Overview with affiliate links. We receive a commission if you click through via us. Read how
Compare

NVIDIA L40S

NVIDIA: specifications, comparisons and alternatives.

The NVIDIA L40S is a data centre GPU based on Ada Lovelace, designed for AI inference, rendering and visual workloads, among other things. According to the manufacturer, it has 48 GB of GDDR6 memory with ECC, 0.864 TB/s of bandwidth and a maximum power of 350 W. It is a dual-slot PCIe card, without NVLink.

Illustratie van NVIDIA L40S19points
Illustration, not a product photo

NVIDIA

48 GB
Memory
GDDR6 (ECC)
Memory type
0,86 TB/s
Memory bandwidth
350 W
Power (board power or TDP)

You can see the current price and availability with the partner as soon as a link has been set up.

How are the points calculated?

Clouwino points are an arithmetic summary of the specifications the manufacturer states, on a scale of 0 to 100. It is not a benchmark and not a judgement on quality or suitability: a product with fewer points may suit your situation better. Products are compared with other products of the same type. For each property we count how far a product sits between the lowest and highest value in the group:

  • FP32 compute (for: consumer), carries heavy weight
  • Memory (for: consumer), carries heavy weight
  • Memory bandwidth (for: consumer), carries heavy weight
  • Compute units (for: consumer), carries light weight
  • Boost clock (for: consumer), carries light weight
  • Memory (for: professional and data centre), carries medium weight
  • Memory bandwidth (for: professional and data centre), carries medium weight
  • FP16/BF16 (for: professional and data centre), carries heavy weight
  • FP8 (for: professional and data centre), carries medium weight
  • FP64 (for: professional and data centre), carries medium weight

If too much data is missing, we do not show points. Points are relative: when a new product is added, they can shift. Read more at how we compare.

Specifications of NVIDIA L40S

All specifications the manufacturer lists, per item. Want to see another product next to it? Pick one below.

Specifications according to the manufacturer
SpecificatieNVIDIA L40S
Overview
Clouwino points19 punten
General
ArchitectureAda Lovelace
Manufacturing processNot stated by the manufacturer
FormPCIe, dual-slot
LaunchNot stated by the manufacturer
Compute
FP32 compute91,6 TFLOPS
FP64Not stated by the manufacturer
FP16/BF16362,05 TFLOPS
FP8733 TFLOPS
FP4Not stated by the manufacturer
Performance notesNVIDIA states dense first, then with sparsity (*). FP16/BF16: 362.05 dense (733 with sparsity). FP8: 733 dense (1,466 with sparsity). TF32 183 dense / 366 sparse. FP32 91.6 TFLOPS without a label. FP64 and FP4 not stated.
Memory
Memory48 GB
Memory typeGDDR6 (ECC)
Memory bandwidth0,86 TB/s
Power and connections
Power (board power or TDP)350 W
GPU interconnectPCIe Gen4 x16: 64 GB/s bidirectional; no NVLink
Multiple instancesNot stated by the manufacturer

Note: Figures per GPU. Max. power consumption 350 W. MIG: not supported; NVLink: not supported. Peak values based on boost clock. Process node and launch date not on the page. The datasheet link was not opened (redirect).

Alternatives to NVIDIA L40S

Products with similar specifications from the same overview. This is not a ranking and not buying advice.

The L40S as a PCIe card

The L40S is built on the Ada Lovelace architecture and comes as a dual-slot PCIe card, with a PCIe Gen4 x16 connection of 64 GB/s bidirectional. That makes it easier to install in common servers than an SXM or OAM module, because no special GPU board is needed.

It is designed for broad use: inference of AI models, graphics applications and rendering. If you want to understand the difference from consumer cards, read graphics card or data centre GPU. The L40S supports neither MIG nor NVLink. The process node and launch date are not on the page.

That can be interesting if you are looking for a GPU for a mix of AI and graphics work in an ordinary rack server.

Because the card uses GDDR6 rather than HBM, it is in many respects a different category from the large training modules, which plays a role in the choice for your application.

What 48 GB GDDR6 and the TFLOPS figures tell you

GDDR6 with ECC is a different memory type from HBM: the bandwidth of 0.864 TB/s is lower, but it can be enough for many inference and graphics tasks. You can read how VRAM works in what is VRAM.

NVIDIA states the dense value first and then the value with sparsity. FP16/BF16 is 362.05 TFLOPS dense and 733 with sparsity. FP8 is 733 TFLOPS dense and 1,466 with sparsity. FP32 is 91.6 TFLOPS without a label. FP64 and FP4 are not mentioned. All values are peak values based on the boost clock and apply per GPU.

With 48 GB, a mid-sized model or a sizeable scene fits in the memory of a single card, although the actual use depends on your software.

What to look out for with the L40S

With a maximum of 350 W, the card needs a server with good airflow and sufficient power. Because there is no NVLink, all communication between GPUs goes via PCIe; for multi-GPU training, that is something to take into account.

  • Check whether your server platform supports dual-slot PCIe cards.
  • Use the guide on comparing GPU clouds if you would rather rent.
  • Compare dense with dense and sparse with sparse.

Also pay attention to the cooling of several cards side by side in one chassis.

You can see the current price and availability at the partner via the button on the page.

Frequently asked questions

Does the L40S have NVLink or MIG?

No, according to the data the L40S supports neither NVLink nor MIG. The connection runs via PCIe Gen4 x16 at 64 GB/s bidirectional. That leaves PCIe for communication between GPUs, which is relevant in multi-GPU setups.

What is the difference between dense and sparse on the L40S?

NVIDIA states the dense value first and then the value with sparsity. For FP8, that is 733 TFLOPS dense and 1,466 TFLOPS with sparsity. You only compare values with the same label.

How much memory and power does the L40S have?

The L40S has 48 GB of GDDR6 memory with ECC and a bandwidth of 0.864 TB/s. The maximum power consumption is 350 W. All figures apply per GPU.

Last updated on

Ask us
Compare Get in touch

Are you sure?