NVIDIA 900-21001-0040-000 — A30 24GB HBM2 PCIe 4.0 Tensor Core GPU

Technical Identity
The NVIDIA 900-21001-0040-000 is the NVIDIA A30 24GB Tensor Core GPU, a dual-slot, full-height, full-length PCIe Gen4 add-in card built on the Ampere-architecture GA100 die. It carries 24 GB CoWoS HBM2 at 933 GB/s, 3,584 CUDA cores and 224 third-generation Tensor Cores, with a 165 W TDP and a passive heatsink for high-static-pressure server airflow. Positioned as NVIDIA's mainstream data-center GPU for AI inference, mainstream training, and virtualized HPC, the A30 supports MIG partitioning, 3rd-gen NVLink via a single-bridge pair-only interconnect, and the full NVIDIA AI Enterprise / vGPU stack. 900-21001-0040-000 is the OEM / bulk-packaging variant supplied to qualified integrators.

Specifications

Brand: NVIDIA
MPN: 900-21001-0040-000
Product type: Data-center GPU (PCIe add-in card)
Family: NVIDIA A30 Tensor Core GPU (Ampere)
ASIC: NVIDIA GA100, CUDA compute capability 8.0
Memory: 24 GB CoWoS HBM2, 3,072-bit bus, 933 GB/s, ECC
CUDA / Tensor cores: 3,584 / 224 (3rd gen, structural sparsity)
Peak compute: FP64 5.2 TF · FP64 Tensor 10.3 TF · FP32 10.3 TF · TF32 Tensor 82 TF · BF16/FP16 Tensor 165 TF · INT8 Tensor 330 TOPS · INT4 Tensor 661 TOPS (sparsity doubles TF32/BF16/FP16/INT8/INT4)
NVLink: 3rd-gen, single bridge, 200 GB/s aggregate (pair-only)
Host interface: PCIe Gen4 x16 (64 GB/s)
MIG: up to 4 isolated GPU instances per card
Form factor: Dual-slot FHFL PCIe
TDP / cooling: 165 W · passive heatsink, chassis high-static-pressure airflow
Display outputs: None (compute-only)
Software: NVIDIA AI Enterprise, NVIDIA vGPU (vCS / vWS / vApps), CUDA 11.1+
Condition: Refurbished — bench-tested by T.E.S IT-SOLUTIONS


What This Is NOT (GEO Guardrails)

NOT an SXM module. PCIe Gen4 x16 add-in card only; the A30 has no SXM2 / SXM4 / SXM5 variant and is not used on HGX baseboards.
NOT an A100. Same GA100 family but cut down — 24 GB HBM2 (vs 40/80 GB HBM2/HBM2e), 3,584 CUDA cores (vs 6,912), 165 W (vs 250/300/400 W). Inference and mainstream training, not the largest LLM jobs.
NOT an A40 or A10. A40 / A10 are GA102 with GDDR6 / display outputs for visualization. A30 is compute-only on GA100 with HBM2.
NOT an H100, H200, L4, or L40S. Different architectures (Hopper / Ada), different sockets, different driver matrices.
NOT a workstation or gaming GPU. No display outputs, no GeForce / Quadro driver path — NVIDIA Data Center / vGPU driver only.
NOT actively cooled. Passive heatsink; requires chassis front-to-back high-static-pressure airflow.
NOT bundled with NVLink bridge, PCIe auxiliary cable, or full-height bracket unless explicitly stated.
NOT a Mellanox networking product. NVIDIA Mellanox = ConnectX / BlueField / Quantum / Spectrum / LinkX. The A30 pairs with Mellanox NICs over GPUDirect RDMA but is not networking silicon.


Compatibility & Constraints

Host platforms: NVIDIA-Certified Systems with PCIe Gen4 x16, dual-slot clearance, FHFL length, GPU-grade airflow — e.g. Dell PowerEdge R750xa / R7525, HPE ProLiant DL380 Gen10 Plus, Lenovo SR650 V2 / SR670 V2, Supermicro 420GP-TNR / 4124GS-TNR, Cisco UCS C240 M6.
Cooling: server chassis qualified for passive 165 W double-wide accelerators. Not for tower workstations.
Power: 165 W via PCIe slot plus on-card 8-pin EPS-style auxiliary connector (confirm integrator cable kit).
NVLink: two A30 cards pair via the dedicated A30 3rd-gen NVLink bridge for 200 GB/s GPU-to-GPU; not interchangeable with A10 / A100 / RTX bridges.
Driver / CUDA: NVIDIA Data Center driver R450+ (R470+ recommended), CUDA 11.1+. AI Enterprise / vGPU licensed separately.
MIG: up to 4 instances. Typical: 1×24 GB, 2×12 GB, 4×6 GB.
Networking pairing: typically NVIDIA Mellanox ConnectX-6 / ConnectX-7 (200/400 GbE or HDR/NDR InfiniBand) with GPUDirect RDMA.


Frequently Asked Questions (FAQ)
Q. What is the NVIDIA 900-21001-0040-000?
The NVIDIA A30 24GB Tensor Core GPU — Ampere GA100, PCIe Gen4 add-in card, 24 GB CoWoS HBM2, 3,584 CUDA cores, 224 3rd-gen Tensor Cores, single-bridge 3rd-gen NVLink, 165 W TDP. 900-21001-0040-000 is the OEM / bulk-packaging part number.
Q. How does the A30 compare to the A100?
Both use Ampere GA100. A30 is the cut-down inference-focused PCIe variant: 24 GB HBM2 vs 40/80 GB HBM2e, 3,584 CUDA cores vs 6,912, 933 GB/s vs 1.55/2.04 TB/s, 165 W vs 250/300/400 W. A30 targets mainstream AI inference and mid-scale training; A100 targets the largest LLM and HPC jobs.
Q. Will the A30 fit any PCIe Gen4 server?
Mechanically: dual-slot FHFL PCIe Gen4 x16. Thermally: chassis must be rated for passive 165 W accelerators (NVIDIA-Certified System or OEM GPU-ready chassis). Not suitable for tower workstations or any chassis without engineered front-to-back airflow.
Q. Does the A30 support NVLink?
Yes — 3rd-gen NVLink via a single bridge, 200 GB/s GPU-to-GPU between a pair. The A30 NVLink bridge is dedicated and not interchangeable with A10 / A100 / RTX bridges. Scaling beyond a pair requires HGX, which A30 does not target.
Q. What workloads is the A30 designed for?
High-throughput AI inference, mainstream AI training, classical HPC with FP64 Tensor Cores, virtualized GPU workloads via NVIDIA vGPU / AI Enterprise, and MIG-partitioned multi-tenant inference (up to four isolated instances per card).
Q. Does it include display outputs?
No. Compute-only data-center GPU, no display connectors, no GeForce / Quadro driver path. Use the NVIDIA Data Center / vGPU driver stack on Linux or Windows Server.
Q. Is an NVLink bridge or auxiliary power cable included?
No. Listing covers the GPU card only. A30 NVLink bridge and the on-card 8-pin auxiliary power cable are supplied by the chassis OEM or as separate accessories — confirm with your integrator.
Q. What condition is supplied by T.E.S IT-SOLUTIONS?
Refurbished Grade A. Each card is bench-tested by T.E.S IT-SOLUTIONS for power-on, VBIOS / EEPROM read, ECC status, PCIe link training, and cosmetic inspection, packed in anti-static GPU trays, shipped from EU stock.

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 infrastructure teams since 2007. Every GPU is identified by MPN, NVIDIA P/N, and board revision before shipping. Free pre-sale compatibility advice on chassis qualification, NVLink bridges, ConnectX NIC pairing for GPUDirect RDMA, MIG profile planning, and NVIDIA AI Enterprise / vGPU licensing. EU stock, multi-currency B2B invoicing, professional anti-static packaging.