What Is the NVIDIA L4?

The NVIDIA L4 is a data centre GPU built for inference rather than training. NVIDIA launched it in 2023 as the successor to the T4, using the same Ada Lovelace architecture as its RTX 40-series and RTX Ada cards. It carries 24 GB of GDDR6 memory, fourth-generation Tensor Cores and dedicated video engines on a low-profile, single-slot PCIe card.

Most data centre GPUs draw hundreds of watts and need large, actively cooled chassis. The L4 GPU draws 72 W, which a PCIe slot can supply without an extra power cable, and it has a passive heatsink that relies on the system’s airflow. That is why rugged computer makers can build it into sealed and compact systems for inference beside the cameras.

It has no display outputs. The L4 is a compute and video card, so a system that needs monitors uses the processor’s graphics or a separate card for them.

NVIDIA aimed the L4 at workloads that run all day on many streams, where performance per watt matters more than peak speed.

Much of the L4’s work at the edge is video. NVIDIA lists two NVENC encoders, four NVDEC decoders and four JPEG decoders on the card, so a single L4 can decode dozens of camera streams in hardware and pass the frames straight to the AI model in GPU memory. AV1 encoding lets a system record or relay those streams at a lower bitrate than H.264 for the same quality.

The 24 GB of GDDR6 runs at around 300 GB/s over a 192-bit bus and connects to the host through PCIe Gen4 x16. That is enough memory to hold several detection and tracking models at once, or a small language model alongside a vision pipeline. Because the card is passive and slot-powered, the same part can be fitted in a 2U server, a fanless rugged chassis or a sealed IP69K enclosure.

The L4 sits between NVIDIA’s entry inference card, the A2, and the larger L40S. Its closest comparison is the T4, which it replaces in the same low-profile, slot-powered format.

Figures are NVIDIA’s published peak values; real throughput depends on the model, the precision and the cooling. FP32 measures general floating-point speed, and our TFLOPS explainer covers why that differs from the TOPS figures quoted for inference.

A passive card has no fan of its own, so it depends on the chassis to push air through its fins. In a server this comes from front-to-back fans. In a rugged edge computer the system is designed around the card, with ducted fans or a sealed enclosure that conducts the heat to the case.

This is why an L4 should not be fitted to an ordinary desktop PC. Without directed airflow it reaches its thermal limit and slows down. In a system built for it, the card holds its clocks at full load across the rated temperature range.

The L4 brings data centre inference to places a server rack cannot go: vehicles, ships, mines and outdoor cabinets. Our maritime surveillance case study and autonomous mining trucks case study use it in sealed computers, and our rugged GPU computers include L4 builds.

Where a project also needs monitor outputs from the GPU, a card such as the RTX 2000 Ada, at a similar power, is often the better fit; the L4 suits headless inference and video.

Our NVIDIA L4 computers range from IP69K and tactical systems to compact inference servers. We fit the L4 in each system as standard, install the drivers and validate the GPU under load before it ships.