers. H100 NVL is the 94 GB card, often used in bridged pairs.

Choose on memory size, form factor, and the rack you can power. An 8-GPU A100 node is still dense, and it is lighter than an H100, H200, or B200/B300 node. An 8-GPU H100 SXM server is a high-density rack load.

What to check on a used unit

On a used A100, separate 40 GB from 80 GB, and SXM from PCIe, before anything else. They are not interchangeable. A100s have been in production longer than Hopper, so hours, utilization, and test records matter. Note air or liquid cooling, whether you have a complete server or loose GPUs, and missing NVLink bridges, mezzanine NICs, or PSUs.

On a used H100, name the version first: SXM, PCIe, or NVL. An SXM module only fits an HGX or DGX-class board. Note hours, air or liquid cooling, complete server versus loose GPUs, test records, and missing NVLink bridges, optics, PSUs, or rails.

On every Replenysh deal, both sides pass KYC, every buyer and destination is screened against export-control and sanctions lists, and we confirm serials and condition. The same NDA and deal terms apply, with a named Replenysh person. Payment, shipping, data wiping, and colocation are arranged based on what each deal needs. How a used GPU deal works.

Questions people ask

Who should pick the A100, and who should pick the H100?

Pick the A100 for Ampere, 40 GB or 80 GB, and a lighter rack. Pick the H100 for Hopper, more memory bandwidth, and 80 GB SXM, 80 GB PCIe, or 94 GB NVL. Tell us which you have or need.

How do A100 and H100 memory and power compare?

The A100 is 40 GB HBM2 or 80 GB HBM2e. The H100 is 80 GB on SXM and PCIe, or 94 GB on NVL, with higher memory bandwidth (3.35 TB/s on SXM). A100 SXM is 400 W, up to 500 W on a custom thermal 80 GB SKU. H100 SXM is configurable up to 700 W. Official DGX H100 systems are 8U at 10.2 kW.

Are these complete servers or loose GPUs?

Either. An A100 can be an 8-GPU server, a 4-GPU server, loose GPUs, or a rack. An H100 can be a complete server, loose modules or cards, or a full rack. We collect which one it is before we match it.

What should I check on a used A100 or H100?

Memory size and form factor first, then hours, cooling, whether the unit is complete, test records, and missing bridges, optics, or power supplies. We confirm serials and condition before we bring the hardware to buyers.

Do you publish a price for the A100 or the H100?

No. We don't publish prices. Tell us what you have or what you need, and vetted buyers make offers. If you're selling, you pick which offer to take.

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.

A100 vs H100: who should pick which

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