SECTIONS
EDITIONS
NETWORK
AIXX Pedia The encyclopedia of compute hardware

AIXX Pedia › Model Entry › NVIDIA L4

NVIDIA L4 low-power data center GPU official product image (vendor image)
Vendor image · AIXX Pedia (NVIDIA official site)

NVIDIA L4

Performance-per-watt Specialist · 24GB · 72W Half-height Single Slot · Transcoding/Inference Value Tier · Model Entry

Overview OVERVIEW

NVIDIA L4 is the 「performance-per-watt」specialist : Ada Lovelace architecture (per official page, the generation with DLSS 3 and ray tracing support[1]), 24GB GDDR6 、 72W Power draw, half-height single slot Form factor — unveiled at GTC in March 2022 alongside the H100 generation (per public reporting[3]), officially positioned as a general-purpose low-power accelerator for "video + AI + graphics": its video pipeline (officially stated 1,040 concurrent 720p30 AV1 streams[1]), generative AI inference, and VDI virtual workstations — three fronts at once.

The math is simple: INT8 Tensor Cores at 485 TOPS (with sparsity; dense 242.5) , FP8 at the same level, FP32 at 30.3 TFLOPS — the absolute compute isn't flagship-class, but 72W with no external power connector and a single-slot half-height profile that packs densely into 1U ) makes it an unbeatable TCO play in transcoding and inference workloads where "power and rack space dominate cost"[1]. The 2× NVENC + 4× NVDEC + 4× JPEG decoder media engine configuration is the densest of any single card in the family.

Positioning boundaries must be addressed first : 24GB of VRAM and 300GB/s bandwidth define what it is is not a large-model training card ; PCIe Gen4 host interface and 30.3 TFLOPS FP32 are likewise not flagship specs — the L4's value proposition only holds in the「video/inference/vision」three scenarios. Match the scenario first, then the price, before buying.

Use Cases FIT

AIXX Index (0–100): estimated by this Pedia via weighted public specifications — memory 40% + interconnect 30% + ecosystem 30%, with weights adjusted by use case. 90+ flagship tier · 75+ Excellent · 60+ usable · below 60 not recommended. The methodology is transparent and verifiable, and does not constitute a performance guarantee.

Video Transcoding and Streaming 95AIXX Index
Officially rated for 1,040 concurrent 720p30 AV1 streams[1]; with 2 NVENC + 4 NVDEC + 4 JPEG, its media engine density is a family first, with native AV1/HEVC encode hardware — the TCO ace for transcoding workloads.
Multimodal and Vision AI 85AIXX Index
4× JPEG hardware decoding feeds vision models, with DLSS 3/RT support for graphics pipelines[1]: a low-power production slot for vision AI like OCR/inspection/visual retrieval.
RAG / Mid-Size Model Inference 80AIXX Index
7B–14B-class FP16 fits resident on a single card, and hundred-B-class models are reachable at 4-bit Quantization (estimated from the 24GB figure); a low-power-bill option for always-on embedding/reranking/recommendation workloads.
VDI / Virtual Workstations 82AIXX Index
Officially listed with vWS support and RTX/DLSS graphics features[1]: 72W single-slot lets a 1U server pack multiple virtual workstations — density is cost.
Large model training 15AIXX Index
The physical ceiling of 24GB/300GB/s/72W (official figures[1]) — training belongs to H100/B300-class HBM big cards, and the L4 was never positioned there.

History HISTORY

March 2022 , L4 launched at GTC: introduced alongside the H100 generation, positioned as a general-purpose low-power accelerator for "video + AI + graphics" (per public reporting[3]).

Second half of 2022 , OEM cards from ASUS/Leadtek and others rolled out, entering the market primarily as OEM server appliance form factors (per public reporting[4]).

In 2023 , Google Cloud launched G2 instances (L4-powered) — the first major public cloud to offer L4 at scale (per public reporting[5]).

2024–2025 , twin tailwinds of AV1 adoption and generative AI inference growth: the video pipeline (transcoding/moderation/subtitling) and mid-size model inference became the workhorse configuration of the public cloud "value tier" (per public reporting[3]).

2026 , still available on the official page, with 1–8 GPU NVIDIA-Certified servers and partner systems listed (verified October 2026[1]).

Technical Specifications SPECIFICATIONS

ItemSpecifications (NVIDIA L4 · official figures, verified October 2026[1])
GPU architectureAda Lovelace generation (per official page, the DLSS 3/ray tracing support tier[1])
Memory24GB GDDR6
Memory Bandwidth300GB/s
INT8 Tensor Core485 TOPS (with sparsity; dense 242.5, official figures[1])
FP8 Tensor Core485 TFLOPS (with sparsity, official figures[1])
FP16/BF16 Tensor Core242 TFLOPS (with sparsity; dense 121-class, official figures[1])
FP3230.3 TFLOPS (official figures[1])
Media EngineNVENC ×2 | NVDEC ×4 | JPEG decoders ×4 (official figures[1])
Video CapabilityOfficially rated 1,040 concurrent 720p30 AV1 streams (official figures[1])
Power72W (officially "72W low-power envelope"[1]) — no external power connector required
Form FactorHalf-height single slot · low profile · PCIe
Host InterfacePCIe Gen4 x16 · 64GB/s (official figures[1])
Server ConfigurationsPartner/NVIDIA-Certified systems, 1–8 GPUs (official figures[1])

Specifications per the official NVIDIA L4 page (verified October 2026[1]); sparse/dense figures transcribed per the official page's "with sparsity" labeling; CUDA core count is not listed on the official page — verify with nvidia-smi before purchasing.

72WL4 · Half-height Single Slot 350WL40S 400WA100 700WH100 Board Power (TBP/TGP) Comparison · NVIDIA official figures (the L4's raison d'être: performance per watt)

Comparison · L4 vs T4 vs L40S CONTEXT

ItemNVIDIA T4 (previous generation)NVIDIA L4NVIDIA L40S
Memory16GB GDDR624GB GDDR648GB GDDR6
Memory Bandwidth320GB/s300GB/s864GB/s
INT8~130 TOPS class (generation figures)485 TOPS (with sparsity[1])~725 TOPS class (with sparsity, see our entry)
Media EngineNVENC ×1 | NVDEC ×2NVENC ×2 | NVDEC ×4 | JPEG ×4[1]NVENC ×3 | NVDEC ×3
Power70W72W350W
Form Factorhalf-height single slothalf-height single slotFull-height dual slot
One-line buying adviceLegacy baselineVideo/inference value tierLarge-VRAM inference tier

Comparisons are based on official product pages and this site's entry data; the T4 serves as the Turing-generation baseline for reference. Media engine and codec support follow the NVENC/NVDEC support matrix[6].

Family Members FAMILY

Ada Lovelace data center family lineup:

L4 (72W half-height single slot · 24GB · video/inference value tier · this entry)L40S (350W full-height dual slot · 48GB · large-VRAM inference tier · entry available on this site)L40 (350W · graphics/rendering focus · previous-generation positioning)RTX 6000 Ada (workstation flagship · entry available on this site)A2 (low-power entry tier · prior form factor)

Cooling & Deployment COOLING

L4's deployment profile: 72W with no external power (PCIe slot power suffices) + half-height single slot + passive cooling — multiple cards can tile densely in a 1U server, and power density is its core deployment advantage over 350W-class cards like the L40S (official form-factor figures[1]). Passive cooling means chassis airflow is a hard prerequisite : tower workstations or modded cases need verified airflow before mounting an L4, or it will throttle from overheating.

PCIe Gen4 host interface — even on Gen5 motherboards, bandwidth still runs at Gen4 (64GB/s, official figures[1]); just match host generation to the card at purchase, no need for a full Gen5 build.

Model Compatibility MODEL FIT

ScenarioFit per card at 24GB (estimated: weights × precision + KV/framework overhead; see the model selector for methodology)
7B–14B inferenceFP16 weights run 14–28GB-class: 7B FP16 or 14B FP8 resident on one card is the comfort zone (estimated)
70B-class at 4-bit Quantizationweights around 35GB-class — out of reach on one card; two-card sharding works (estimated, throughput limited by bandwidth)
Embeddings/RAG/Rerankingembedding and reranker models are generally under 10GB-class: an always-on, low-power production slot (estimated)
Vision/Multimodal4× JPEG hardware decode feeds image streams[1]: an acceleration slot for OCR/inspection/content moderation pipelines
Video transcodingOfficially rated 1,040 concurrent 720p30 AV1 streams[1] — standard fit for FFmpeg NVENC pipelines
Not applicableLarge-model training (dual ceiling of 300GB/s bandwidth and 24GB capacity, official figures[1])

Methodology: memory footprint ≈ parameter count × bytes per parameter (FP16=2B / FP8=1B / INT4=0.5B), KV and framework overhead excluded; same source as the selector's "memory budget," verifiable.

NVIDIA L4 official product image
NVIDIA L4 · NVIDIA official materials (AIXX Pedia watermark)

See the official NVIDIA page[1] and OEM vendors' (ASUS/Leadtek, etc.) product pages for low-profile single-slot form factor and heatsink details.

Video Overview VIDEO

Video walkthrough of this entry is in the production queue (see the video slot schedule for Chinese/multilingual workflow) — this section updates automatically once it goes live.

Software & Ecosystem ECOSYSTEM

The dual ecosystem of full-stack CUDA plus media: Inference side : standard support for TensorRT-LLM/Triton Inference Server/vLLM (FP8 via Ada Tensor Cores); Media side FFmpeg NVENC/NVDEC pipelines with native AV1 encode hardware support (codec matrix in official documentation[6]); Cloud side Public cloud instances such as GCP G2 offer the L4 as a standard "value tier" supply (per public reporting[5]).

Transcoding vs consumer cards (4090-class): data center cards carry no NVENC concurrent-session limit (per driver policy), so for 7×24 video pipelines the L4 over a consumer card is a dual choice of compliance and stability — a point low-bid proposals often conveniently omit.

Purchasing & Inspection PROCUREMENT

1
First question on channel form
The L4 ships primarily as OEM server appliances (official 1–8 GPU system figures[1]), with little loose-card retail circulation — beware pulled cards and engineering samples when buying loose cards; cross-check serial numbers against the appliance's customs declaration.
2
Confirm airflow first
: passive cooling means no onboard fan — verify chassis airflow and 1U packing layout before installation; thermal throttling in modified environments is the most common root cause of「underperformance」.
3
Verify codec requirements
For AV1/HEVC/region-specific codec requirements, verify each item against the NVENC/NVDEC support matrix[6] — unsupported codecs are a hidden Fail point in transcoding procurement.
4
Session-limit compliance
: a core reason to pick a data center card for commercial transcoding pipelines — no consumer-card NVENC session limit (per driver policy); write the compliance risk of cheap consumer-card proposals out of the contract.
5
Match host generation
PCIe Gen4 x16 interface[1] — on older platforms (Gen3), halved bandwidth impacts inference throughput; write host generation into acceptance criteria.
6
On-machine measurement prevails
Transcoding throughput (measured stream counts at target format), inference latency (your own model on TRT-LLM), Burn-in Test stability (no throttling within the 72W envelope) — software screenshots don't count.

Common risks

Scenario-mismatch risk
Using the L4 as a training card or large-model card is the first mismatch — the 24GB/300GB/s physical ceiling (official figures[1]) means it serves only video/inference/vision, three scenarios.
Insufficient-airflow risk
: a passively cooled card in a poorly ventilated case = long-term throttling — acceptance should include a 30-minute Full Load temperature log.
Loose-card provenance risk
: low retail circulation = pulled parts masquerading freely — triple-check serial number, cosmetic wear, and connector traces.

For loose-card purchases, third-party inspection or platform Buyer Fund Verification and inspection flows are recommended — Marketplace Search ›

How to Stress Test BENCH TEST

Judgment tools: nvidia-smi/DCGM (power/temperature/ECC status), bandwidthTest (memory bandwidth vs. the official 300GB/s rating[1]), gpu-burn (Burn-in Test, 72W envelope validation), FFmpeg benchmark (practical stream counts at target formats vs. the official 1,040-stream 720p30 AV1 figure[1]), TRT-LLM/vLLM benchmark (inference throughput and latency with your own models). Acceptance for a transcoding card hinges on two metrics: "measured stream count + no sustained throttling." Dedicated test video samples for this card are in production.

Shipping & Delivery LOGISTICS

StageKey Points
Form Factor & WeightHalf-height half-length compact single card — anti-static bag plus anti-static foam is sufficient; low shock-resistance requirements for full-box shipping
HeatsinkPassive cooling, no fan — avoid crushing the heatsink fins during transport; inspect exterior on arrival
Cross-border ComplianceCompute tier below consumer-flagship sensitivity to export controls — still defer to current lists and customs classification
InsuranceInsure by cargo value; for server-level delivery, acceptance is based on the complete server system

OEM & Variants VARIANTS

Variants / OEMForm Factor & Positioning
NVIDIA L4Low-profile single-slot standard form factor (official figures[1])
OEM cardPartner form factors from ASUS/Leadtek/Gigabyte and others (per public reporting[4])
NVIDIA-Certified servers1–8 GPU system integration (official figures[1])
Cloud instancesPublic cloud L4 instances such as GCP G2 (per public reporting[5])

Limitations & Controversies LIMITS

  • 24GB/300GB/s ceiling : low on both memory and bandwidth (official figures[1]) — out of the running for training and large-model scenarios; even 70B-class Quantization inference needs two-card sharding.
  • PCIe Gen4, not Gen5 : 64GB/s host bandwidth (official figures[1]) — data-throughput-sensitive inference services gain no bandwidth benefit on Gen5 platforms.
  • FP32 only 30.3 TFLOPS : mediocre graphics/simulation performance (official figures[1]) — it's a "video + AI" specialist card, not an all-rounder.
  • Limited retail availability : channel is appliance-dominant (official 1–8 GPU system figures[1]) — a small loose-card market with high risk of pulled cards.
  • 72W performance ceiling : the price of performance-per-watt is absolute performance — high-throughput transcoding farms need to stack cards by stream count; a single card can't compare to the L40S.

Market Notes MARKET

DateTypeMarketPrice BandBasis
2026-10Card priceOverseas channelsNo standing official card price · actual price subject to the seller's quote (primarily OEM appliance form[1])Channel basis
2026-10Cloud RentalOverseas public cloudBilled per GPU-hour (GCP G2 and similar instances[5]; actual price subject to the seller's quote on cloud pricing pages, varies by region)Third-Party Metrics
2026-10Marketplace StatusMainland ChinaFew loose cards in circulation · mostly servers/pulled partsMarketplace Search ›

Note: this database holds no standing transaction data for this model; "actual price subject to the seller's quote" means we do not anchor on one-sided asking prices. Market and inventory: AIXX Exchange (Zero-commission Matching; not a party to transactions).

Timeline MILESTONES

TimeEvent
2022-03L4 unveiled at GTC (sharing the stage with the H100 generation, per public reporting[3])
H2 2022OEM cards and server systems rolled out (per public reporting[4])
In 2023Google Cloud G2 instances launched — first major cloud to offer it at scale (per public reporting[5])
2024–2025AV1 adoption + generative AI inference growth: dual mainstays of video pipelines and "value tier" inference (per public reporting[3])
2026-10Available on the official page: 1–8 GPU certified systems listed (verified October 2026[1])

FAQ FAQ

What's the upgrade over the previous-gen T4?
One calculation says it all: INT8 jumps from the 130-TOPS class to 485 TOPS (with sparsity[1]), the media engine goes from 1+2 to 2 NVENC + 4 NVDEC + 4 JPEG[1], and AV1 encode hardware plus FP8/RT/DLSS 3 arrive — all while power barely moved (70W→72W). Both transcoding and inference are generational upgrades.
Can the L4 run a 70B model?
Barely, but not recommended: 4-bit Quantization of 70B weights is roughly 35GB-class (estimated), which doesn't fit on a single 24GB card; two-card sharding works — but at 300GB/s bandwidth, throughput will look ugly. 7B–14B resident models are its comfort zone; for 100B-class, look to the L40S/A100/H100.
How many streams can it actually transcode?
Officially rated 1,040 concurrent 720p30 AV1 streams[1] (raw stream passthrough basis) — real workloads with filters/watermarks/AI processing will drop that count significantly; benchmark your own pipeline.
L4 or L40S?
Choose by memory and power: for inference within 24GB plus heavy transcoding, pick the L4 (72W low power, dense 1U packing); for 48GB large-memory inference, pick the L40S (350W). For video farms sensitive to power and rack space, L4 usually wins on TCO; for single-card performance and large-model loading, L40S wins.
Why not use a consumer card (4090) for cheaper transcoding?
Three reasons: consumer cards have NVENC concurrent session limits (driver policy), no data-center-grade warranty for 7×24 sustained operation, and lack data-center features (ECC/virtualization) — video pipelines are production environments, and the L4 is the card built for them.

References REFERENCES

[1]NVIDIA official · L4 product page (24GB/300GB/s/485 TOPS with sparsity/72W/half-height single-slot/Gen4/NVENC2+NVDEC4+JPEG4/1,040-stream AV1 basis) · https://www.nvidia.com/en-us/data-center/l4/ (verified 2026-10)
[2]NVIDIA official · NVENC/NVDEC codec support matrix · https://developer.nvidia.com/video-codec-support (official documentation)
[3]Public reporting · GTC 2022 L4 launch and video/AI scaling momentum · https://www.reuters.com/technology/ (per public reporting)
[4]Public reporting · OEM card and server form factor availability (ASUS/Leadtek etc.) · https://www.reuters.com/technology/ (per public reporting)
[5]Google Cloud · GPU documentation (L4 on G2 instances) · https://cloud.google.com/compute/docs/gpus (third-party figures)
[6]This site · L40S entry (family comparison data source) · /wiki/l40s.html (This Site's Entry)
[7]Selector tool · VRAM budgeting methodology (parameters × precision + KV/framework overhead) · /wiki/xuanxing.html (this site's methodology)

Specs are anchored to NVIDIA's official page; "public reporting" items are media accounts and this database does not vouch for their accuracy; official figures such as transcoding stream counts are lab conditions — rely on real-world tests.

This Library
Editorial PolicyAbout This Library (Disclosure of Interests)Corrections Channel
Editorial Constitution Summary
Every claim sourcedDrawbacks always stated · corrections publicly loggedNo paid entries acceptedZero traffic-funneling or sales elements within entriesMarket data attached only as data modules
AIXX Family
AIXX PediaAIXX ExchangeAIXX News
Market overview: AIXX Exchange · © 2026 AIXX Pedia · An AIXX Product