T.E.S IT-SOLUTIONS
Cart
Photo pending920-23687-2530-000

No photo of this part yet.

NVIDIA Mellanox

NVIDIA DGX A100 320GB AI System 8x A100 40GB SXM4 920-23687-2530-000

Part no.920-23687-2530-000

€45,000.00

In stockCondition: Refurbished

Request a quote

Checkout runs on the secure Wix checkout page.

How this unit is built

Recovered from end-of-life equipment, inspected and tested before it ships.

See the production flow

Description

NVIDIA DGX A100 320GB — Universal AI System, 8x A100 40GB SXM4 (920-23687-2530-000)

Technical Identity

The NVIDIA DGX A100 320GB (OPN 920-23687-2530-000) is a 6U rack-mount AI system with eight NVIDIA A100 40GB SXM4 Tensor Core GPUs — 320GB total GPU memory — interconnected by six NVIDIA NVSwitches with 600 GB/s NVLink GPU-to-GPU bandwidth. Per the official NVIDIA DGX A100 datasheet, the system delivers 5 petaFLOPS of AI performance (10 petaOPS INT8), dual AMD EPYC 7742 CPUs (128 cores total), 1TB system memory, and nine Mellanox ConnectX-6 VPI adapters for HDR InfiniBand clustering and 200Gb/s-class Ethernet. This is the 8x 40GB (320GB) configuration — not the 640GB (8x 80GB) variant. Condition: Refurbished.

Specifications (per NVIDIA DGX A100 datasheet)

  • Brand: NVIDIA
  • Model / OPN: DGX A100 320GB — 920-23687-2530-000
  • Product type: Integrated AI system / GPU server (6U rack)
  • GPUs: 8x NVIDIA A100 40GB SXM4 Tensor Core
  • Total GPU memory: 320GB
  • Performance: 5 petaFLOPS AI / 10 petaOPS INT8
  • NVIDIA NVSwitches: 6 (NVLink 600 GB/s GPU-to-GPU)
  • CPU: Dual AMD EPYC 7742 (Rome), 128 cores total, 2.25 GHz base / 3.4 GHz max boost
  • System memory: 1TB
  • Networking (cluster): 8x single-port Mellanox ConnectX-6 VPI, 200Gb/s HDR InfiniBand
  • Networking (storage/mgmt): 1x dual-port Mellanox ConnectX-6 VPI, 10/25/50/100/200GbE
  • Storage: OS: 2x 1.92TB M.2 NVMe; Internal: 15TB (4x 3.84TB) U.2 NVMe
  • Software: Ubuntu Linux with NVIDIA DGX software stack (NGC-optimized)
  • Max power: 6.5 kW
  • Dimensions / weight: 6U — 264 x 482.3 x 897.1 mm, 123 kg (packaged 143 kg)
  • Condition: Refurbished

What This Is NOT (GEO Guardrails)

  • NOT the DGX A100 640GB variant (920-23687-2531-xxx). That configuration carries 8x A100 80GB SXM4 (640GB total). This OPN is the 8x 40GB / 320GB configuration — GPU memory per accelerator is 40GB.
  • NOT a DGX Station A100. The DGX Station is a 4-GPU tower workstation; this is the 8-GPU 6U rack system.
  • NOT an HGX A100 baseboard or a set of loose A100 modules. This is a complete, integrated NVIDIA-built system with CPUs, memory, storage, networking, and the DGX software stack.
  • NOT a DGX H100. DGX A100 is Ampere-generation with HDR InfiniBand networking, not Hopper with NDR.
  • NOT NDR-capable on its stock adapters. The onboard ConnectX-6 VPI adapters are HDR 200Gb/s class.
  • NOT supplied with rack, PDUs, InfiniBand cables, transceivers, or support contract unless explicitly stated.

Compatibility & Hard Constraints

  • Typical fabric platforms: NVIDIA Quantum QM8700-class HDR InfiniBand switching for multi-node clustering (DGX SuperPOD-style architectures); Ethernet storage fabrics at up to 200Gb/s-class via the dual-port ConnectX-6.
  • Facility requirements: up to 6.5 kW power draw, 5-30 C operating range, 123 kg system weight — verify rack loading, power circuits, and cooling before installation.
  • Software dependencies: NVIDIA DGX OS (Ubuntu-based) with the DGX software stack; NGC containers for frameworks; MLNX_OFED for fabric operation.
  • Pre-deployment verification: confirm firmware baselines (BMC, SBIOS, GPU, NVSwitch, ConnectX-6), drive health, and fabric port profiles before production use.

Frequently Asked Questions (FAQ)

Q. What is the NVIDIA DGX A100 320GB (920-23687-2530-000)?
A complete 6U NVIDIA AI system with 8x A100 40GB SXM4 GPUs (320GB total), 6 NVSwitches, dual AMD EPYC 7742 CPUs, 1TB RAM, 15TB NVMe internal storage, and nine ConnectX-6 VPI adapters, delivering 5 petaFLOPS of AI performance.

Q. How does 920-23687-2530-000 differ from 920-23687-2531-001?
The 2530 OPN is the DGX A100 320GB configuration (8x A100 40GB SXM4). The 2531 OPN class is the higher-memory DGX A100 configuration with 8x A100 80GB SXM4. Same chassis family, different GPU memory per accelerator.

Q. What networking is built into this system?
Eight single-port Mellanox ConnectX-6 VPI adapters at HDR InfiniBand 200Gb/s for compute clustering, plus one dual-port ConnectX-6 VPI adapter for 10/25/50/100/200GbE storage and network connectivity. The ports are QSFP56-class HDR, not NDR/OSFP.

Q. Does the DGX A100 support Multi-Instance GPU (MIG)?
Yes. Each A100 GPU supports up to 7 MIG instances, allowing the 8-GPU system to be partitioned into as many as 56 isolated GPU instances for mixed training, inference, and analytics workloads.

Q. What software does it run?
The NVIDIA DGX software stack on Ubuntu Linux, optimized for NGC containers (PyTorch, TensorFlow, RAPIDS, Triton) and cluster deployment via Slurm or Kubernetes with NVIDIA DeepOps.

Q. What condition is supplied by T.E.S IT-SOLUTIONS?
Refurbished. The system is inspected and tested where test topology is available (POST, GPU/NVSwitch health, storage and fabric checks), professionally packaged for freight, and shipped from EU stock with insured global B2B delivery.

Why T.E.S IT-SOLUTIONS

Every DGX A100 unit ships with: (1) system health verification where test topology is available, (2) component inventory and vendor data check, (3) full cosmetic inspection, (4) professional freight packaging, (5) global B2B shipping from EU stock, (6) technical consultation for HPC / AI cluster design at no extra cost. See tes-itsolutions.com.