Calculate the TPS of the L4 on local AI models
Every model in our catalogue assessed against this card at the context length and minimum quality you choose. Speed is an estimate for a single request, calculated from this card's memory bandwidth and the size of each model once compressed.
Calculated for this card
721 models in our catalogue altogether
Largest model it holds
Mixtral 8x7B
46.7B · Q3_K_M · 26.6 tok/s
Fastest model
Gemma 3 QAT 1B
127 tok/s · 1B
Which AI models can run on a L4?
Set the inputs, read the answer
More context means more memory for the conversation cache. Speed is for a fresh conversation and does not change with this setting.
Hides models that would only fit by being compressed below this point.
532 models match
Calculating| Quantisation | Fit | ||||||
|---|---|---|---|---|---|---|---|
|
127
tok/s
108–153 |
Gemma 3 1B | 1B | Mar 2025 | 1.8 GB | 33k tokens | Q8_0 | Comfortable |
|
127
tok/s
108–153 |
Gemma 3 QAT 1B | 1B | Apr 2025 | 1.8 GB | 33k tokens | Q8_0 | Comfortable |
|
127
tok/s
76–203 · low confidence |
HGRN 1B (WT 103) ≈ | 1B | Nov 2023 | 1.8 GB | 131k tokens ? | Q8_0 | Comfortable |
|
127
tok/s
76–203 · low confidence |
LLama 3..2 Typhoon 2 1B ≈ | 1B | Dec 2024 | 1.8 GB | 131k tokens ? | Q8_0 | Comfortable |
|
127
tok/s
76–203 · low confidence |
OLMo-1B ≈ | 1B | Feb 2024 | 1.8 GB | 131k tokens ? | Q8_0 | Comfortable |
|
127
tok/s
76–203 · low confidence |
Pythia-1b ≈ | 1B | Apr 2023 | 1.8 GB | 131k tokens ? | Q8_0 | Comfortable |
|
118
tok/s
71–188 · low confidence |
OpenELM-1.1B ≈ | 1.1B | May 2024 | 1.9 GB | 131k tokens ? | Q8_0 | Comfortable |
|
116
tok/s
69–185 · low confidence |
DeciCoder-1B ≈ | 1.1B | Aug 2023 | 1.9 GB | 131k tokens ? | Q8_0 | Comfortable |
|
116
tok/s
69–185 · low confidence |
SantaCoder ≈ | 1.1B | Jan 2023 | 1.9 GB | 131k tokens ? | Q8_0 | Comfortable |
|
116
tok/s
69–185 · low confidence |
TinyLlama-1.1B (1T token checkpoint) ≈ | 1.1B | Oct 2023 | 1.9 GB | 131k tokens ? | Q8_0 | Comfortable |
|
116
tok/s
69–185 · low confidence |
TinyLlama-1.1B (3T token checkpoint) ≈ | 1.1B | Oct 2023 | 1.9 GB | 131k tokens ? | Q8_0 | Comfortable |
|
106
tok/s
64–169 · low confidence |
EXAONE 4.0 (1.2B) ≈ | 1.2B | Jul 2025 | 2.0 GB | 131k tokens ? | Q8_0 | Comfortable |
|
106
tok/s
64–169 · low confidence |
LFM2-1.2B ≈ | 1.2B | Jul 2025 | 2.0 GB | 131k tokens ? | Q8_0 | Comfortable |
|
106
tok/s
64–169 · low confidence |
MinerU2.5 ≈ | 1.2B | Sep 2025 | 2.0 GB | 131k tokens ? | Q8_0 | Comfortable |
|
106
tok/s
64–169 · low confidence |
Pleias 1.0 1.2B ≈ | 1.2B | Dec 2024 | 2.0 GB | 131k tokens ? | Q8_0 | Comfortable |
|
106
tok/s
64–169 · low confidence |
Pleias-RAG-1B ≈ | 1.2B | Apr 2025 | 2.0 GB | 131k tokens ? | Q8_0 | Comfortable |
|
103
tok/s
88–124 |
Llama 3.2 1B | 1.2B | Sep 2024 | 2.2 GB | 131k tokens | Q8_0 | Comfortable |
|
102
tok/s
61–163 · low confidence |
MiniCPM-1.2B ≈ | 1.2B | Jun 2024 | 2.0 GB | 131k tokens ? | Q8_0 | Comfortable |
|
97.8
tok/s
59–156 · low confidence |
DeepSeek Coder 1.3B ≈ | 1.3B | Jan 2024 | 2.1 GB | 131k tokens ? | Q8_0 | Comfortable |
|
97.8
tok/s
59–156 · low confidence |
DeepSeek-VL-1.3B ≈ | 1.3B | Mar 2024 | 2.1 GB | 131k tokens ? | Q8_0 | Comfortable |
|
97.8
tok/s
59–156 · low confidence |
DigiRL ≈ | 1.3B | Jun 2024 | 2.1 GB | 131k tokens ? | Q8_0 | Comfortable |
|
97.8
tok/s
59–156 · low confidence |
GLA Transformer 1.3B ≈ | 1.3B | Aug 2024 | 2.1 GB | 131k tokens ? | Q8_0 | Comfortable |
|
97.8
tok/s
59–156 · low confidence |
Janus 1.3B ≈ | 1.3B | Oct 2024 | 2.1 GB | 131k tokens ? | Q8_0 | Comfortable |
|
97.8
tok/s
59–156 · low confidence |
Kosmos-2.5 ≈ | 1.3B | Aug 2024 | 2.1 GB | 131k tokens ? | Q8_0 | Comfortable |
|
97.8
tok/s
59–156 · low confidence |
Otter ≈ | 1.3B | May 2023 | 2.1 GB | 131k tokens ? | Q8_0 | Comfortable |
Speeds are estimates for a single request — one conversation at a time — calculated from memory bandwidth, model size and quantisation. Real throughput varies with the inference runtime and its version. Figures published by hardware vendors measure many simultaneous requests and are much higher.
On record
L4 full specification
Everything on record for this board, ordered by how much it bears on running a language model rather than by how a spec sheet would list it. Memory comes first because it decides the outcome; the rest is context.
Memory
The two specifications that decide what this card can run and how quickly. Capacity sets which models fit; bandwidth sets how many tokens per second they produce once they do.
- Memory size
- 24 GB
- Memory bandwidth
- 300 GB/s
- Memory type
- GDDR6
- Memory bus width
- 192 bit
- Memory clock
- 1.56 GHz
The chip
Which processor is on the board and how it was manufactured. A smaller process size generally means more performance for the same power.
- Graphics processor
- AD104
- Architecture
- Ada Lovelace
- Generation
- Server Ada(Lxx)
- Foundry
- TSMC
- Process size
- 5 nm
- Transistors
- 35.8 billion
- Transistor density
- 121,800 K/mm²
- Die size
- 294 mm²
- Released
- 21 March 2023
Clock speeds
How fast the processor runs. Worth far less here than on a gaming benchmark: generating text is limited by memory bandwidth, so a higher clock barely moves the result.
- Base clock
- 795 MHz
- Boost clock
- 2.04 GHz
Processing units
What the chip contains. These drive graphics performance and matter mainly for processing a long prompt rather than for producing the answer.
- Shading units
- 7,424
- Texture mapping units
- 240
- Render output units
- 80
- Streaming multiprocessors
- 60
- Tensor cores
- 240
- Ray tracing cores
- 60
- L1 cache
- 128 KB
- L2 cache
- 48 MB
Theoretical performance
Peak arithmetic rates published for the board. These are ceilings that no real workload reaches, and generating text reaches a small fraction of them because it is limited by memory rather than arithmetic.
- Half precision (FP16)
- 30.3 TFLOPS
- Single precision (FP32)
- 30.3 TFLOPS
- Double precision (FP64)
- 473.3 GFLOPS
- Pixel rate
- 163 GPixel/s
- Texture rate
- 490 GTexel/s
The board
What it takes to physically install and power the card — the practical constraints that decide whether it fits the machine you already own.
- Power draw (TDP)
- 72 W
- Suggested power supply
- 250 W
- Power connectors
- None
- Bus interface
- PCIe 4.0 x16
- Slot width
- Single-slot
- Dimensions
- 169 mm
Software support
Which graphics and compute interfaces the card supports. CUDA compute capability is the one that bears on inference: below 7.0 there are no tensor cores, and modern inference software falls back to slower code paths.
- CUDA compute capability
- 8.9
- DirectX
- 12.2
- OpenGL
- 4.6
- Vulkan
- 1.4
- OpenCL
- 3.0
- Shader model
- 6.8
Listings
Where to buy a L4
No vendor is currently listing this card. Listings come from vendors who publish them here directly — browse the vendor directory to see who is selling what.
What the numbers mean
Memory: the specification that decides everything
Memory
24 GB
Bandwidth
300 GB/s
Largest model
Mixtral 8x7B
L4 carries 24 GB of GDDR6. That covers the mid-sized models most people actually run. Once the runtime and driver reserve their working space, roughly this much is left: 21.6 GB.
Memory bandwidth reaches 300 GB/s across a bus of 192 bits. That is the number governing generation speed — arithmetic per byte read is small enough that the bus, not the cores, is what everything waits on.
Bandwidth is clock times bus width, and this card clocks its memory at 1.56 GHz. Widening the bus and raising the clock are the two levers a manufacturer has, which is why a card with unremarkable cores can still generate quickly.
The biggest thing it holds is Mixtral 8x7B, 46.7B, compressed to Q3_K_M and generating around 26.6 tokens per second.
The chip and how it was built
L4 is built on the graphics processor AD104, using the architecture Ada Lovelace from NVIDIA, as part of the generation Server Ada(Lxx).
The chip is manufactured by TSMC, on a process of 5 nm, with a die measuring 294 mm², holding 35.8 billion transistors. A smaller process generally means more performance for the same power, though for language models it matters far less than the memory subsystem.
It was released in March 2023, roughly 3.4840818262504 years ago. Inference software support tends to follow hardware by a year or two, so a card of this age generally has mature, well-optimised code paths available to it.
Compute throughput, and why it matters less than it looks
FP16
30.3 TFLOPS
FP64
473.3 GFLOPS
Tensor cores
240
On paper L4 reaches 30.3 TFLOPS at half precision, and 30.3 TFLOPS at single precision. These are peak figures no real workload sustains, and generating text reaches only a small fraction of them — decoding is limited by memory rather than arithmetic, which is why a card can look enormously powerful here and still produce tokens at an ordinary rate.
Double-precision throughput reaches 473.3 GFLOPS. It has no bearing on running a language model — no inference runtime uses it — but it separates datacentre parts from consumer ones, since the latter deliberately restrict it.
The card carries 240 tensor cores across 60 streaming multiprocessors. These accelerate the matrix arithmetic at the heart of a transformer, and they are what make prompt processing — reading a long document before answering — dramatically faster than it would otherwise be.
Clocks run from a base of 795 MHz to a boost of 2.04 GHz. Worth far less here than on a gaming benchmark: raising the clock speeds up the arithmetic, and the arithmetic is not what generation is waiting on.
Cache and processing units
L4 has an L1 cache of 128 KB, backed by an L2 cache of 48 MB. Cache absorbs a share of the memory traffic that would otherwise hit the main bus, which is the one place on this page where a number other than bandwidth quietly affects generation speed — a large L2 lets more of the working set stay close to the cores.
There are 7,424 shading units, 240 texture mapping units, and 80 render output units. These drive graphics workloads and contribute to prompt processing, but they sit idle for much of the time a model spends generating a reply.
Power, size and installation
Power draw
72 W
L4 is rated at 72 W, and the suggested system power supply is 250 W. Running a language model keeps a card busy in bursts rather than continuously — it draws hard while generating and idles between requests — so sustained draw over a working day is usually well below the rated figure.
The board occupies single-slot, measuring 169 mm long. Worth checking against the case and power supply already in the machine, since the largest cards need considerably more of both than a typical desktop provides.
It connects over PCIe 4.0 x16. The interface governs how quickly a model is loaded from disk into the card, not how fast it runs once there, so a narrower link costs a few seconds at startup and nothing thereafter.
The extremes
The largest AI models that run on a L4
The biggest open-weight models that fit on this card, newest first. Each is shown at the best compression the card can hold.
The fastest AI models on a L4
Where this card produces tokens quickest. Smaller models dominate here, because generating each token means reading the whole model out of memory once.
Step by step
How to work out the tokens per second of a L4
You do not have to calculate anything by hand — the gputps.com calculator on this page has already worked it out for every model this card can hold. Reading off the answer takes six steps.
-
01
Search for the model you want
The table lists 532 models this card runs. Search narrows the list by name or by size.
-
02
Decide how long your conversations run
Set the context to your real working length. Short questions cost almost nothing, but a long document can consume a large share of 24 GB it is often what pushes a large model over the edge.
-
03
Choose how far you will compress
By default the table picks the least-compressed copy that fits. Setting a floor removes models that only qualify through heavy compression.
-
04
Read the speed and the range
Each speed is an estimate for a single conversation, with a range beneath it. The top end here is 127 tok/s on Gemma 3 QAT 1B. The same card and model vary by thirty to fifty per cent between inference runtimes.
-
05
Check the memory column before committing
The fit column separates models that just fit from those with room to spare — worth checking before settling on one, against an available 24 GB.
-
06
Check the same model from the other side
Each model page repeats this calculation for the whole catalogue. Worth a look before deciding: it shows what else runs the same model, alongside L4.
Answers
L4 — common questions
L4— can it run a model that does not fit in its memory?
It can be split, with the overflow held in system memory beyond the card's 24 GB drags the whole thing down, and none of the figures on this page assume it.
Would two L4 cards be twice as fast?
Pairing them buys headroom rather than pace: 48 GB which lets you hold models neither could hold alone, but generation does not split that way. These figures describe one card.
L4— which AI models can it run?
532 of the 721 open-weight language models we track fit on this card and can be run locally. The table on this page lists every one, with the memory it needs, the quantisation it runs at and an estimated generation speed.
L4— what is the largest AI model it can run?
The largest model in our catalogue that fits is Mixtral 8x7B at 46.7B parameters, compressed to Q3_K_M. It generates roughly 26.6 tokens per second and needs about 21.0 GB of the card's memory.
L4— how many tokens per second does it produce?
It depends on the model. The fastest model we track here is Gemma 3 QAT 1B at about 127 tokens per second, while larger models run proportionally slower because each token requires reading the whole model out of memory once. Speeds are estimates for a single conversation at a time.
L4— can it run 7B models?
Yes. For example it runs Gemma 4 E4B at Q8_0, using about 9.8 GB of memory and generating around 28.2 tokens per second.
L4— can it run 13B models?
Yes. For example it runs DeepSeekMoE-16B at Q8_0, using about 17.5 GB of memory and generating around 44.1 tokens per second.
L4— can it run 30B models?
Yes. For example it runs Nemotron 3-Nano-30B-A3B at Q4_K_M, using about 18.1 GB of memory and generating around 51.6 tokens per second.
L4— how much memory does it have?
This card has 24 GB of GDDR6. Around a tenth is reserved by the inference runtime and the driver, leaving roughly 21.6 GB available for a model and its conversation.
L4— what is its memory bandwidth?
Memory bandwidth reaches 300 GB/s across a bus of 192 bits. This is the single best predictor of how fast it generates text, because producing each token means reading the entire model out of memory once.
L4— what type of memory does it use?
It uses GDDR6 clocked at 1.56 GHz. HBM types are found on datacentre accelerators and carry far more bandwidth than the GDDR used on desktop cards, which is why they generate tokens considerably faster at the same capacity.
L4— who makes it?
This is a product of NVIDIA, with the chip manufactured by TSMC, on a process of 5 nm.
L4— when was it released?
It was released in March 2023.
L4— how much power does it use?
Rated board power is 72 W, and the suggested system power supply is 250 W. Generating text draws hard in bursts and idles between requests, so average consumption over a working session is normally well below the rated figure.
L4— how much cache does it have?
The L1 cache is 128 KB, and the L2 cache is 48 MB. Cache absorbs part of the memory traffic that would otherwise reach the main bus, so a larger L2 gives a modest lift to generation speed beyond what bandwidth alone predicts.
L4— what are its TFLOPS?
It is rated at 30.3 TFLOPS at half precision and 30.3 TFLOPS at single precision. These are peak arithmetic ceilings rather than achievable rates, and text generation reaches only a small fraction of them because it is limited by memory bandwidth instead.
L4— how many tensor cores does it have?
It has 240 tensor cores across 60 streaming multiprocessors. They accelerate the matrix arithmetic a transformer is built from, which mainly speeds up processing a long prompt rather than producing the reply.
L4— does it support CUDA?
Yes. It reports CUDA compute capability 8.9. Capability 7.0 and above has tensor cores, which modern inference software uses; below that it falls back to slower code paths for quantised models.
L4— what bus interface does it use?
It uses PCIe 4.0 x16. This governs how fast a model is loaded onto the card rather than how fast it runs once loaded, so it costs a few seconds at startup and nothing during generation.
L4— is it good for running local AI models?
Its memory comfortably covers the mid-sized models most people run locally though its bandwidth means generation will feel slow on larger models. In total it runs 532 of the models we track. Whether that is enough depends entirely on which model you want — the table above answers that directly.
The other direction
Looking at it from the other side?
This page starts from the hardware. If you already know which model you want and need to know what it takes to run it, start from the model instead.