NVIDIA L40S
Expert rating8.9 / 10
- GPU Brand:
- NVIDIA
- VRAM:
- 48GB GDDR6
- Memory Type:
- GDDR6
- TDP:
- 350W
Consumercarts is reader-supported. Learn more
Disclaimer: The price and specification of the given product may differ from the actual product. Please confirm on the retailer's website before buying.
Manufacturer: NVIDIA Last updated: 2026-09-12
NVIDIA L40S Review
If you're deciding between the NVIDIA L40S and the L40, the L40S is NVIDIA's tuned-up follow-up released in mid-2023, specifically optimized for higher generative AI training and inference throughput in the data center.
It shares the same 18,176 CUDA cores and 48GB of GDDR6 as the L40, but runs at a higher 350W TDP with improved sustained clocks and structural sparsity support tuned for transformer-based AI workloads, making it NVIDIA's go-to mainstream data center GPU for both AI and graphics duty when full H100/H200 budgets aren't available.
Performance & Use Case
It handles generative AI fine-tuning and inference, virtual workstation graphics, and rendering-farm workloads well, with its structural sparsity support giving a real boost in transformer-based AI training compared to the L40. It's built for data centers running mixed AI training/inference and graphics workloads at scale, not for desk-side workstation use.
Comparison to Alternatives
Against the NVIDIA L40 (nvidia-l40), the L40S offers meaningfully higher sustained AI training and inference throughput at the cost of 50W more power draw, with identical VRAM and CUDA core counts - for new deployments the L40S is generally the better buy unless power budget is the deciding factor. Against the NVIDIA RTX PRO 6000 Blackwell Server Edition (nvidia-rtx-pro-6000-blackwell-server-edition), the L40S is a full architecture generation behind and has half the VRAM (48GB vs 96GB), but remains widely deployed and considerably cheaper for teams not needing the newest Blackwell silicon.
Specifications
Platform
| GPU Brand | NVIDIA |
|---|---|
| Board Partner | NVIDIA Reference (data center/OEM only, passive cooling) |
Identity and type
| Series | RTX Ada Generation |
|---|---|
| Tier | Flagship |
Performance
| VRAM | 48 GB |
|---|---|
| Memory Type | GDDR6 |
| Boost Clock | 2520 MHz |
| Ray Tracing Support | Yes |
Physical & Power
| TDP | 350 W |
|---|---|
| Card Length | 267 mm |
| Slot Width | 2 |
Silicon
| Architecture | Ada Lovelace |
|---|---|
| Process Node | TSMC 4N |
| Shader Units | 18176 |
Memory
| Memory Bus Width | 384 bit |
|---|---|
| Memory Bandwidth | 864 GB/s |
Power
| Recommended PSU | 600 W |
|---|---|
| Power Connector | 1x 16-pin (12VHPWR) |
Host interface
| PCIe Generation | PCIe 4.0 x16 |
|---|
Market and workstation
| MSRP | 8500 USD |
|---|---|
| Launch Date | 2023-08 |
| Base / Game Clock | 735 MHz |
|---|---|
| Display Outputs | 4x DisplayPort 1.4a (rarely used in server deployment) |
| Media Engine | 3x NVENC (8th gen), 3x NVDEC, AV1 encode/decode |
Expert Review
Reviewed by Burooj Alam. Rated by us, not by users.
Expert rating8.9 / 10
One of the most widely deployed data center GPUs of its generation - a genuinely strong middle ground between pure graphics cards and dedicated AI accelerators.
Best for
- generative AI fine-tuning and inference at scale
- mixed graphics + AI data center deployments
- rendering-as-a-service infrastructure
Not for
- desktop workstation builds (passive cooling, server-only)
- gaming
- teams needing Blackwell-generation VRAM capacity (96GB+)
FAQs
What are the key specifications of the NVIDIA L40S?
The NVIDIA L40S is a graphics card from NVIDIA. Key specs: GPU Brand NVIDIA, Tier Flagship, Board Partner NVIDIA Reference (data center/OEM only, passive cooling), VRAM 48GB GDDR6, Memory Type GDDR6.
Who is the NVIDIA L40S best suited to?
It suits generative ai fine-tuning and inference at scale, mixed graphics + ai data center deployments, rendering-as-a-service infrastructure.
What are the drawbacks of the NVIDIA L40S?
It is a poor fit for desktop workstation builds (passive cooling, server-only), gaming, teams needing blackwell-generation vram capacity (96gb+).