NVIDIA · Ada Lovelace · 2023
NVIDIA L40S. Rental by the hour, on demand.
From $0.55/hr per GPU on an 8-GPU machine, pay as you go with no commitment. 48 GB of GDDR6 and 362 TFLOPS FP16 tensor, for training and inference.
L40S on demand
on an 8-GPU machine
- Availability
- Available now · 6 regions
- By region and size
- $0.55 to $2.15/hr
- Billing
- By the second on most GPUs, a few by the minute
- 8+ GPUs or a month
- Launch any size yourself, or prepay a machine for 1 to 12 months at a discount. Custom quotes on request.
The card
48 GB of GDDR6. At 864 GB/s.
- Memory
- 48 GB
- GDDR6
- Bandwidth
- 864 GB/s
- Memory to compute
- FP16 tensor
- 362
- TFLOPS
- Per node
- 4 GPUs
- PCIe, 350 W each
Price across clouds
NVIDIA L40S, cloud by cloud. Sorted by price.
On-demand rates per GPU-hour from each cloud's public price page on 8 Nov 2026: RunPod's Secure Cloud, and Lambda's lowest per-GPU rate across its cluster sizes. Runcrate's row is today's live starting rate.
Launch it
Ask your agent for an L40S. Or read the call it makes.
Claude Code or Codex runs it for you: it starts the instance, copies your files over and runs the job. The SDKs make the same request.
Connect your agent first
Add Runcrate's MCP server to yourself: run curl -fsSL "https://runcrate.ai/mcp.md" and follow it.
Start an L40S on Runcrate called my-gpu, copy train.py to it and run it.
create_instance { name: "my-gpu", gpu: "L40S" }instance_status { id: "my-gpu" }file_upload { id: "my-gpu", content: "<train.py>", path: "/root/workspace/train.py" }ssh_execute { id: "my-gpu", command: "nohup python /root/workspace/train.py > train.log 2>&1 &" }Spec sheet
How the L40S compares. Against its closest alternatives.
Benchmarks
The L40S, measured. Real runs, dated, with what they cost.
- LLM inference
Qwen2.5-7B on vLLM, three GPUs (run 2026-10-26); Prefix caching on vs off (run 2026-10-27)
- A100 vs H100 vs H200
One script on six GPUs (run 2026-10-25)
- Speech to text
Whisper large-v3 audio hours per dollar (run 2026-10-28)
- Text embeddings
BGE embedding throughput per dollar (run 2026-10-29)
- Training and fine-tuning
Full fine-tune vs LoRA vs QLoRA on one L40S (run 2026-10-28); torch.compile vs eager for training (run 2026-10-29)
- Image generation
SDXL images per dollar on four GPUs (run 2026-10-27)
- GPU startup time
Launch to first CUDA matmul (run 2026-10-26); Hugging Face download vs local cache (run 2026-10-29)
Alternatives
Alternatives to the L40S. Same account, different price and performance.
| GPU | Memory · FP16 tensor | Stands out | Price gap | From |
|---|---|---|---|---|
| NVIDIA L40SThis GPU | 48 GB GDDR6 · 362 TFLOPS | $0.55/hr | ||
| AMD MI250X | 128 GB HBM2e · 383 TFLOPS | +80 GB memory | Not in stock right now | |
| NVIDIA A100 80GB | 80 GB HBM2e · 312 TFLOPS | +32 GB memory | +215% | $1.73/hr |
| NVIDIA A100 40GB | 40 GB HBM2 · 312 TFLOPS | +262% | $1.99/hr | |
| NVIDIA RTX 5090 | 32 GB GDDR7 | -16 GB memory | +31% | $0.72/hr |
Full specs
NVIDIA L40S specifications.
Memory
- Memory
- 48 GB
- Type
- GDDR6
- Bandwidth
- 864 GB/s
Compute
- FP32
- 91.6 TFLOPS
- FP16 tensor
- 362 TFLOPS
- INT8
- 733 TOPS
- Tensor cores
- 568
- CUDA cores
- 18,176
Power and form
- TDP
- 350 W
- Form factor
- PCIe
- Architecture
- Ada Lovelace
- Released
- 2023
Node
- GPUs per node
- 4
- Interconnect
- PCIe
- Maker
- NVIDIA
Questions
Renting the L40S. What buyers ask first.
Other GPUs
Browse every GPU by architecture.
NVIDIA Blackwell
NVIDIA Hopper
NVIDIA Ampere
NVIDIA Ada Lovelace
NVIDIA Volta