Home › Servers › NVIDIA

NVIDIA DGX A100

End of saleServers6U rackmount
NVIDIA DGX A1006U rackmount2 socketsmax 2.05 TB

The NVIDIA DGX A100 is a 6U eight-GPU training and inference server built around eight A100 SXM GPUs linked by six NVSwitches. It uses two 64-core AMD EPYC 7742 processors, 1 TB or 2 TB of memory and NVMe storage, and shipped in 320 GB and 640 GB GPU-memory versions. It is the predecessor of the DGX H100 and is traded second-hand as clusters retire. Buyers should confirm which GPU memory version they are getting, because the 40 GB and 80 GB A100 variants are priced differently.

Also written as: NVIDIA DGX A100 320GB, NVIDIA DGX A100 640GB

Buy NVIDIA DGX A100Sell yours

Specifications

BrandNVIDIA
FamilyDGX
Form factor6U rackmount
Released2020
Cpu sockets2
Cpu platform2x AMD EPYC 7742 (Rome), 64 cores each, 128 cores total
Memory max (GB)2,048
Memory typeDDR4, 1 TB per 16 DIMMs
Gpu support8x NVIDIA A100 with third-generation NVLink and six NVSwitches, in 320 GB or 640 GB GPU memory models
Gpu memory (GB)640
Drive bays2x 1.92 TB NVMe M.2 (OS, RAID 1) plus NVMe U.2 SED data cache, 15 TB or 30 TB as shipped, upgradeable to 60 TB
Network onboardConnectX-6 or ConnectX-7 single-port InfiniBand cluster cards up to 200 Gb/s and dual-port 200GbE storage cards; up to 9 or 10 ports as shipped
Power supplySix power supplies in 3+3 redundancy, 3000 W each at 200-240 V
Weight kg123.16

Generations

Works together

How the NVIDIA DGX A100 connects to servers, storage and the network: 4 ways, with the parts and products on each path.

Fibre Channel SANiSCSI and NVMe/TCP over EthernetDirect-attached SASFile storage (NFS and SMB)

Fibre Channel SAN

Match the speed: 16Gb and 32Gb FC HBAs, switches and array ports work together, and each generation also runs one or two speeds below its own (a 32Gb port runs at 16Gb and 8Gb).

Switches

Storage

Optics

Fibre Channel SAN: how it fits together

iSCSI and NVMe/TCP over Ethernet

Keep the speed the same at all three points: a 25GbE server, a 25GbE switch port and a 25GbE array port. Mixed speeds work but the slowest link sets the pace.

Adapters for the server (check the fit list)

Switches

Storage

Optics

iSCSI and NVMe/TCP over Ethernet: how it fits together

Direct-attached SAS

12Gb SAS uses Mini-SAS HD (SFF-8644) connectors; older 6Gb gear uses Mini-SAS (SFF-8088). Adapter cables exist, but the link runs at the slower speed.

Storage

Direct-attached SAS: how it fits together

File storage (NFS and SMB)

Front-end ports face the clients; back-end ports (on scale-out systems) connect the nodes to each other and need their own switch pair.

Switches

Storage

Optics

File storage (NFS and SMB): how it fits together

Plan and size it

Check compatibility with NVIDIA

Before you buy, confirm the exact configuration against the vendor's own lists: firmware levels, operating systems, adapters and optics that NVIDIA has tested with the NVIDIA DGX A100.

Equivalent models from other brands

Buying a used DGX A100: what to check

Check the GPU memory version (320 GB or 640 GB) in nvidia-smi, confirm all six power supplies and both ConnectX card sets are present, and ask whether the unit carries an NVIDIA DGX support entitlement, which usually does not transfer. Test all eight GPUs and the NVSwitch fabric with the DGX diagnostics before paying.

FAQ

What is the difference between the DGX A100 320GB and 640GB?

The 320GB model uses eight 40 GB A100 GPUs and ships with 1 TB of system memory and 15 TB of NVMe data cache. The 640GB model uses 80 GB GPUs and ships with 2 TB of memory and 30 TB of cache.

What CPUs does the DGX A100 use?

Two AMD EPYC 7742 (Rome) processors with 64 cores each, 128 cores in total.

How much power does the DGX A100 need?

It has six 3000 W power supplies in 3+3 redundancy, each rated at 200 to 240 V, and NVIDIA lists 6.5 kW as the maximum for the system.

Buying or selling DGX A100?

New, refurbished or used. Tell us the quantity and condition and we connect you with a vetted reseller in your region, usually within one working day.

Sources