GPU Use Cases
L4 use cases
The L4 is an Ada Lovelace data-center GPU optimized for efficiency. At just 72W TDP, it delivers 24 GB of GDDR6 memory and 60.6 TFLOPS of FP16 throughput in an exceptionally low power envelope, making it ideal for high-density inference and edge-adjacent deployments.
Efficient inference
The L4 is well-suited for inference serving of models up to 13B parameters where power efficiency and rack density matter. Its 24 GB GDDR6 fits 13B models at FP16 or larger models with quantization. At 72W TDP, it enables dense deployments without specialized cooling. Its 300 GB/s bandwidth is modest, so autoregressive generation of larger models will be memory-bound.
Edge and low-power deployments
The L4's 72W TDP makes it practical for edge-adjacent deployments where power constraints are significant. It is a common choice for inference at the edge, video processing, and vision pipelines where Ada Lovelace efficiency is needed without the cost of larger GPUs.
Video and vision pipelines
The L4 includes Ada Lovelace hardware-accelerated video encoding and decoding, making it well-suited for video transcoding, computer vision, and media processing pipelines. Its low power consumption enables dense video processing deployments.
Batch scoring and embeddings
For asynchronous batch scoring, embedding generation, and retryable inference queues, the L4 delivers reliable throughput at a low cost per watt. Its Ada Lovelace architecture provides better efficiency than older Turing and Ampere GPUs for the same workloads.