Overview
The NVIDIA L4 is a low-power GPU designed for cloud AI inference, released in 2023 and based on the Ada Lovelace architecture. With just 72W TDP, it supports passive cooling, delivers 485 TFLOPS FP8 compute, and offers 3× the inference performance of the T4. The L4 supports AV1 hardware encode/decode, making it suitable for cloud gaming, real-time video AI, and generative AI inference.
Core Specifications
| Item | Spec |
|---|
| Architecture | Ada Lovelace AD104 |
| Process | TSMC 4N |
| Memory | 24 GB GDDR6 |
| Memory Bandwidth | 300 GB/s |
| CUDA Cores | 7,680 |
| Tensor Cores | 240 (4th Gen) |
| FP32 | 30.3 TFLOPS |
| FP8 Tensor Core | 485 TFLOPS |
| INT8 Tensor Core | 485 TOPS |
| TDP | 72 W (passive cooling capable) |
| PCIe | Gen 4 x16 |
| Video Codec | 2× NVENC (AV1) + 4× NVDEC (AV1) |
| Metric | T4 (Turing) | L4 (Ada) | Improvement |
|---|
| FP16 Compute | 65 TFLOPS | 121 TFLOPS | 1.9× |
| INT8 Compute | 130 TOPS | 242 TOPS | 1.9× |
| FP8 Compute | N/A | 485 TFLOPS | New |
| Memory | 16 GB | 24 GB | +50% |
| Bandwidth | 320 GB/s | 300 GB/s | Slightly lower |
| TDP | 70 W | 72 W | Same |
| Video Codec | 1× NVENC | 2× NVENC (AV1) | +100% |
Use Cases
- Cloud AI inference (generative AI, LLM)
- Cloud gaming (high-density deployment)
- Video AI (transcoding, content moderation, AV1 streaming)
- Edge AI inference