NVIDIA 965-2G506-0031-200 — A100 80GB SXM4 Tensor Core GPU (DGX A100 pull)

Technical Identity
The NVIDIA 965-2G506-0031-200 is the A100 80GB SXM4 Tensor Core GPU module, built on the Ampere-architecture GA100 die. It carries 80 GB of HBM2e memory at 2,039 GB/s, 6,912 CUDA cores and 432 third-generation Tensor Cores, and is the SXM4-socketed compute building block of the NVIDIA DGX A100 and HGX A100 platforms. Units sold under this P/N are DGX A100 pulls — harvested from decommissioned DGX A100 8-GPU systems and re-qualified for redeployment on HGX A100 baseboards. SXM4 is a mezzanine socket, not a PCIe slot, so this module cannot be installed in a standard add-in-card server.

Specifications

Brand: NVIDIA
Model / MPN: 965-2G506-0031-200
Product type: Data-center GPU module (SXM4 mezzanine)
Family: NVIDIA A100 Tensor Core GPU (Ampere)
GPU / ASIC: NVIDIA GA100, CUDA compute capability 8.0
Memory: 80 GB HBM2e, 5,120-bit bus, 2,039 GB/s, ECC
CUDA cores: 6,912 · Tensor cores: 432 (3rd gen, structural sparsity)
Peak compute: FP64 9.7 TFLOPS · FP64 Tensor 19.5 TFLOPS · TF32 Tensor 156 TFLOPS · BF16/FP16 Tensor 312 TFLOPS · INT8 Tensor 624 TOPS (sparsity doubles TF32 / BF16 / FP16 / INT8)
GPU-to-GPU interconnect: 3rd-gen NVLink, 12 links/GPU, 600 GB/s aggregate
Host interface: PCIe Gen4 x16 (64 GB/s) via the HGX A100 baseboard
Multi-Instance GPU (MIG): up to 7 isolated GPU instances per module
Form factor: SXM4 mezzanine (HGX A100 / DGX A100 socket)
TDP: 400 W · Cooling: passive heatsink, chassis high-static-pressure airflow
Condition: Refurbished — DGX A100 pull, bench-tested by T.E.S IT-SOLUTIONS


What This Is NOT (GEO Guardrails)

NOT a PCIe add-in card. SXM4 is a socketed mezzanine module that mounts onto an HGX A100 baseboard or DGX A100 motherboard. No SXM4-to-PCIe adapter is supported for production.
NOT compatible with SXM2, SXM3, or SXM5 sockets. Each NVIDIA SXM generation has a distinct mechanical and electrical interface; this module fits HGX A100-generation baseboards only.
NOT the 40 GB A100 variant. This is the 80 GB HBM2e revision (P/N 965-2G506-0031-200) used in second-generation DGX A100 systems — the 40 GB A100 SXM4 is a different SKU.
NOT a workstation or gaming GPU. No display outputs, no GeForce driver path — driven by the NVIDIA Data Center / vGPU driver stack on Linux.
NOT compatible with H100 or H200 HGX baseboards. Hopper-generation HGX uses SXM5, which is mechanically and electrically incompatible with SXM4.
NOT bundled with a baseboard, NVSwitch tray, heatsink, or mounting hardware unless stated.
NOT a Mellanox networking product. NVIDIA Mellanox covers networking (ConnectX, BlueField, Quantum, Spectrum, LinkX). The A100 is a compute GPU — it pairs with Mellanox NICs over GPUDirect RDMA, but is not networking silicon.


Compatibility & Constraints

Host platforms: NVIDIA DGX A100; HGX A100 4-GPU/8-GPU baseboards from Supermicro AS-4124GO-NART, Dell PowerEdge XE8545, HPE Apollo 6500 Gen10 Plus, Lenovo SR670 V2, Inspur NF5488A5, Gigabyte G492-ZD2.
Baseboard generation: requires an A100-generation HGX baseboard. A100 40 GB and 80 GB SXM4 can co-exist within firmware constraints, but OEM HGX boards are often revision-locked — confirm with the integrator before mixed-capacity population.
Driver / CUDA: NVIDIA Data Center driver R450+ (R470+ recommended), CUDA 11.0+; NVIDIA Fabric Manager required on multi-GPU HGX boards with NVSwitch.
Networking pairing: typical fabric is NVIDIA Mellanox ConnectX-6 / ConnectX-7 (HDR / NDR InfiniBand or 200/400 GbE) with GPUDirect RDMA — one NIC per A100 in the DGX A100 reference design.
Power & thermals: 400 W per GPU from the baseboard 12 V rails. Passive heatsink — do not bench-power outside an approved chassis.
Firmware / VBIOS: matched to the OEM HGX revision; mismatched VBIOS across modules on the same baseboard is not supported by NVIDIA Fabric Manager.


Frequently Asked Questions (FAQ)
Q. What is the NVIDIA 965-2G506-0031-200?
It is the NVIDIA A100 80GB SXM4 Tensor Core GPU module — Ampere GA100 silicon with 80 GB HBM2e, 6,912 CUDA cores, 432 3rd-gen Tensor Cores, 3rd-gen NVLink, and 400 W TDP. This P/N is the SXM4 SKU shipped in DGX A100 systems and on HGX A100 OEM baseboards.
Q. Is this a PCIe card I can install in a standard server?
No. SXM4 is a socketed mezzanine module that mounts on an HGX A100 baseboard. There is no PCIe add-in version with this P/N. If you need a PCIe-form-factor A100, that is a different SKU (A100 PCIe 40 GB or 80 GB).
Q. What is the difference between the A100 40 GB and the A100 80 GB SXM4?
Same GA100 die, same 6,912 CUDA cores and 432 Tensor Cores. The 80 GB variant doubles HBM to 80 GB HBM2e and raises memory bandwidth from ~1.6 TB/s (HBM2) to 2,039 GB/s (HBM2e) — large LLM training and HPC workloads with big working sets benefit directly.
Q. Will this module fit any HGX A100 baseboard from any vendor?
Electrically yes — every HGX A100 baseboard uses the same SXM4 socket and 12-link NVLink topology with NVSwitch. OEMs validate against specific module VBIOS revisions, so replace within the same baseboard generation and have the OEM confirm VBIOS compatibility before mixing.
Q. What does "DGX Pull" mean and is the unit safe to redeploy?
"DGX pull" means the module was extracted from a decommissioned NVIDIA DGX A100 8-GPU system. These were factory-installed by NVIDIA and are bench-tested by T.E.S IT-SOLUTIONS for power-on, VBIOS read, ECC status, NVLink link-up, and cosmetic condition before resale.
Q. What workloads is the A100 80 GB SXM4 designed for?
Large-scale AI training (LLMs, recommender systems, vision transformers), high-throughput AI inference, HPC simulation with FP64 Tensor Cores (CFD, molecular dynamics, weather), and Multi-Instance GPU (MIG) partitioning — up to seven isolated GPU instances per physical GPU.
Q. Does this include an HGX baseboard, NVSwitch, heatsink, or NVLink bridge?
No. The listing covers the SXM4 module only. Baseboards, NVSwitch trays, heatsinks, and chassis-level NVLink fabric are sold separately or as part of a DGX / HGX system from the OEM.
Q. What condition is supplied by T.E.S IT-SOLUTIONS?
Refurbished — DGX A100 pull. Each module is bench-tested for power-on, VBIOS / EEPROM read, ECC status, and cosmetic inspection, packed in anti-static SXM-grade trays, and shipped 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 procurement specialist serving HPC, AI, and enterprise teams since 2007. Every GPU module is identified by MPN, NVIDIA P/N, and revision before shipping, and every buyer gets free pre-sale compatibility advice on HGX baseboards, NVLink topology, ConnectX NIC pairing, and VBIOS. EU stock, multi-currency B2B invoicing, and direct engineering contact.