NVIDIA L40S GPU
Universal 48 GB Ada Lovelace GPU for inference, fine-tuning, graphics and video
- Model
- L40S
- Condition
- New, original packaging
- Availability
- Quoted on request
Price on request. Tell us the quantity and destination and we reply within one working day.
How we supply this item
Supplied through licensed partners. Server, networking and data-centre hardware is supplied through licensed partners of InTechnologies FZCO. The partner issues the contract and invoice; we handle sourcing, consolidation in Dubai and delivery coordination. Export compliance
Export compliance. This product is subject to US export regulations. How we handle it is described on the export compliance page. Export compliance
The NVIDIA L40S is a universal data-centre GPU based on the Ada Lovelace architecture, combining 48 GB of GDDR6 with ECC, 18,176 CUDA cores, 568 fourth-generation Tensor Cores and 142 third-generation RT Cores on a 350 W, dual-slot PCIe Gen4 card. FP8 Tensor throughput of 733 TFLOPS dense makes it a capable inference and fine-tuning accelerator for models into the tens of billions of parameters, while its RT Cores and three NVENC and three NVDEC engines with AV1 cover rendering, video processing and Omniverse workloads that pure AI accelerators cannot run.
The L40S is popular with enterprises consolidating generative AI, VDI and visualisation on one platform, with service providers hosting inference at scale, and with media companies running cloud rendering farms. It installs in nearly every mainstream 1U to 4U server, usually at four to eight GPUs per node, and it is a frequent choice where H100-class parts are unavailable or oversized for the workload. It is fully supported by NVIDIA vGPU software and the NVIDIA AI Enterprise suite.
InTechnologies FZCO sources NVIDIA L40S cards and L40S-equipped servers for corporate customers in Dubai, the UAE and nearby markets. Contracts and invoices for GPU hardware are issued by our licensed partners; InTechnologies takes care of sourcing, consolidation in Dubai and DDP or DDU delivery. The L40S is subject to US export regulations, and end-user verification of the customer and destination is completed before every order is confirmed.
Specifications
| Architecture | NVIDIA Ada Lovelace |
|---|---|
| CUDA cores | 18,176 |
| Tensor Cores / RT Cores | 568 (4th generation) / 142 (3rd generation) |
| GPU memory | 48 GB GDDR6 with ECC |
| Memory bandwidth | 864 GB/s |
| FP8 Tensor Core | 733 TFLOPS dense, 1,466 TFLOPS with sparsity |
| FP32 performance | 91.6 TFLOPS |
| Max power | 350 W |
| Interface | PCIe Gen4 x16 |
| Form factor | Dual-slot, 4.4 in × 10.5 in, passive cooling |
| Video engines | 3× NVENC, 3× NVDEC, AV1 encode and decode |
| NVLink | Not supported |
Questions buyers ask
When should I choose L40S over H100 or H200?
Choose L40S when the workload mixes inference with graphics or video, when models fit in 48 GB, or when the server only offers PCIe Gen4 slots and 350 W of cooling per card. For large-model training, HBM-based Hopper or Blackwell GPUs are the better fit.
Can several L40S cards share memory?
No. The L40S has no NVLink, so multi-GPU jobs communicate over PCIe. Tensor-parallel inference across cards works, but with lower interconnect bandwidth than NVLink systems.
What is the lead time for L40S?
L40S is a mature product with regular channel availability; typical lead time is a few weeks and is confirmed with the quotation.
Is the L40S subject to export controls?
Yes. It exceeds the US performance thresholds for controlled GPUs, so end-user verification is required and some destinations need a licence.
Also searched as
L40S · L40 S · L40S wholesale · buy L40S wholesale · L40S price Dubai · L40S bulk order · L40S supplier UAE · NVIDIA L40S · buy L40S GPU Dubai · L40S 48GB UAE · NVIDIA L40S supplier Middle East · Ada Lovelace data centre GPU Gulf · L40S server GPU distributor UAE
Related products
NVIDIA HGX B200 8-GPU Platform
Eight Blackwell GPUs with 1.4 TB of HBM3e for large-scale AI training and inference
NVIDIA HGX B300 8-GPU Platform
Eight Blackwell Ultra GPUs with 2.1 TB of HBM3e for reasoning-model inference at scale
NVIDIA DGX B200 System
Turnkey 10U AI system with eight Blackwell GPUs, 1.4 TB HBM3e and 72 PFLOPS FP8 training
NVIDIA GB200 NVL72 Rack-Scale System
Liquid-cooled rack with 72 Blackwell GPUs and 36 Grace CPUs in one 130 TB/s NVLink domain