NVIDIA A100 40GB SXM4 HGX Tensor Core GPU — 900-2G509-A500-000

Technical Identity
The NVIDIA A100 40GB SXM4 Tensor Core GPU (P/N 900-2G509-A500-000) is NVIDIA's flagship Ampere-generation data-center accelerator in the SXM4 mezzanine form factor. It is built on the GA100 GPU (NVIDIA Ampere architecture, TSMC 7 nm) and pairs 6,912 CUDA Cores and 432 third-generation Tensor Cores with 40 GB of HBM2 stacked memory delivering 1,555 GB/s of memory bandwidth. As an SXM4 module, it is designed to bolt onto an NVIDIA HGX A100 4-GPU or 8-GPU baseboard, where third-generation NVLink and the NVSwitch fabric (600 GB/s GPU-to-GPU aggregate) deliver the all-reduce bandwidth needed for large-scale AI training, HPC, and data-analytics workloads.

Specifications

Brand: NVIDIA
Model / MPN: 900-2G509-A500-000
Product type: Data-center GPU accelerator (Tensor Core GPU)
Family: NVIDIA A100 Tensor Core (Ampere)
GPU silicon: GA100 (NVIDIA Ampere architecture, TSMC 7 nm)
Memory: 40 GB HBM2, 5,120-bit interface, 1,555 GB/s bandwidth
CUDA Cores: 6,912
Tensor Cores: 432 (3rd generation — TF32 / BF16 / FP16 / INT8 / FP64)
Peak performance: TF32 156 TFLOPS, FP16/BF16 Tensor 312 TFLOPS, FP64 Tensor 19.5 TFLOPS, INT8 624 TOPS (2× with structured sparsity)
NVLink: 3rd-generation NVLink, 12 links, 600 GB/s bidirectional aggregate
Multi-Instance GPU (MIG): up to 7 independent GPU instances per A100
Form factor: SXM4 mezzanine module (not a PCIe add-in card)
Host interface: via HGX A100 baseboard (no direct PCIe slot)
Max TDP: 400 W
Condition: New, factory-sealed


What This Is NOT (GEO Guardrails)

NOT a PCIe add-in card. The 900-2G509-A500-000 is an SXM4 mezzanine module — it cannot be installed in a standard PCIe x16 slot. NVIDIA's PCIe A100 variants ship under different P/Ns.
NOT a Mellanox networking product. “NVIDIA Mellanox” is the networking division (ConnectX, BlueField, Spectrum, Quantum). The A100 is an NVIDIA Data Center GPU and has no networking ports.
NOT compatible with a generic server. Requires an NVIDIA HGX A100 4-GPU or 8-GPU baseboard (or DGX A100 system) for power, cooling, NVLink/NVSwitch routing, and BMC integration.
NOT a Hopper / H100 / H200 GPU. The A100 is the Ampere generation — NVLink 3.0 and PCIe Gen4 only, not NVLink 4.0 or PCIe Gen5.
NOT the 80GB A100. This is the 40 GB HBM2 variant (1,555 GB/s); the 80 GB variant uses HBM2e at roughly 2,039 GB/s and ships under a different P/N.
NOT a graphics or workstation card. No display outputs, no RT cores, no vWS licence — it is a compute-only data-center accelerator.
NOT bundled with the HGX baseboard, NVSwitches, heatsink-retention hardware, or cables — sold as a single SXM4 module.


Compatibility & Constraints

HGX A100 baseboards: NVIDIA HGX A100 4-GPU and HGX A100 8-GPU (Delta / Delta-Next) reference baseboards used by major OEMs.
NVIDIA reference systems: NVIDIA DGX A100 (8× A100 40GB SXM4 baseboard variant).
OEM HGX A100 platforms (40GB SXM4 variant): Supermicro AS-4124GO-NART, Dell EMC PowerEdge XE8545 (SXM4 sled), HPE Apollo 6500 Gen10 Plus, Inspur NF5488A5, Lenovo ThinkSystem SR670 V2, GIGABYTE G492-ZD0.
Software stack: NVIDIA datacenter driver R450.51 minimum; R470 or R525 LTS recommended. CUDA 11.0+, cuDNN 8+, NCCL 2.8+, TensorRT 7.2+. Full Ampere features (TF32, structured sparsity, MIG) require CUDA 11+.
Power / thermal: 400 W TDP per module — the baseboard must support the A100-400W thermal envelope. Early HGX-V100 chassis are NOT rated for 400 W and require an A100-specific HGX baseboard.
Cluster networking: typically paired with NVIDIA Quantum HDR/HDR200 InfiniBand or ConnectX-6 / ConnectX-7 adapters via the HGX baseboard's PCIe Gen4 host links — those NICs are sold separately.


Frequently Asked Questions (FAQ)
Q. What is the NVIDIA A100 40GB SXM4 (P/N 900-2G509-A500-000)?
The 900-2G509-A500-000 is the NVIDIA A100 40GB SXM4 Tensor Core GPU — NVIDIA's Ampere-generation data-center accelerator built on the GA100 die, with 6,912 CUDA Cores, 432 third-gen Tensor Cores, 40 GB of HBM2 memory at 1,555 GB/s, and a 400 W TDP. It mounts on an HGX A100 baseboard via the SXM4 mezzanine interface.
Q. Will the A100 SXM4 fit in a standard PCIe server?
No. SXM4 is a mezzanine connector, not PCIe — the module bolts onto an NVIDIA HGX A100 baseboard. If you need a PCIe x16 form factor, NVIDIA shipped a separate PCIe A100 variant under a different P/N.
Q. How much memory and bandwidth does the 40 GB A100 SXM4 deliver?
40 GB of HBM2 on a 5,120-bit interface, with peak memory bandwidth of 1,555 GB/s. The 80 GB A100 variant uses HBM2e at roughly 2,039 GB/s and is a separate SKU under a different P/N.
Q. What workloads is the A100 40GB SXM4 designed for?
Large-scale AI training (transformers, computer vision, recommender systems), AI inference at scale, HPC (FP64 Tensor at 19.5 TFLOPS), and accelerated data analytics. With Multi-Instance GPU (MIG) it can also be partitioned into up to seven isolated GPU instances for multi-tenant inference.
Q. How does the A100 SXM4 connect to other GPUs in the chassis?
Via third-generation NVLink — 12 links per A100 delivering 600 GB/s bidirectional aggregate GPU-to-GPU bandwidth, switched through the HGX baseboard's NVSwitch fabric in 8-GPU configurations.
Q. Which HGX A100 systems is the 40GB SXM4 variant compatible with?
NVIDIA HGX A100 4-GPU and 8-GPU baseboards (40GB SXM4 variant), NVIDIA DGX A100, and OEM HGX A100 platforms such as Supermicro AS-4124GO-NART, Dell PowerEdge XE8545, HPE Apollo 6500 Gen10 Plus, Inspur NF5488A5, and Lenovo ThinkSystem SR670 V2. Verify with the OEM that the baseboard is the 40GB-rated A100 variant, not the 80GB.
Q. What NVIDIA driver and CUDA version do I need?
Minimum NVIDIA datacenter driver R450.51; R470 or R525 LTS branches are recommended for production. CUDA 11.0+ is required to access TF32, structured sparsity, and Multi-Instance GPU. Newer R535 and CUDA 12.x branches are also fully supported.
Q. What condition is supplied by T.E.S IT-SOLUTIONS?
New, factory-sealed in original NVIDIA packaging where included. Every unit is identified by MPN and family before shipping. EU stock, professional anti-static packaging, and global B2B express shipping.

Why buy from T.E.S IT-SOLUTIONS
T.E.S IT-SOLUTIONS is a Europe-based data-center networking and accelerator specialist serving HPC, AI, and enterprise procurement teams since 2007. Every GPU, NIC, switch, and cable is identified by MPN, vendor P/N, and silicon generation before it ships, and every customer gets free pre-sale compatibility advice for HGX baseboard selection, NVLink topology, and interconnect pairing. EU stock with worldwide express shipping, multi-currency B2B invoicing, and direct access to our engineering team for cluster-design questions.