NVIDIA L40S · Ada Lovelace
L40S dedicated server 48 GB of ECC memory, three AV1 encoders, root on your own OS.
An L40S dedicated server is one physical machine with an NVIDIA L40S GPU passed through to the operating system you install, billed by the month. The lowest complete price is $532.67 a month, for Intel Xeon Silver 4110 8 Core 2.10 GHz · 16 GB DDR4 · 2 × SATA-SSD 240 GB (RAID 1) · 1 Gbps with the GPU fitted.
- Single tenant
- PCIe passthrough
- Unmetered port
- AV1 NVENC
- No contract
The L40S is the GPU we fit for mixed work: LLM inference and fine-tuning, ray-traced rendering and hardware video encoding on one GPU. This page gives its datasheet figures, the price of every platform that takes it, the PCIe link it gets on each, and where an A100 or H100 suits the job better.
Datasheet
NVIDIA L40S specifications
The L40S is a passive, dual-slot PCIe data-centre GPU on the Ada Lovelace architecture, the same AD102 chip family as the RTX 6000 Ada, with ECC memory and a 350 W board limit. NVIDIA quotes its Tensor figures twice, dense and with structured sparsity; the table gives both and says which is which.
| Specification | NVIDIA L40S |
|---|---|
| Architecture | Ada Lovelace |
| GPU memory | 48 GB GDDR6 with ECC |
| Memory bandwidth | 864 GB/s |
| CUDA cores | 18,176 |
| Tensor Cores | 568, fourth generation |
| RT Cores | 142, third generation |
| Interface | PCIe 4.0 x16 |
| Board power | 350 W |
| Compute | Dense, and with sparsity |
|---|---|
| FP32 | 91.6 TFLOPS |
| TF32 Tensor | 183 TFLOPS 366 with sparsity |
| FP16 / BF16 Tensor | 362 TFLOPS 733 with sparsity |
| FP8 Tensor | 733 TFLOPS 1,466 with sparsity |
| INT8 Tensor | 733 TOPS 1,466 with sparsity |
| RT Core | 212 TFLOPS |
| Video engines | 3 NVENC, 3 NVDEC, AV1 encode and decode |
| MIG / NVLink | Neither |
The L40S also supports NVIDIA vGPU software and secure boot with a hardware root of trust, and it is NEBS Level 3 ready. On our servers the full GPU goes to your operating system, so splitting it with vGPU would be your own licence and your own hypervisor.
From the order catalogue
L40S server prices, with the machine the GPU goes in
The L40S is on 2 platforms. Each row is the full monthly bill for the cheapest machine on one of them: processor, memory, the disks the order form opens on, an unmetered port and one IPv4 address, plus the GPU. Configure opens that platform with the L40S picked, where you add memory, disks or a faster port.
| Platform | PCIe link | Cheapest machine | Cities | Per month |
|---|---|---|---|---|
| Intel Xeon Silver / GoldL40S $449/mo | PCIe 3.0 x16GPU is 4.0 | Intel Xeon Silver 4110 8 Core 2.10 GHz · 16 GB DDR4 · 2 × SATA-SSD 240 GB (RAID 1) · 1 Gbps | New York, Bucharest, San Francisco, Miami | $532.67Configure |
| AMD EPYCL40S $449/mo | PCIe 4.0 x16full speed | AMD EPYC 7413 24 CORE 2.65 GHz 128MB L3 CACHE · 32 GB DDR4 · 1 × SATA-SSD 240 GB · 1 Gbps | New York, Bucharest | $666.59Configure |
The L40S is a PCIe 4.0 GPU. On a PCIe 3.0 platform the link runs at 16 GB/s each way instead of 32, which slows loading weights and streaming batches from host memory but not work that stays in the GPU's own 864 GB/s memory. A build with a GPU fitted pays a one-time setup of $29. Paying 3, 6 or 12 months ahead takes up to 15% off, and the price does not rise at renewal.
L40S vs A100 vs H100
L40S, A100 or H100: the three data-centre GPUs we fit
The A100 and H100 carry HBM with more than twice the bandwidth and NVLink connectors, and they have no video encoder. The L40S has less memory bandwidth, adds RT Cores and three NVENC engines, and costs less per month. All three are the PCIe versions, the only form our chassis take.
| Specification | L40S | A100 80 GB | H100 80 GB PCIe |
|---|---|---|---|
| Architecture | Ada Lovelace | Ampere | Hopper |
| GPU memory | 48 GB GDDR6, ECC | 80 GB HBM2e, ECC | 80 GB HBM2e, ECC |
| Memory bandwidth | 864 GB/s | 1,935 GB/s | 2,039 GB/s |
| FP8 | Yes | No | Yes |
| RT Cores | 142 | None | None |
| Video encode | 3 NVENC with AV1 | None | None |
| NVLink | No | Bridge connector | Bridge connector |
| Multi-Instance GPU | No | Up to 7 | Up to 7 |
| Board power | 350 W | 300 W | 350 W |
| Interface | PCIe 4.0 x16 | PCIe 4.0 x16 | PCIe 5.0 x16 |
| GPU line, per month | $449 | $899 | $1,599 |
| Complete server from | $532.67 | $972.47 | $1,682.67 |
Pick the A100 or H100 when a model needs more than 48 GB on one GPU, when training is memory-bandwidth bound, or when two GPUs must exchange data over NVLink. Both are on A100 and H100 servers.
Workloads
What an NVIDIA L40S GPU server runs well
Start from memory. The weights, the KV cache and the runtime have to fit in 48 GB; a job that does not fit is not saved by a faster chip.
AI
Inference and fine-tuning
Serving 13B models at 16-bit with room to spare, 30B at 8-bit or 70B at 4-bit with a short context, and LoRA or QLoRA fine-tuning. FP8 Tensor Cores speed up inference where your runtime supports FP8.
Graphics
Rendering and 3D
RT Cores for renderers that use them through OptiX, such as Blender Cycles, and 48 GB for scenes that run a 24 GB consumer GPU out of memory.
Video
Encoding and transcoding
Three NVENC and three NVDEC engines with AV1, so FFmpeg, OBS or a transcoder encodes in hardware and leaves the CUDA cores for the rest of the pipeline.
Will the model's weights fit in 48 GB?
| Model weights | 16-bit | 8-bit | 4-bit |
|---|---|---|---|
| 7-8 billion parameters | fits16 GB | fits8 GB | fits5 GB |
| 13-14 billion parameters | fits28 GB | fits14 GB | fits8 GB |
| 30-34 billion parameters | no68 GB | tight34 GB | fits19 GB |
| 70 billion parameters | no140 GB | no70 GB | tight38 GB |
Weights only. "Fits" means half the weights' size again is still free for the KV cache and the runtime. "Tight" loads, but a long prompt or a second user can push the 48 GB over. For a given model, LLM hosting names the machine, and for live channels in hardware, GPU transcoding servers run Flussonic or Wowza on the GPU.
Included
L40S bare metal: what you control and how it is billed
No virtual machine and no shared GPU: the L40S belongs to one machine and the machine to you. The machine and the GPU are two monthly lines on one invoice, and changing the GPU later moves the bill by the difference while your disks and IP addresses stay.
Root and a remote console
Your operating system, installed clean, with root, plus the machine's out-of-band controller (iLO, iDRAC or BMC) for the screen, virtual media and power.
Drivers you pin
The NVIDIA driver, the CUDA release and the framework stay at the versions you choose, and nothing updates them for you. To have a driver installed before handover, write its version on the order.
An unmetered port
Traffic is not metered in either direction, so datasets, checkpoints and encoded video add nothing to the bill. Faster ports are on the order form and on unmetered bandwidth.
Month to month
Pay a month at a time and stop when the job ends. Hourly billing is only for a machine already racked and ordered unchanged on instant servers; a build stays monthly.
Delivery
How long an L40S server takes
It depends on where the GPU is when you order, and you are told which case applies before you pay. Each time below counts from approval of the order.
- 30 min
Already racked
A machine that carries an L40S is waiting in the rack.
- 4–24 h
Fitted from our stock
We hold the GPU and fit it to the machine you configured.
- 5–10 days
Bought in
Working days, when we have to buy the L40S in first.
- 10–15 days
On allocation
Working days, when the supplier sets the date.
GPU marketplace
L40S machines listed by their owners
Owners also list their own L40S machines on our marketplace, each at the monthly price shown on its listing. Listings from our sister company, Primcast, show here too and are rented there.
Answers
Questions about L40S servers
What is an L40S dedicated server?
A physical server with one NVIDIA L40S 48 GB GPU passed through to the operating system you install, with root access, an unmetered port and nobody else on the machine. It is billed by the month, with the machine and the GPU as two lines on one invoice.
How much does an L40S server cost?
The lowest monthly price for a complete L40S server is $532.67. That machine is Intel Xeon Silver 4110 8 Core 2.10 GHz · 16 GB DDR4 · 2 × SATA-SSD 240 GB (RAID 1) · 1 Gbps, and the GPU is $449 of the total. A build with the GPU fitted adds a one-time setup of $29, and the builds table lists every platform with its own price.
How many CUDA cores does the NVIDIA L40S have?
18,176 CUDA cores, with 568 fourth-generation Tensor Cores and 142 third-generation RT Cores, on the Ada Lovelace architecture. It carries 48 GB of GDDR6 with ECC at 864 GB/s, connects over PCIe 4.0 x16 and draws up to 350 W.
Is the L40S better than the A100 for AI?
For inference on a model that fits in 48 GB, often yes: it has FP8 Tensor Cores, which the A100 lacks, and it costs less per month. For training large models the A100 80 GB usually wins, with 80 GB of HBM2e at 1,935 GB/s against 48 GB at 864 GB/s, and an NVLink connector. The H100 is ahead of both on training.
Can an L40S run a 70B model?
At 4-bit, yes, with little room: the weights take about 38 GB of the 48, which leaves space for a short context and one user. At 8-bit, 70 billion parameters need about 70 GB, which takes two GPUs or an 80 GB one.
Does the L40S support NVLink or MIG?
Neither. The L40S has no NVLink connector and no Multi-Instance GPU, so two L40S GPUs in one server talk over PCIe, and one L40S runs as one GPU. If the job needs either feature, look at the A100 or H100.
Is the L40S good for video encoding?
Yes. It has three NVENC and three NVDEC engines with AV1 encode and decode, so FFmpeg, OBS or a transcoder encodes in hardware. The A100 and H100 have no NVENC at all.
Can I rent an L40S on demand, by the hour?
Yes, but only for an L40S machine already racked and ordered unchanged from instant servers. It costs its monthly price divided by 730 hours, times 1.3, metered by the second until you release it, with one hour as the minimum. Every build in the table above is monthly.
Is this an L40S cloud instance or bare metal?
Bare metal. There is no hypervisor between your operating system and the GPU, and the machine has one tenant. You see the full L40S with its 48 GB, not a slice of it.
Can I put two or more L40S GPUs in one server?
Yes, as a build quoted by hand rather than ticked at checkout: the chassis, its PCIe lanes and its power supply set how many fit. The GPUs talk over PCIe, since the L40S has no NVLink. Ask in chat before you order, or see multi-GPU servers.
How long does an L40S server take to deliver?
About 30 minutes after approval when a machine with an L40S is already racked, 4 to 24 hours when we fit a GPU we hold, 5 to 10 working days when the GPU is bought in, and 10 to 15 when it is on allocation. We tell you which of the four applies before you pay.
Who answers when I need help?
Start a chat: our AI replies in seconds and can look up your account and your L40S server as you talk. Say you want a person and one of us takes over the same conversation.
In plain text
Every figure on this page
- GPU
- NVIDIA L40S, Ada Lovelace, 48 GB GDDR6 with ECC, 864 GB/s, PCIe 4.0 x16, 350 W
- Cores
- 18,176 CUDA cores, 568 fourth-generation Tensor Cores, 142 third-generation RT Cores
- Compute
- FP32 91.6 TFLOPS; FP16 and BF16 Tensor 362 TFLOPS (733 with sparsity); FP8 733 TFLOPS (1,466 with sparsity)
- Video
- 3 NVENC and 3 NVDEC engines, AV1 encode and decode
- L40S server from
- $532.67 a month, complete: Intel Xeon Silver 4110 8 Core 2.10 GHz · 16 GB DDR4 · 2 × SATA-SSD 240 GB (RAID 1) · 1 Gbps with the GPU
- GPU line
- $449 a month on top of the machine
- Platforms
- Intel Xeon Silver / Gold (PCIe 3.0); AMD EPYC (PCIe 4.0)
- Cities
- New York, Bucharest, San Francisco, Miami
- Setup
- $29 once for a build with a GPU fitted
- Term
- Monthly, no contract; up to 15% off for 3, 6 or 12 months paid ahead
- Traffic
- Unmetered port, no egress charge
- GPU access
- PCIe passthrough to your own OS, root, single tenant, no hypervisor
- Delivery
- About 30 minutes racked; 4 to 24 hours fitted from stock; 5 to 10 working days bought in; 10 to 15 on allocation
- Not on the L40S
- NVLink, Multi-Instance GPU, HBM
Elsewhere
Other GPU questions
Pick the platform, then configure the L40S server
Open the configurator with the L40S already chosen, or tell us the model, the scene or the channel count in chat. The AI replies within seconds; a person takes over whenever you ask.