deploybase

L4 on AWS: Pricing, Specs & How to Rent

Deploybase · April 2, 2025 · GPU Pricing

Contents

L4 GPU Specifications

NVIDIA L4: low-power inference GPU. 24GB GDDR6. Great cost-per-performance for production inference on AWS g6 instances.

Key specifications:

  • 24GB GDDR6 memory
  • 7,424 CUDA cores
  • 30.3 TFLOPS peak FP32 performance
  • 300 GB/s memory bandwidth
  • Maximum power consumption: 72W
  • Single-slot form factor
  • Hardware video encoding (NVENC)
  • TensorRT optimization support

The L4's power efficiency makes it ideal for dense deployments where power and cooling are constraints. Unlike training-focused GPUs, the L4 balances performance and cost for inference operations.

AWS L4 Pricing & Availability

AWS offers L4 GPUs through multiple instance families, each targeting different workload patterns and scaling needs.

Standard L4 instance pricing:

  • g6.xlarge (1x L4): approximately $0.80/hour on-demand
  • g6.12xlarge (4x L4): approximately $3.22/hour on-demand
  • Regional variation: US regions typically cheapest
  • Spot pricing: 60-70% discount off on-demand rates
  • 1-year savings plans: 20-30% discount
  • 3-year savings plans: 40-45% discount

As of March 2026, AWS L4 pricing remains competitive for production inference. The g6 instance family integrates well with AWS services like SageMaker and Lambda.

Compare to Lambda GPU pricing which offers fixed rates independent of commitment, or RunPod pricing for alternative spot-like flexibility.

How to Rent L4 on AWS

Deploying L4 instances on AWS EC2 follows standard EC2 provisioning workflows:

  1. Log into AWS Console
  2. Go to EC2 Dashboard
  3. Select "Launch Instances"
  4. Choose Amazon Machine Image (Ubuntu, Amazon Linux, or custom)
  5. Select instance type (g6.xlarge or g6.12xlarge)
  6. Configure instance details (VPC, subnet, IAM role)
  7. Add storage (EBS volume for OS and model storage)
  8. Configure security group (enable SSH, ports for inference API)
  9. Review and launch
  10. Select key pair or use AWS Systems Manager Session Manager

AWS provides NVIDIA drivers pre-installed on GPU-optimized AMIs. Users can deploy inference frameworks like TensorRT, ONNX Runtime, or vLLM directly on launch.

Integration with AWS SageMaker simplifies multi-instance deployment. The managed service handles scaling, load balancing, and endpoint monitoring automatically.

AWS L4 vs. Alternative Inference GPUs

L4 competes with various GPU options across clouds, each suited to different inference scenarios.

Vs. T4: Older Tesla T4 GPUs offer lower cost but 3x slower throughput. L4 preferred for modern workloads despite higher cost.

Vs. A10: A10 on Lambda costs around $0.86/hour with stronger performance than L4 for many models. A10 better for mixed workloads, L4 for pure inference.

Vs. L40: L40S specifications include 48GB memory versus L4's 24GB. L40S better for large models, L4 for cost-efficient smaller model inference.

Vs. Google Cloud GPU pricing: Google's L4 pricing similar to AWS. Choice depends on ecosystem integration preferences.

The L4 strikes balance between cost and performance, making it popular with startups and teams scaling inference from development to production.

L4 Performance Benchmarks

L4 performance varies significantly by workload, with particular strength in inference scenarios.

LLM inference (with quantization):

  • Llama 2 7B: 45-60 tokens/second
  • Mistral 7B: 50-70 tokens/second
  • TinyLlama 1B: 150+ tokens/second
  • Throughput improves with batch size (up to 8)

Image generation:

  • Stable Diffusion: 4-6 images/minute at 512x512
  • SDXL: 2-3 images/minute at 1024x1024
  • Real-time improvement with model optimization

Video processing:

  • H.264 encoding: Full HD real-time (30fps)
  • H.265 encoding: Full HD at 15-20fps
  • Decoding: supports 4-8 concurrent HD streams

Recommendation systems:

  • Dense embedding lookup: 1000s requests/second
  • Feature extraction: 100-200 batch inference/second

The L4's memory bandwidth enables efficient batching. Real-world deployments often saturate the GPU with batch sizes of 8-16, achieving near-peak throughput.

FAQ

Is L4 suitable for real-time inference? Yes, L4 latency ranges from 10-50ms depending on model size. Perfect for API endpoints requiring sub-100ms response times.

Can I run multiple models on one L4? Yes, NVIDIA's Multi-Process Service (MPS) allows multiple inference processes per GPU. Carefully partition memory to avoid contention.

Does AWS L4 support TensorRT optimization? Yes, TensorRT is fully supported. Optimization typically yields 2-5x speedup over baseline CUDA kernels.

What's cheaper: AWS L4 or spot GPU markets? Spot instances on AWS cost 60-70% less. Vast.ai offers similar discounts but with availability variability. Trade stability for cost savings.

How does L4 compare to RTX 4090 for inference? L4 is datacenter-grade with better cooling and ECC memory. RTX 4090 offers more raw compute but lacks production support.

Sources