NVIDIA

L4

Name: NVIDIA L4
Brand: NVIDIA
Availability: InStock

L4 — illustration of the card's form factor

VRAM

24 GB

FP32 TFLOPS

30.3 TFLOPS

TDP

72 W

Memory

GDDR6

Provider Marketplace

Cheapest

$0.25/hour

Starting from

Google Cloud Visit

Best Value

$0.60/hour

Starting from

RedSwitches Visit

Enterprise Choice

$2.20/hour

Starting from

Yotta Data Services Visit

All Cloud Providers

11 Options available

Google CloudCheapest

Spot · preemptible

$0.25/ hour

Estimated Cost

Provision

RunPod

On-Demand

$0.44/ hour

Estimated Cost

Provision

Jarvis Labs

On-Demand

$0.44/ hour

Estimated Cost

Provision

Lightning AI

On-Demand

$0.48/ hour

Estimated Cost

Provision

E2E Networks

On-Demand

$0.57/ hour

Estimated Cost

Provision

RedSwitches

On-Demand•MontrealCanada

$0.60/ hour

Estimated Cost

Provision

Modal

On-Demand

$0.80/ hour

Estimated Cost

Provision

Baseten

On-Demand

$0.85/ hour

Estimated Cost

Provision

OVHcloud

On-Demand

$1.00/ hour

Estimated Cost

Provision

Runcrate

On-Demand

$1.05/ hour

Estimated Cost

View Provider

Yotta Data Services

On-Demand

$2.20/ hour

Estimated Cost

Provision

Estimates only — rates are collected automatically from public provider pages and may be out of date. Prices vary by region, commitment term, and availability, and typically exclude storage, egress, and tax. Confirm current pricing with the provider before purchasing. Last collected Sep 1, 2026.

Compute Performance

FP3230.3 TFLOPS

TF32120 TFLOPS

FP16242 TFLOPS

BF16242 TFLOPS

FP8485 TFLOPS

INT8485 TOPS

Architecture

MicroarchitectureAda Lovelace

Base Clock795 MHz

Boost Clock2,040 MHz

Sparse AccelerationSupported (Shown with sparsity)

Dynamic PrecisionSupported (FP8/FP16/BF16/TF32/INT8/FP32)

Memory & VRAM

Memory TypeGDDR6

Total Capacity24 GB

Bandwidth300 GB/sec

Bus Width192 bits

ECC SupportEnabled (by default); can be disabled using software

Connectivity & Scaling

InterconnectPCIe

GenerationPCIe Gen4

IB Bandwidth64GB/s

PCIe InterfaceGen4 xx16

Max GPUs/Node8

Virtualization

SR-IOVSupported: 32 VF (virtual functions)

vGPU ReadinessSupports vGPU 15.2 or later

GPU SharingSR-IOV (32 VF) and vGPU (vGPU 15.2 or later)

Power & Efficiency

TDP72 W

Peak Power72

Thermal LimitsPassive cooling; requires system airflow; ambient operating temperature 0°C to 50°C (short term -5°C to 55°C); NEBS-3 capable

Physical Design

Form FactorPCIe

Slot WidthSingle-slot

WeightBoard: 270 Grams (excluding bracket)

CoolingPassive

Thermals & Cooling

Temp Range0°C to 50°C

DC Heat72W low-power envelope

Software Ecosystem

CUDACUDA 12.0 or later

PyTorchSupports PyTorch inference

Driver StabilityDrivers R525 or later (Linux and Windows)

Server & Deployment

PreconfiguredPartner and NVIDIA-Certified Systems with 1–8 GPUs

Edge DeployCloud and at the Edge

System Compatibility

Required PCIePCIe Gen4 x16

MotherboardHalf-height (low profile), half-length (HHHL), single-slot PCIe (NVIDIA Form Factor 5.5)

Rack Power72 W maximum per board

BIOS LimitsSBIOS and software support in the OS instance or hypervisor required to enable SR-IOV

OS CompatLinux (drivers R525 or later); Windows (drivers R525 or later); Windows 10, Windows 11, Windows Server 2019, Windows Server 2022

Benchmarks & Throughput

Structured Sparsity

Shown with sparsity. Specifications are one-half lower without sparsity.

Inference Benchmarks

Claims include up to 120X higher AI video performance vs CPU; 2.5X image-generation performance vs T4 (512x512 Stable Diffusion v2.1, FP16, TensorRT 8.5.2); up to 1,040 concurrent AV1 720p30 streams.

Multi-GPU Scalability

Scaling Characteristics

ParallelismSR-IOV: 32 virtual functions (VFs); Virtual GPU: supports vGPU 15.2 or later; PCIe: Gen4 x16 interface (physical x16 lanes).

Workload Readiness

Diffusion Models

2.5X

Multimodal AI

supported

HPC / Simulation

supported

Edge Inference

supported

Market Authority

Cloud Adoption

Google Cloud (early access)

Key Strengths

Limitations

Also in the Lineup

Expert Insight

The L4 represents a strategic leap in AI compute. When comparing cloud providers, consider not just the hourly rate, but also the interconnect bandwidth (InfiniBand/NVLink) and regional availability which can significantly impact total cost of ownership for large-scale training.

Glossary Terms

FP32 TFLOPS

VRAM

TDP

Cores

Information updated daily. Cloud pricing subject to vendor availability.