GPUDatacenterReleased Mar 21, 2023

NVIDIA L4

192-bit · 7,424 cores

A low-profile, single-slot Ada Lovelace datacenter GPU that draws its full 72 W from the PCIe slot without an auxiliary power connector. Announced in March 2023 as the successor to the Turing-based T4, it targets AI inference, video transcoding, virtual desktops and cloud gaming in mainstream and edge servers, with dedicated media engines that support AV1 encoding and decoding.

General

Manufacturer: NVIDIA
Model: L4
Segment: Datacenter
GPU chip: AD104
Architecture: Ada Lovelace
Release date: Mar 21, 2023
Price: Search on Amazon

Memory

Memory size: 24 GB
Memory type: GDDR6
Bus width: 192-bit
Bandwidth: 300 GB/s
Memory clock: 1,563 MHz

Clocks & silicon

GPU clock: 795 MHz
Boost clock: 2,040 MHz
Shader units: 7,424 CUDA Cores
Tensor / AI cores: 232
RT cores: 58
Transistors: 35.8 billion
Technology (nm): 5 nm
TSMC 4N
TDP / TBP: 72 W

Interface & APIs

Bus interface: PCIe 4.0 x16
SLI / multi-GPU: None (no NVLink)
DirectX: 12 Ultimate (12_2)
OpenGL: 4.6
Vulkan: 1.4
Other APIs: CUDA 8.9, OpenCL 3.0

AI support

Precisions: FP8FP16BF16INT8INT4TF32FP64
AI accelerator: 4th Gen Tensor Cores (232) + 3rd Gen RT Cores (58)
Peak AI TOPS: 485 TOPS
Notes: Fourth-generation Tensor Cores with FP8 support. Peak throughput with structured sparsity: FP8 and INT8 485, FP16 and BFLOAT16 242 TFLOPS, TF32 120 TFLOPS; dense rates are half. FP32 is 30.3 TFLOPS. Media engines: 2 NVENC, 4 NVDEC and 4 JPEG decoders, with AV1 encode and decode.