NVIDIA A100 Specs & Datasheet (80GB SXM · HGX & DGX A100)

The NVIDIA A100 is a data-center GPU built on NVIDIA’s Ampere architecture. The mainstream data-center variant is the A100 SXM 80GB — an SXM4-form GPU with 80 GB of on-package HBM2e memory, used for AI training, inference, and HPC.

The A100 is almost always deployed as an 8-GPU building block, and it appears in two forms:

• HGX A100 — the 8-GPU NVLink baseboard that integrators (e.g. Supermicro) drop into their own server chassis. You buy it as part of an OEM system.
• NVIDIA DGX A100 — NVIDIA’s own turnkey appliance. The DGX A100 640GB is exactly eight A100 SXM 80GB GPUs (640 GB HBM2e per system) on the same 8-GPU NVLink domain — the DGX is the integrated system, the A100 SXM is the chip inside it.

A100 is a prior-generation part: it predates the Hopper H100/H200 and the current Blackwell generation. It remains widely deployed and available on the secondary market as fleets rotate to newer silicon — a proven, brand-recognized platform rather than the latest one.

The NVIDIA A100 SXM 80GB (Ampere) specification. The per-GPU figures below are fixed characteristics of the chip; the 8-GPU column is the same GPU on an 8-GPU NVLink domain — the form the HGX A100 node and the DGX A100 640GB system both take. Chassis, host CPU, cooling, site power, and final configuration vary by integrator and are confirmed at quote.

Each A100 SXM carries 80 GB of HBM2e — the on-package GPU memory (VRAM) available to the model — at 2.04 TB/s per GPU of memory bandwidth.

Across the 8-GPU NVLink domain that both the HGX A100 node and the DGX A100 640GB system use, that is 640 GB of HBM2e per node (8 × 80 GB), pooled and addressable over NVLink so a large model and its KV cache can span all eight GPUs. This is why the DGX A100 640GB carries that number in its name: it is eight A100 SXM 80GB GPUs, nothing more exotic.

Inside an 8-GPU node the A100s are wired over third-generation NVLink at 600 GB/s per GPU, giving all-to-all connectivity across the 8-GPU NVLink domain so the node behaves as one tightly-coupled accelerator rather than eight discrete cards.

That 8-GPU NVLink domain is the fixed characteristic shared by the HGX A100 baseboard and the DGX A100 system; scale-out networking (InfiniBand adapters, GPU:NIC ratio) is where the integrator builds differ, and is confirmed per build at quote.

The A100 SXM 80GB ships in two 8-GPU forms below — the Supermicro HGX A100 8-GPU node and the NVIDIA DGX A100 640GB turnkey appliance. Both carry the same A100 SXM 80GB GPU, the same 640 GB of HBM2e per node, and the same 8-GPU third-generation NVLink domain; chassis, host CPU, networking, warranty, and lead time differ by build. Each links to its full specification.

The A100 is Ampere — a prior-generation part. If you want the current silicon, the honest pointer is up a generation:

• Hopper H100 / H200 — the A100’s direct successors. The H200 in particular lifts per-GPU memory to 141 GB of HBM3e at far higher bandwidth, a large gain for memory-bound inference over the A100’s 80 GB HBM2e.
• Blackwell (B200 and up) — the current top of the stack; see Blackwell vs Hopper for how the generations line up.

What keeps the A100 in service is not raw spec leadership but availability and value: it is a proven, brand-recognized 8-GPU building block that is widely deployed and trades actively on the secondary market as fleets rotate to Hopper and Blackwell. For workloads that fit 80 GB per GPU (or 640 GB across a node) it remains a capable training and inference platform — just not the newest one.

Both A100 forms — the HGX A100 8-GPU node and the DGX A100 640GB system — are quoted per configuration on request; the figure depends on the build, host CPU, networking, unit grade, and quantity, all confirmed at quote. Tell us the build you are standing up and we will return availability and a facility-fit review. Browse the GPU catalog to compare integrations, or step up a generation to the HGX H200 if you want current Hopper silicon.