NVIDIA L4
Run inference, image and video workloads on a 24 GB GPU. A practical starting point for models and batch jobs with smaller memory requirements.
From $0.44/GPU/hour · USD on-demand rate · Storage extra
Workloads
What to run on L4.
Choose for your workload’s memory needs, software support and measured runtime.
Smaller-model inference
Serve models that fit within 24 GB, allowing room for runtime overhead and request cache.
Image workflows
Run supported image-generation pipelines and adjust resolution and batch size to available memory.
Video processing
Use supported GPU-accelerated video software. Validate codec and application compatibility for your pipeline.
Choose your setup
Start with the environment you need.
Use a preconfigured container or manage your operating system in a GPU VM. The dashboard shows current GPU and region availability.
GPU Templates
Start with PyTorch, ComfyUI or another supported environment. Drivers and framework dependencies are preconfigured.
Explore templates↗GPU VMs
Get SSH and full root access when you need control over the operating system and runtime. Available configurations vary by GPU and region.
Explore VMs↗Persistent storage
Keep datasets and checkpoints on a filesystem you can attach to your own instances in the same region.
Explore filesystems↗Sizing
Start with memory. Then measure performance.
Model weights are only part of the working set. Leave room for everything the job needs while it runs.
For inference
Account for model precision, context length, KV cache and concurrent requests. A model loading successfully does not tell you how much traffic it can serve.
For training
Include activations, gradients and optimizer state. Batch size, sequence length and checkpointing change memory use.
For multiple GPUs
GPU memory is not automatically pooled. Use a framework and parallelism strategy that distribute the workload across devices.
Hardware specifications: NVIDIA L4. Software and workload affect realized performance.
Compare options
Compare memory and hourly rates.
Use this as a shortlist, then test your workload. Lower hourly pricing does not always mean a lower total job cost.
| GPU | Memory | Architecture | From / GPU / hour |
|---|---|---|---|
| NVIDIA H200 | 141 GB HBM3e | Hopper | $3.99 |
| NVIDIA H100 | 80 GB HBM3 | Hopper | $2.69 |
| NVIDIA B200 | 180 GB per GPU¹ | Blackwell | Request a quote |
| NVIDIA RTX PRO 6000 | 96 GB GDDR7 | Blackwell | $1.89 |
| NVIDIA A100 80 GB | 80 GB HBM2e | Ampere | $1.49 |
| NVIDIA A100 40 GB | 40 GB HBM2 | Ampere | $0.89 |
| NVIDIA L4 | 24 GB GDDR6 | Ada Lovelace | $0.44 |
Before you start
Common questions.
Practical details for choosing and using this product.
How do I get started with NVIDIA L4?
Choose a template or VM, select the GPU and region, and review the configuration and price in the dashboard before launch.
Is storage included in the GPU rate?
Storage is billed separately. Retained storage continues to incur charges while an instance is paused. Review the full configuration price before launching.
Will my model fit on one GPU?
This GPU has 24 GB GDDR6. Fit depends on weights, precision, framework overhead and workload state. For inference, also account for context and concurrency; for training, include activations and optimizer state.
Can I get more than eight GPUs?
For workloads spanning nodes, explore GPU clusters. Reserved capacity can be planned at 128, 256, 1,024 GPUs and beyond, subject to configuration and availability.
Get started with NVIDIA L4.
Launch a GPU instance or talk to our team about the right setup for your workload.