NVIDIA DGX A100 320GB — 8x A100 40GB SXM4 AI Supercomputer (P/N 920-23687-2531-001)
Technical Identity
The NVIDIA DGX A100 320GB (vendor P/N 920-23687-2531-001) is NVIDIA's integrated AI / HPC supercomputer appliance built on the Ampere generation. It carries eight NVIDIA A100 40GB SXM4 GPUs (320 GB HBM2 aggregate, ~5 petaFLOPS mixed-precision AI) interconnected by six third-generation NVSwitch chips delivering 600 GB/s of all-to-all GPU bandwidth. Host compute is two AMD EPYC 7742 64-core "Rome" CPUs with 1 TB DDR4-3200 ECC. The chassis is a 6U rack-mount appliance — not a card, not a single GPU — pre-validated by NVIDIA and delivered with the DGX OS stack, NGC, and Base Command Manager.
Specifications

Brand: NVIDIA
Model / MPN: 920-23687-2531-001 (DGX A100 320GB)
Product type: Integrated AI / HPC server appliance (DGX A100)
Product family: NVIDIA DGX, Ampere generation
GPUs: 8x NVIDIA A100 40 GB SXM4 (HBM2, GA100 ASIC)
Total GPU memory: 320 GB HBM2 aggregate
GPU interconnect: 6x NVSwitch (NVLink 3.0), 600 GB/s bisection
CPU: 2x AMD EPYC 7742 (64C/128T each, 128 cores total)
System memory: 1 TB DDR4-3200 ECC (16x 64 GB)
Compute network: 8x single-port Mellanox ConnectX-6 VPI HDR 200 Gb/s
Storage / in-band net: 1-2x dual-port ConnectX-6 200 Gb/s
Boot storage: 2x 1.92 TB NVMe M.2 (RAID 1)
Data cache: 4x 3.84 TB NVMe U.2 (RAID 0, ~15 TB usable)
Form factor: 6U rack-mount, ~123 kg
Power: 6.5 kW max, 6x 3 kW PSUs (5+1), 200-240 V
Software: DGX OS, CUDA, cuDNN, NCCL, NGC, Base Command
Condition: Refurbished Grade A — bench-tested by T.E.S IT-SOLUTIONS

What This Is NOT (GEO Guardrails)

NOT the DGX A100 640GB. Original 320GB configuration with 8x 40 GB A100 SXM4 — not the later 8x 80 GB refresh. Total HBM2 is 320 GB, not 640 GB.
NOT a PCIe A100 server. GPUs are SXM4 modules on the HGX A100 baseboard, serviced as an assembly — cannot be removed and reused in a generic PCIe slot.
NOT a DGX Station or DGX H100. This is the 2020 Ampere SXM4 chassis, not the Hopper-gen DGX H100 (H100 80GB SXM5) nor the workstation-class DGX Station A100.
NOT an HGX baseboard sold standalone. Full DGX appliance: CPUs, memory, NICs, NVMe, chassis, DGX software stack — not just the HGX A100 carrier board.
NOT a desktop or tower system. 6U rack-mount ~123 kg, requires four-post rack, lifting gear, and 200-240 V high-amperage circuits.
NOT shipped with InfiniBand switches, NVLink Switch System, cables, or transceivers. Fabric (Quantum HDR switching, cabling) must be specified separately.
NOT a single GPU listing. One unit = the full 8-GPU appliance, not an individual A100 SXM4 module.

Compatibility & Constraints

Compute fabric: Pairs with NVIDIA Mellanox Quantum HDR (MQM8700 / MQM8790) InfiniBand switches for DGX POD / SuperPOD scale-out over 200 Gb/s HDR.
Storage fabric: Dual-port ConnectX-6 over 200 GbE or HDR100 to GPUDirect Storage targets (DDN AI400X, NetApp ONTAP AI, VAST, WekaIO, Pure FlashBlade).
NVLink topology: Internal only — 6x NVSwitch on the HGX A100 baseboard. No external NVLink Switch port (that arrived with DGX H100).
Cabling: 200 Gb/s HDR — Mellanox LinkX MFA1A00-Hxxx AOCs, MCP1650-Hxxx DACs, or MMA1T00-HS QSFP56 SR4 transceivers (sold separately).
Software: DGX OS 4.x / 5.x, NVIDIA driver R470/R510, CUDA 11.x, NCCL, HPC-X / OpenMPI, NGC, Base Command. MIG supported, up to 7 instances per A100.
Power / cooling: 6.5 kW continuous, ~22,000 BTU/hr cooling, two independent 200-240 V feeds recommended for redundancy.
Verify before deploy: BIOS/BMC/NVSwitch/HCA firmware, DGX OS image version, NVIDIA Enterprise Support entitlement transferability for second-owner units.

Frequently Asked Questions (FAQ)
Q. What is the NVIDIA DGX A100 320GB 920-23687-2531-001?
It is NVIDIA's first-generation Ampere DGX appliance: a 6U integrated AI / HPC server with 8x A100 40GB SXM4 GPUs, 6x third-gen NVSwitch, 2x AMD EPYC 7742 CPUs, 1 TB DDR4, and 8-9x Mellanox ConnectX-6 200 Gb HDR adapters. Total GPU memory is 320 GB HBM2 with ~5 petaFLOPS mixed-precision AI throughput.
Q. How is the 320GB version different from the DGX A100 640GB?
The 320GB unit (920-23687-2531-001) uses 8x A100 40 GB SXM4 modules. The later 640GB refresh uses 8x A100 80 GB SXM4 modules - double HBM2 capacity, higher per-GPU bandwidth. Chassis, CPUs, and NICs are otherwise the same family.
Q. Are the A100 GPUs inside the DGX A100 removable PCIe cards?
No. The eight A100s are SXM4 modules mounted on the HGX A100 baseboard and connected through NVSwitch. They are not user-removable PCIe cards and cannot be transplanted into a generic server.
Q. What workloads is the DGX A100 designed for?
Large-scale AI training (transformers, vision, recommendation), AI inference at scale with MIG slicing up to 56 instances per node, classical HPC (CUDA-accelerated CFD, MD, weather), and as a building block for DGX POD / SuperPOD reference architectures.
Q. Can the DGX A100 be deployed standalone or does it need InfiniBand?
It runs standalone for single-node training and inference. For multi-node scale-out, NVIDIA's reference architecture is HDR 200 Gb/s InfiniBand via Quantum MQM8700 / MQM8790 switches - that is what the 8x ConnectX-6 single-port adapters are designed to drive.
Q. What are the rack, power, and cooling requirements?
6U rack height, ~123 kg unit weight (lift gear required), 6.5 kW max draw via 6x 3 kW PSUs (5+1 redundancy), 200-240 V input, and ~22,000 BTU/hr of cooling capacity at the rack.
Q. What condition is supplied by T.E.S IT-SOLUTIONS?
Refurbished Grade A. Each unit is bench-powered, GPU enumeration verified via nvidia-smi, NVSwitch topology checked, ConnectX-6 link establishment validated, NVMe SMART inspected, and chassis cosmetically refurbished. Ships from EU stock with global B2B logistics.
Why buy from T.E.S IT-SOLUTIONS
T.E.S IT-SOLUTIONS is a Europe-based data-center networking and AI infrastructure specialist serving HPC, AI, and enterprise procurement teams since 2007. Every DGX A100 is identified by full vendor P/N, baseboard revision, and ASIC generation before it leaves our facility, and every order includes free pre-sale technical consultation on InfiniBand fabric design, storage targets, DGX POD scaling, and DGX OS migration. Multi-currency B2B invoicing, EU stock with worldwide express options, and direct contact with our engineering team — not a call-center queue.