GPU server · NVIDIA L40S

L40S dedicated server for AI, rendering and live video on one GPU

An L40S dedicated server from Primcast is a whole physical machine with one NVIDIA L40S GPU passed straight to your own operating system: 48 GB GDDR6 ECC, three AV1 hardware encoders and ray-tracing cores. You rent it by the month, with no contract and an unmetered port that is yours alone.

Build your L40S server See builds and prices

nvidia-l40s Ada Lovelace

GPU memory: 48 GB GDDR6 ECC
Memory bandwidth: 864 GB/s
CUDA cores: 18,176
Video engines: 3 NVENC + 3 NVDEC, AV1
Interface, power: PCIe Gen 4 x16, 350 W
NVLink: None

01What it is for

Four jobs the L40S GPU does well

Most GPUs are built for one kind of work. The NVIDIA L40S can run a model, render a frame and encode a stream on the same board, which is why people pick it for mixed pipelines.

01Video

L40S for video encoding

Three encoders and three decoders handle AV1, HEVC and H.264 in hardware. You can take in a feed, transcode it and serve it while a model on the same GPU decides what to do with the frames. The A100 and H100 have no video encoder at all, so on those GPUs video falls back to the processor. For channel counts and the full machine, see live video transcoding servers.

02AI

Inference on a model that fits in 48 GB

FP8 tensor support and 48 GB leave room for a 7B or 13B model at full 16-bit precision, or a 30B model at 4-bit. For an endpoint answering real users at a modest batch size, this GPU is plenty. The sizing table further down shows where it gets tight.

03Rendering

Long renders you do not want to repeat

Every gigabyte of memory is error-corrected, and the third-generation ray-tracing cores do the heavy lifting. Consumer GPUs cost less per frame, but a flipped bit on an unattended overnight job means running it again. ECC is what lets you start a render and walk away.

04Desktops

Virtual workstations

The L40S supports NVIDIA vGPU and has four DisplayPort outputs, so you can split it between your own virtual desktops. The A100 and H100 have no display outputs. Slicing is your choice on your machine: we hand the GPU over whole.

02Before you order

Two limits of the NVIDIA L40S

Both come from NVIDIA's own specification, and both are part of why the L40S costs less than the accelerators beside it.

Memory bandwidth, not stacked memory

The L40S reads its memory at 864 GB/s. The A100 reads at 1,555 GB/s with 40 GB and 1,935 GB/s with 80 GB, roughly twice as fast. When you generate tokens at a large batch size, that read rate sets the ceiling. So an A100 serves more requests a second from the same model, even though the L40S has more raw compute. If serving throughput is your whole problem, choose the accelerator.

No NVLink

Two L40S GPUs in one machine are two separate 48 GB pools that talk over the PCIe slot. They never become one 96 GB pool. The A100 and H100 can pool memory; the L40S cannot. If you need one address space larger than 48 GB, the comparison table below shows the GPUs that give you that.

Across the 57 cards in the catalogue: L40S holds 48 GB, 4 cards we fit hold more, and 8 cost more per month. Read the whole ladder on the whole catalogue, priced.

03Specification

L40S GPU specification

Ada Lovelace architecture, fourth-generation Tensor cores and third-generation RT cores. The figures are NVIDIA's, from its L40S product page.

L40S

Architecture

Ada Lovelace

Card memory

48 GB GDDR6 ECC

CUDA cores

18,176

Memory bandwidth

864 GB/s

Board power

350 W

Card interface

PCIe Gen 4 x16

Hardware video encoders

3 NVENC, 3 NVDEC, AV1

Error-correcting memory

Yes

NVIDIA also lists 91.6 teraFLOPS of FP32, 212 teraFLOPS of ray-tracing throughput and 1,466 teraFLOPS of FP8 tensor throughput with sparsity, from 142 RT cores and 568 Tensor cores. Those are peaks under ideal conditions. In practice, memory bandwidth usually decides what you get.

04Model sizing

Will your model fit in 48 GB?

Yes for 7B and 13B models at any precision, and for 30B at 4-bit. A 70B model only loads at 4-bit, with little room left over for context.

Weights only, rounded up. **Fits** means the card’s 48 GB holds the weights 1.5× over — room for the KV cache, the activations and the runtime, and the only verdict the LLM sizing page recommends a card on. **Tight** clears 1.25× only: the model loads, and a long context or a second user runs the card out.
Model	16-bit	8-bit	4-bit
7-8 billion parameters	16 GB — fits	8 GB — fits	5 GB — fits
13-14 billion parameters	28 GB — fits	14 GB — fits	8 GB — fits
30-34 billion parameters	68 GB — does not fit	34 GB — tight	19 GB — fits
70 billion parameters	140 GB — does not fit	70 GB — does not fit	38 GB — tight

Starting from the model instead? LLM server sizing works out how much GPU memory a model needs and which GPU clears it.

05The server around it

Which L40S GPU server to pick

The L40S is a PCIe Gen 4 x16 GPU. Each server it fits gives it a different link, and the difference matters for some jobs and not for others.

Intel Xeon Silver / Gold

L40S

Slot: PCIe 3.0, 48 lanes per socket. Card: PCIe Gen 4.

About half the bus the card was designed for. That costs a training loop streaming batches off the host; it costs an inference server or a transcoder nothing, because the weights are already on the card.

AMD EPYC

L40S

Slot: PCIe 4.0, 128 lanes per socket. Card: PCIe Gen 4.

Full width. The slot is at least the generation the card was designed for, so nothing is left on the table.

For inference or transcoding, the slot generation hardly matters: the model weights stay on the GPU and video frames are small. It matters for training that streams fresh batches from the host on every step. For that, pick the server whose slot matches the GPU.

06Compare

L40S vs A40, RTX 6000 Ada and A100

These are the GPUs either side of the L40S in our own catalogue, each with the cheapest complete machine it goes in.

Cheapest complete machine first. Every figure read from the order catalogue when this page was served.
Card	Card memory	The card	Cheapest complete machine
L40S	48 GB GDDR6 ECC	$449	$532.67
A40	48 GB GDDR6	$499	$572.47
RTX 6000 ADA	48 GB GDDR6 ECC	$625	$698.47
A100 40GB	40 GB HBM2	$699	$772.47
A100 80GB	80 GB HBM2e	$899	$972.47

The A40 is the Ampere GPU the L40S replaced: the same 48 GB, an older architecture, no AV1 and no FP8. The RTX 6000 Ada is the workstation GPU with the same memory and the same generation. Move to an A100 when bandwidth or pooled memory is what holds you back, and read A100 and H100 servers for the detail.

07Prices

L40S dedicated server prices

The GPU is one line on your invoice and the server is another. Each row shows the full machine, what the GPU adds and the monthly total, read from the same catalogue the checkout uses.

Read from the order catalogue when this page was served. The card is a separate line; the machine does not get dearer because of what is plugged into it.
Card	The machine it goes in	The card	Complete, per month
L40S	Intel Xeon Silver / Gold Intel Xeon Silver 4110 8 Core 2.10 GHz · 16 GB DDR4 · 2 × SATA-SSD 240 GB (RAID 1) · 1 Gbps · PCIe 3.0	$449	$532.67 Configure this build →
L40S	AMD EPYC AMD EPYC 7413 24 CORE 2.65 GHz 128MB L3 CACHE · 32 GB DDR4 · 1 × SATA-SSD 240 GB · 1 Gbps · PCIe 4.0	$449	$666.59 Configure this build →

PortUnmetered traffic in both directions, on a port that serves your machine alone.
AddressOne IPv4 address included.
TermMonthly by default, with no contract.
HelpOur chat assistant answers in seconds and can check your account and server. Ask for a person any time.

Want to compare every GPU we fit? GPU servers, all priced lists them the same way. Need a machine today? Instant servers lists what is already racked.

08Questions

L40S questions people ask before ordering

01How many CUDA cores does the NVIDIA L40S have?

The NVIDIA L40S has 18,176 CUDA cores, 568 fourth-generation Tensor cores and 142 third-generation RT cores. They sit beside 48 GB GDDR6 ECC read at 864 GB/s, on a PCIe Gen 4 x16 board that draws up to 350 W.

02Is the L40S good for video encoding?

Yes. It has three NVENC encoders and three NVDEC decoders, and all of them handle AV1 in both directions. The RTX 30 series cannot encode AV1, and the A100 and H100 cannot encode video in hardware at all. For a live transcoding setup built around a GPU, see live video transcoding servers.

03L40S or A100 for serving a model?

If the model fits in 48 GB and your traffic is moderate, the L40S costs less and also does video and graphics. If you run the GPU flat out at a large batch size, the A100 reads its weights about twice as fast and serves more requests a second. If the model needs more than 48 GB, pick the A100 80 GB.

04Can I pool two L40S GPUs into 96 GB?

No. The L40S has no NVLink, so two of them are two separate 48 GB pools that talk over the PCIe slot. Your framework can still split a larger model across them, with the slot as the link. For a single memory space above 48 GB, choose the A100 80 GB from the comparison table.

05Is the GPU shared, sliced or virtualised?

Not by us. You are the only tenant on the physical machine, and the GPU goes straight to your operating system with no hypervisor, no MIG partition, no vGPU profile and no time-slicing. You install the driver and CUDA version you want. The L40S supports vGPU, so you can slice it for your own virtual desktops if you choose.

06How soon can I have an L40S server?

It depends on where the GPU is, and you are told before you pay. A GPU we already hold is fitted within four to twenty-four hours. One we have to buy in takes five to ten working days. To start today, instant servers lists the machines already racked.

07Who helps me if something goes wrong?

Open the chat on any page. Our assistant replies in seconds, can look at your account and your server, and passes you to one of our team whenever you ask. You keep full root access to the machine the whole time.

Get an L40S server for your pipeline

Pick the server, add the L40S and see the monthly total before you pay. No contract, and the port is unmetered.

Build your L40S server Compare every GPU