900-2G133-2780-030 NVIDIA L40S 48GB GDDR6 PCIe Gen4 Passive GPU

NVIDIA datacenter GPU / accelerator — H-series (Hopper) / A-series (Ampere) / L-series (Ada Lovelace) / B-series (Blackwell) / Tesla (Legacy Volta/Pascal/Maxwell) for AI training / inference / HPC / VDI deployments.

Specifications

• Part Number: 900-2G133-2780-030
• Brand: NVIDIA
• Family: gpu
• Memory: 48 GB GDDR6 with ECC
• Form Factor: Full-height, full-length (FHFL), dual-slot / double-width, 4.4 in (H) x 10.5 in (L), PASSIVE heatsink (no fan)
• Interface: PCI Express Gen4 x16 (64 GB/s bidirectional)
• TDP: 250W (A10/L4) / 300W (A100 PCIe) / 400W (H100 PCIe) / 700W (H100/H200 SXM5) / 1000W+ (B200 SXM6)

Compatibility

NVIDIA datacenter GPUs require GPU-validated chassis with proper power + cooling + airflow. Verify chassis GPU-support matrix + power-supply capacity (typically 1100W+ per GPU socket pair) before order.

• H-Series (Hopper, 2022+): H100 (80GB), H200 (141GB HBM3e), H800 (China-only) — transformer engine + FP8 — for large-model LLM training/inference
• A-Series (Ampere, 2020+): A100 (40/80GB), A30, A40, A10, A2 — general-purpose AI/HPC; A100 remains workhorse for training
• L-Series (Ada Lovelace, 2022+): L40 (48GB graphics+inference), L40S (graphics+gen-AI), L4 (24GB low-power inference) — visual computing + inference
• B-Series (Blackwell, 2024+): B100, B200, GB200 NVL72 (Grace+Blackwell) — flagship LLM training
• Tesla legacy: V100 (16/32GB HBM2 Volta), P100 (Pascal), P40/P4 (Pascal inference), M40/M60 (Maxwell), K80 (Kepler) — EOL but production-stable
• NVIDIA-Certified Systems: Dell PowerEdge R750xa/R760xa, HPE Apollo 6500/DL380a Gen11, Lenovo SR670/SR675 V3, Cisco UCS C-series M7, Supermicro SuperGPU — validated airflow + driver stack + NGC catalog
• HGX baseboards (multi-GPU): HGX H100 / HGX H200 / HGX B200 with 4x/8x SXM modules + NVLink fabric — for DGX-class density

Use Cases

AI training (large transformer models, LLM fine-tuning — GPT/Llama/Mistral on 8x H100 / 8x H200), AI inference (LLM serving via Triton, NIM, TensorRT-LLM), HPC simulation (CFD, FEA, molecular dynamics, weather modeling), scientific compute (CUDA/cuBLAS/cuDNN), VDI acceleration (vGPU via NVIDIA AI Enterprise), media analytics (real-time video inference). H200 = best for memory-bound LLM serving (141GB HBM3e). B200 = next-gen flagship for LLM training scale-out.

Ships worldwide from ICD, Cairo, Egypt. All parts are tested, inspected, and backed by ICD's quality guarantee.