Calculate the TPS of the L40S 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
Largest model it holds
Qwen3-Coder-Next
80B · IQ4_XS · 62.4 tok/s
Fastest model
Gemma 3 QAT 1B
366 tok/s · 1B
What AI models can a L40S run?
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.
576 models match
Calculating| Quantisation | Fit | ||||||
|---|---|---|---|---|---|---|---|
|
366
tok/s
311–439 |
Gemma 3 1B | 1B | Mar 2025 | 1.8 GB | 33k tokens | Q8_0 | Comfortable |
|
366
tok/s
311–439 |
Gemma 3 QAT 1B | 1B | Apr 2025 | 1.8 GB | 33k tokens | Q8_0 | Comfortable |
|
366
tok/s
220–585 · low confidence |
HGRN 1B (WT 103) ≈ | 1B | Nov 2023 | 1.8 GB | 131k tokens ? | Q8_0 | Comfortable |
|
366
tok/s
220–585 · low confidence |
LLama 3..2 Typhoon 2 1B ≈ | 1B | Dec 2024 | 1.8 GB | 131k tokens ? | Q8_0 | Comfortable |
|
366
tok/s
220–585 · low confidence |
OLMo-1B ≈ | 1B | Feb 2024 | 1.8 GB | 131k tokens ? | Q8_0 | Comfortable |
|
366
tok/s
220–585 · low confidence |
Pythia-1b ≈ | 1B | Apr 2023 | 1.8 GB | 131k tokens ? | Q8_0 | Comfortable |
|
339
tok/s
203–542 · low confidence |
OpenELM-1.1B ≈ | 1.1B | May 2024 | 1.9 GB | 131k tokens ? | Q8_0 | Comfortable |
|
333
tok/s
200–532 · low confidence |
DeciCoder-1B ≈ | 1.1B | Aug 2023 | 1.9 GB | 131k tokens ? | Q8_0 | Comfortable |
|
333
tok/s
200–532 · low confidence |
SantaCoder ≈ | 1.1B | Jan 2023 | 1.9 GB | 131k tokens ? | Q8_0 | Comfortable |
|
333
tok/s
200–532 · low confidence |
TinyLlama-1.1B (1T token checkpoint) ≈ | 1.1B | Oct 2023 | 1.9 GB | 131k tokens ? | Q8_0 | Comfortable |
|
333
tok/s
200–532 · low confidence |
TinyLlama-1.1B (3T token checkpoint) ≈ | 1.1B | Oct 2023 | 1.9 GB | 131k tokens ? | Q8_0 | Comfortable |
|
305
tok/s
183–488 · low confidence |
EXAONE 4.0 (1.2B) ≈ | 1.2B | Jul 2025 | 2.0 GB | 131k tokens ? | Q8_0 | Comfortable |
|
305
tok/s
183–488 · low confidence |
MinerU2.5 ≈ | 1.2B | Sep 2025 | 2.0 GB | 131k tokens ? | Q8_0 | Comfortable |
|
305
tok/s
183–488 · low confidence |
Pleias 1.0 1.2B ≈ | 1.2B | Dec 2024 | 2.0 GB | 131k tokens ? | Q8_0 | Comfortable |
|
305
tok/s
183–488 · low confidence |
Pleias-RAG-1B ≈ | 1.2B | Apr 2025 | 2.0 GB | 131k tokens ? | Q8_0 | Comfortable |
|
298
tok/s
253–357 |
Llama 3.2 1B | 1.2B | Sep 2024 | 2.2 GB | 131k tokens | Q8_0 | Comfortable |
|
293
tok/s
176–469 · low confidence |
MiniCPM-1.2B ≈ | 1.2B | Jun 2024 | 2.0 GB | 131k tokens ? | Q8_0 | Comfortable |
|
281
tok/s
169–450 · low confidence |
DeepSeek Coder 1.3B ≈ | 1.3B | Jan 2024 | 2.1 GB | 131k tokens ? | Q8_0 | Comfortable |
|
281
tok/s
169–450 · low confidence |
DeepSeek-VL-1.3B ≈ | 1.3B | Mar 2024 | 2.1 GB | 131k tokens ? | Q8_0 | Comfortable |
|
281
tok/s
169–450 · low confidence |
DigiRL ≈ | 1.3B | Jun 2024 | 2.1 GB | 131k tokens ? | Q8_0 | Comfortable |
|
281
tok/s
169–450 · low confidence |
GLA Transformer 1.3B ≈ | 1.3B | Aug 2024 | 2.1 GB | 131k tokens ? | Q8_0 | Comfortable |
|
281
tok/s
169–450 · low confidence |
Janus 1.3B ≈ | 1.3B | Oct 2024 | 2.1 GB | 131k tokens ? | Q8_0 | Comfortable |
|
281
tok/s
169–450 · low confidence |
Kosmos-2.5 ≈ | 1.3B | Aug 2024 | 2.1 GB | 131k tokens ? | Q8_0 | Comfortable |
|
281
tok/s
169–450 · low confidence |
Otter ≈ | 1.3B | May 2023 | 2.1 GB | 131k tokens ? | Q8_0 | Comfortable |
|
281
tok/s
169–450 · low confidence |
Phi-1 ≈ | 1.3B | Oct 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
L40S 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
- 48 GB
- Memory bandwidth
- 864 GB/s
- Memory type
- GDDR6
- Memory bus width
- 384 bit
- Memory clock
- 2.25 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
- AD102
- Architecture
- Ada Lovelace
- Generation
- Server Ada(Lxx)
- Foundry
- TSMC
- Process size
- 5 nm
- Transistors
- 76.3 billion
- Transistor density
- 125,300 K/mm²
- Die size
- 609 mm²
- Released
- 13 October 2022
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
- 1.11 GHz
- Boost clock
- 2.52 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
- 18,176
- Texture mapping units
- 568
- Render output units
- 192
- Streaming multiprocessors
- 142
- Tensor cores
- 568
- Ray tracing cores
- 142
- 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)
- 91.6 TFLOPS
- Single precision (FP32)
- 91.6 TFLOPS
- Double precision (FP64)
- 1.4 TFLOPS
- Pixel rate
- 484 GPixel/s
- Texture rate
- 1,431 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)
- 300 W
- Suggested power supply
- 700 W
- Power connectors
- 1x 16-pin
- Bus interface
- PCIe 4.0 x16
- Slot width
- Dual-slot
- Dimensions
- 267 mm
- Display outputs
- 1x HDMI 2.1, 3x DisplayPort 1.4a
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 L40S
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
Why memory is the number that matters here
Memory
48 GB
Bandwidth
864 GB/s
Largest model
Qwen3-Coder-Next
The L40S carries 48 GB of GDDR6, which covers the mid-sized models most people actually run — about 43.2 GB of it after the runtime and driver reserve their working space.
Bandwidth is 864 GB/s across a 384-bit bus. Generating a token means reading every weight once, so that figure sets the pace more than any other number here, and at this level text arrives faster than most people read.
That comes from a 2.25 GHz memory clock across the bus width above. 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 practical ceiling is Qwen3-Coder-Next at 80B, held at IQ4_XS and running at roughly 62.4 tokens per second.
The chip and how it was built
The L40S is built on the AD102 graphics processor, using NVIDIA's Ada Lovelace architecture, as part of the Server Ada(Lxx) generation.
The chip is manufactured by TSMC, on a 5 nm process, with a die measuring 609 mm², holding 76.3 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 October 2022, roughly 3 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
91.6 TFLOPS
FP64
1.4 TFLOPS
Tensor cores
568
On paper the L40S reaches 91.6 TFLOPS at half precision and 91.6 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 is 1.4 TFLOPS. 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 568 tensor cores across 142 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 1.11 GHz at base to 2.52 GHz boosted. 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
The L40S has 128 KB of L1 cache, backed by 48 MB of L2. 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 18,176 shading units, 568 texture mapping units, and 192 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
300 W
The L40S is rated at 300 W, with a 700 W power supply suggested for the whole system. 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 a dual-slot, measuring 267 mm long, and needs 1x 16-pin. 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 a L40S can run
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 L40S
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 L40S
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 576 models this L40S can run. Search by name, or by size — typing 27b matches on the parameter count even when the name never states it.
-
02
Set the context length you will actually use
Set the context to your real working length. Short questions cost almost nothing; a long document can consume a large share of the card's 48 GB.
-
03
Pin the comparison to one quality level
Compression is what lets bigger models fit. The quality control drops any model that needs more of it than you are willing to give.
-
04
Take the range as the answer
Speeds come with error bars for a reason. The best case here is 366 tok/s on Gemma 3 QAT 1B, and which inference software you use moves that by thirty to fifty per cent.
-
05
Check the memory column before committing
Compare what each model needs with the 48 GB this card provides. Tight means it works today; comfortable means it still works when the conversation grows.
-
06
Check the same model from the other side
Every model name in the table links to its own page, which runs the same calculation across every card we hold. That is where you see whether the L40S is the right buy for it or merely a card that fits.
Answers
L40S — common questions
How much memory does a L40S have?
A L40S has 48 GB of GDDR6 memory. Around a tenth of that is reserved by the inference runtime and the driver, leaving roughly 43.2 GB available for a model and its conversation.
What is the memory bandwidth of a L40S?
The L40S has 864 GB/s of memory bandwidth, across a 384-bit memory bus. 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.
What type of memory does a L40S use?
It uses GDDR6 clocked at 2.25 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.
Who makes the L40S?
The L40S is a NVIDIA product, with the chip manufactured by TSMC, on a 5 nm process.
When was the L40S released?
The L40S was released in October 2022.
How much power does a L40S use?
The L40S has a rated board power of 300 W, and a 700 W system power supply is suggested. Generating text draws hard in bursts and idles between requests, so average consumption over a working session is normally well below the rated figure.
How much cache does a L40S have?
The L40S has 128 KB of L1 cache, and 48 MB of L2 cache. 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.
What are the TFLOPS of a L40S?
The L40S is rated at 91.6 TFLOPS at half precision and 91.6 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.
How many tensor cores does a L40S have?
The L40S has 568 tensor cores across 142 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.
Does the L40S support CUDA?
Yes. The L40S 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.
What bus interface does the L40S 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.
Is the L40S good for running local AI models?
Its memory is large enough for models most desktop hardware cannot touch and its bandwidth is high enough to generate text faster than most people read. In total it runs 576 of the models we track. Whether that is enough depends entirely on which model you want — the table above answers that directly.
Can a L40S run a model that does not fit in its memory?
Only partly. Layers beyond the 48 GB sit in system memory and run at a fraction of the speed, so a mostly-offloaded model is rarely worth using. Every figure here assumes it is fully resident on the card.
Would two L40S cards be twice as fast?
Pairing L40S cards buys headroom rather than pace: 96 GB of combined memory, at roughly the same generation speed as one.
What AI models can a L40S run?
576 of the 679 open-weight language models we track fit on a L40S and can be run locally on it. The table on this page lists every one, with the memory it needs, the quantisation it runs at and an estimated generation speed.
What is the largest AI model a L40S can run?
The largest model in our catalogue that fits on a L40S is Qwen3-Coder-Next at 80B parameters, compressed to IQ4_XS. It generates roughly 62.4 tokens per second and needs about 38.8 GB of the card's memory.
How many tokens per second does a L40S produce?
It depends on the model. On a L40S the fastest model we track is Gemma 3 QAT 1B at about 366 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.
Can a L40S run a 7B model?
Yes. For example a L40S runs Multi-Token Prediction 7B at Q8_0, using about 7.9 GB of memory and generating around 54.6 tokens per second.
Can a L40S run a 13B model?
Yes. For example a L40S runs DeepSeekMoE-16B at Q8_0, using about 17.5 GB of memory and generating around 127 tokens per second.
Can a L40S run a 30B model?
Yes. For example a L40S runs ERNIE-4.5-VL-28B-A3B at Q8_0, using about 29.2 GB of memory and generating around 72.6 tokens per second.
Can a L40S run a 70B model?
Yes. For example a L40S runs Qwen3-Coder-Next at IQ4_XS, using about 38.8 GB of memory and generating around 62.4 tokens per second.
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.