Rent NVIDIA L4 GPU by the hour
The NVIDIA L4 is an Ada Lovelace architecture datacenter GPU equipped with 22 GB of VRAM and a 72 W power envelope. It delivers about 24.2 TFLOPS of single-precision floating-point performance and about 221 GB/s of memory bandwidth with CUDA compute capability 8.9. Rent it on ONIE with per-hour billing, full SSH access, and ready templates for PyTorch/CUDA, vLLM, Ollama, and ComfyUI.
Specifications
VRAM per GPU
Architecture
Compute capability
FP32 compute
Memory bandwidth
Power limit
Values reported by the infrastructure provider for machines currently listed; real performance varies by machine.
Available configurations
Live stock and prices, updated continuously. The final price is confirmed before you order.
Loading live prices…
What the L4 is good for
Language Model Inference
Run quantized or small-to-medium language models in the 22 GB VRAM buffer using pre-configured vLLM or Ollama templates.
Image Generation Workflows
Deploy ComfyUI pipelines that fit comfortably within 22 GB of memory while utilizing Ada Lovelace architecture features.
Lightweight Fine-Tuning
Execute parameter-efficient fine-tuning and adapter training jobs in PyTorch/CUDA environments without needing a high-wattage datacenter card.
Batch Media and Vision
Process computer vision inference and video workloads on a datacenter GPU rated at 72 W.
Is it the right GPU for you?
Choose the L4 when you need 22 GB of VRAM and Ada Lovelace compute capability 8.9 for inference or moderate training workloads. Move to a larger datacenter GPU if your models exceed 22 GB or require higher than about 221 GB/s of memory bandwidth, or pick a smaller card for basic development tasks that need less memory.
Compare with similar GPUs
Frequently asked questions
Is the NVIDIA L4 suitable for LLM inference?
Yes, the 22 GB of VRAM allows you to host small and medium language models or quantized weights. You can launch vLLM or Ollama templates on ONIE to start serving models right away.
How does the L4 compare to larger datacenter GPUs?
The L4 is designed for efficiency with a 72 W power draw, about 24.2 TFLOPS of single-precision compute, and about 221 GB/s of memory bandwidth. Larger datacenter GPUs offer more VRAM and higher memory bandwidth for heavy training, while the L4 focuses on cost-effective inference and light workloads.
What software templates are available for the L4 on ONIE?
You can deploy the L4 using ready templates for PyTorch/CUDA, vLLM, Ollama, and ComfyUI. All instances include SSH access so you can customize your software environment and CUDA compute capability 8.9 workflows.
How does billing work for L4 instances?
ONIE bills for L4 instances on a per-hour basis. You only pay for the time your instance is active, making it practical for both short experiments and ongoing inference tasks.
Rent the L4 by the hour
Create an account, choose a configuration and connect over SSH. Billing stops when you delete the GPU.
Rent L4