UL2 TPS calculator

Open weights Google Research,Google Brain 20B parameters May 2022

Each card below is assessed against this model at the context length and minimum quality you choose. Speed is an estimate for a single request, calculated from the card's memory bandwidth and the size of the model once compressed.

Calculated for this model

293 cards that can run it

818 cards we hold specifications for

Smallest card that fits

Quadro K6000

12 GB · Q3_K_M · 14.0 tok/s

Fastest card

B200

169 tok/s · 180 GB

Which GPUs can run UL2?

Set the inputs, read the answer

Context length tokens

A longer conversation needs more memory, which can push this model off smaller cards.

Minimum quality

Hides cards that would only fit the model by compressing it below this point.

Search cards

Card memory

Also show cards that cannot run it

293 cards match

Calculating

					Needs	Quantisation	Fit
169 tok/s 102–271 · low confidence	B200 NVIDIA	180 GB	8,000 GB/s	Jan 2024	22.1 GB	Q8_0	Comfortable
169 tok/s 102–271 · low confidence	B300 NVIDIA	288 GB	8,000 GB/s	Sep 2025	22.1 GB	Q8_0	Comfortable
135 tok/s 81–216 · low confidence	Radeon Instinct MI350X AMD	288 GB	8,190 GB/s	Jan 2025	22.1 GB	Q8_0	Comfortable
135 tok/s 81–216 · low confidence	Radeon Instinct MI355X AMD	288 GB	8,190 GB/s	Jan 2025	22.1 GB	Q8_0	Comfortable
108 tok/s 65–173 · low confidence	Radeon Instinct MI300 AMD	128 GB	6,550 GB/s	Jan 2023	22.1 GB	Q8_0	Comfortable
104 tok/s 62–166 · low confidence	H200 NVL NVIDIA	141 GB	4,890 GB/s	Nov 2024	22.1 GB	Q8_0	Comfortable
104 tok/s 62–166 · low confidence	H200 SXM 141 GB NVIDIA	141 GB	4,890 GB/s	Nov 2024	22.1 GB	Q8_0	Comfortable
99.1 tok/s 59–159 · low confidence	Radeon Instinct MI325X AMD	256 GB	6,000 GB/s	Oct 2024	22.1 GB	Q8_0	Comfortable
88.0 tok/s 53–141 · low confidence	Radeon Instinct MI300A AMD	128 GB	5,325 GB/s	Dec 2023	22.1 GB	Q8_0	Comfortable
88.0 tok/s 53–141 · low confidence	Radeon Instinct MI300X AMD	192 GB	5,325 GB/s	Dec 2023	22.1 GB	Q8_0	Comfortable
88.0 tok/s 53–141 · low confidence	Radeon Instinct MI308X AMD	192 GB	5,325 GB/s	Dec 2023	22.1 GB	Q8_0	Comfortable
83.4 tok/s 50–134 · low confidence	H100 NVL 94 GB NVIDIA	94 GB	3,940 GB/s	Mar 2023	22.1 GB	Q8_0	Comfortable
71.2 tok/s 43–114 · low confidence	H100 PCIe 96 GB NVIDIA	96 GB	3,360 GB/s	Mar 2023	22.1 GB	Q8_0	Comfortable
71.2 tok/s 43–114 · low confidence	H100 SXM5 80 GB NVIDIA	80 GB	3,360 GB/s	Oct 2022	22.1 GB	Q8_0	Comfortable
71.2 tok/s 43–114 · low confidence	H100 SXM5 94 GB NVIDIA	94 GB	3,360 GB/s	Mar 2023	22.1 GB	Q8_0	Comfortable
71.2 tok/s 43–114 · low confidence	H100 SXM5 96 GB NVIDIA	96 GB	3,360 GB/s	Mar 2023	22.1 GB	Q8_0	Comfortable
71.2 tok/s 43–114 · low confidence	H800 SXM5 NVIDIA	80 GB	3,360 GB/s	Mar 2023	22.1 GB	Q8_0	Comfortable
55.2 tok/s 33–88 · low confidence	Tesla V100 SXM2 16 GB NVIDIA	16 GB	1,130 GB/s	Nov 2019	12.8 GB	Q4_K_M	Tight
54.2 tok/s 33–87 · low confidence	Radeon Instinct MI250 AMD	128 GB	3,280 GB/s	Nov 2021	22.1 GB	Q8_0	Comfortable
54.2 tok/s 33–87 · low confidence	Radeon Instinct MI250X AMD	128 GB	3,280 GB/s	Nov 2021	22.1 GB	Q8_0	Comfortable
52.1 tok/s 31–83 · low confidence	GeForce RTX 3080 12 GB NVIDIA	12 GB	912 GB/s	Jan 2022	10.5 GB	Q3_K_M	Tight
52.1 tok/s 31–83 · low confidence	GeForce RTX 3080 Ti NVIDIA	12 GB	912 GB/s	May 2021	10.5 GB	Q3_K_M	Tight
46.9 tok/s 28–75 · low confidence	GeForce RTX 5080 NVIDIA	16 GB	960 GB/s	Jan 2025	12.8 GB	Q4_K_M	Tight
45.2 tok/s 27–72 · low confidence	Data Center GPU Max 1550 Intel	128 GB	3,280 GB/s	Jan 2023	22.1 GB	Q8_0	Comfortable
44.2 tok/s 27–71 · low confidence	Data Center GPU Max Subsystem Intel	128 GB	3,210 GB/s	Jan 2023	22.1 GB	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

Full specification

Everything on record for this model. Most of it describes how it was trained rather than how it runs — useful context for judging how much work went into it, and how it compares with models built at a different scale.

Origin

Who built this model, where, and when it was published.

Organisation: Google Research,Google Brain
Organisation type: Industry,Industry
Country: United States of America
Published: 10 May 2022
Authors: Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Neil Houlsby, Donald Metzler

What it does

The problem areas the model was built for. A model can carry several of each.

Domain: Language
Task: Language modeling/generation, Question answering, Text summarization
Numerical format: BF16

Size

How large the model is and how much data it was trained on. Parameters are the figure that decides whether it fits on a given graphics card.

Parameters: 20B
Training data: 1,000,000,000,000 tokens
Batch size: 524,288

Training compute

The arithmetic performed to train the model, measured in floating-point operations. It is a measure of what the training run cost, not of how fast the finished model answers you.

Training compute: 1.2 × 10²³ FLOP
How it was established: Hardware,Operation counting

The training run

What it physically took to train: which chips, how many, for how long, and what that drew from the wall.

Training hardware: Google TPU v4
Chips used: 512
Chip-hours: 380,928
Wall-clock time: 744 hours (31 days)
Hardware utilisation: MFU 29.9%
Power draw: 349.6 kW
Compute cost: $126,786

Availability

Whether you can obtain the model and run it on your own hardware, which is what decides if any of the graphics-card figures on this page apply.

Weights: Open — downloadable
Model access: Open weights (unrestricted)
Training code: Unreleased
Hugging Face: google

How it is classified

Labels the source dataset applies when tracking notable models, and how confident it is in the entry.

Foundation model: Yes
Likely above 10²³ FLOP: Yes
Why it is tracked: SOTA improvement
Record confidence: Confident
Citations: 387

Sources

Where this record came from and when it was last checked.

Reference: Unifying Language Learning Paradigms
Last updated: 25 May 2026

The extremes

The ten fastest GPUs that run UL2

Ranked by estimated tokens per second, newest card first where speeds tie. Because generation is bound by memory bandwidth, this ordering follows bandwidth rather than any gaming benchmark.

The smallest GPUs that still run UL2

The cheapest route in, by memory capacity. A tight fit runs the model but leaves nothing spare for a longer conversation.

What the numbers mean

What you need to run it

Minimum card

Quadro K6000

Memory needed

10.5 GB

Fastest

169 tok/s

UL2 reaches a parameter count of 20B. That lands in the range a serious desktop card can handle once the weights are compressed. The number of cards we track that can run it: 293.

The least hardware that works is Quadro K6000, with a memory capacity of 12 GB, running it at a compression of Q3_K_M and producing around 14.0 tokens per second.

Top of the range is B200, generating roughly 169 tokens per second on the strength of a memory bandwidth of 8,000 GB/s.

What this model is

UL2 was published by Google Research,Google Brain, in the country recorded as United States of America, during May 2022. It comes out of an organisation categorised as industry,Industry.

It works in the domain of Language, and is recorded as performing the task of language modeling/generation, Question answering, Text summarization.

Because its weights were released, nothing about running it depends on a provider staying available — it is yours once downloaded. On Hugging Face it is published under the organisation google.

What decides the speed

Half the cards that hold it manage more than 20.6 tokens per second. Exceeding reading speed outright: 248 of them.

Being dense, it reads all of itself per token, which is why the ordering by speed below follows the ordering by memory bandwidth so closely.

Its internal architecture is not on file, so memory is approximated from the parameter count and marked accordingly. Expect the real figure to differ, more so at long context.

Training and provenance

Training it took a computation budget of roughly 1.2 × 10²³ FLOP, on hardware recorded as Google TPU v4. That figure measures what producing the model cost, and has no bearing on how fast it answers.

Training consumed a corpus of around 1,000,000,000,000 tokens of text.

It is tracked in the underlying dataset for one reason in particular: sOTA improvement.

Step by step

How to choose a GPU for UL2

The table above has already assessed every card we hold specifications for against this model. Getting to your answer takes six steps.

01

Read the memory figure first

Every card here has been checked against UL2, needing around 10.5 GB at a compression of Q3_K_M. No amount of processing power compensates for a card that cannot hold it.
02

Set the context length you will work at

Set the context to what you will actually use. The cache grows with the conversation, and it is the usual reason a card that seemed fine stops fitting UL2.
03

Decide how much compression you will accept

The quantisation column varies by card, because a bigger card holds a more accurate copy, reaching a compression of Q3_K_M on the smallest card that fits. Setting a minimum quality drops the cards that only manage it by squeezing further than you would want, and holds the comparison at one level.
04

Rank by throughput rather than spec sheet

The speed ordering is effectively an ordering by memory bandwidth, for UL2. It will not match a gaming ordering, because generation is bound by memory bandwidth. The card topping the list is B200, at 169 tok/s.
05

Look at the headroom, not just the fit

A tight fit runs, but leaves nothing spare for a longer conversation, in the case of UL2. Comfortable means you can grow the context later. That difference matters more than a few tokens per second, so buy for comfortable if you expect to.
06

Check the card from the other side

Each card page repeats this sweep for every model we hold, answering what else the hardware is good for beyond UL2.

Answers

UL2 — common questions

UL2— can I run it on a GPU holding 12 GB?

Yes. The card GeForce RTX 3080 Ti, holding 12 GB, runs it at a compression of Q3_K_M, using about 10.5 GB and generating roughly 52.1 tokens per second. The fit is tight.

UL2— can I run it on a GPU holding 16 GB?

Yes. The card Tesla V100 SXM2 16 GB, holding 16 GB, runs it at a compression of Q4_K_M, using about 12.8 GB and generating roughly 55.2 tokens per second. The fit is tight.

UL2— can I run it on a GPU holding 24 GB?

Yes. The card GeForce RTX 5090 D V2, holding 24 GB, runs it at a compression of Q6_K, using about 17.5 GB and generating roughly 41.2 tokens per second. The fit is comfortable.

UL2— is it open source?

Its weights are published, so it can be downloaded and run on your own hardware. Note that open weights is not the same as open source in the full sense — it says nothing about the training data, the training code, or the commercial terms attached.

UL2— how many parameters does it have?

It has a parameter count of 20B. Taken from Directory of LLMs. That figure is the total, and it is what decides how much memory the model needs — roughly half a gigabyte per billion at the compression most people use.

UL2— who created it?

It was published by Google Research,Google Brain, based in United States of America, an organisation categorised as industry,Industry.

UL2— when was it released?

It was published in May 2022. Capability per parameter has improved considerably since, so a newer model of the same size is often the better use of the same hardware.

UL2— what is it used for?

It works in the domain of Language, and is recorded as handling the task of language modeling/generation, Question answering, Text summarization. These are the areas it was designed around; they describe intent rather than a hard boundary.

UL2— where can I download it?

Its weights are published on Hugging Face, under the organisation google. We do not host model files — this site calculates what hardware is needed to run them.

UL2— how much compute was used to train it?

Training consumed around 1.2 × 10²³ FLOP, on hardware recorded as Google TPU v4. That measures what producing the model cost and says nothing about how quickly it answers once trained — inference speed comes from memory bandwidth, not from the training budget.

UL2— can I run it if it does not fit in my GPU?

Partly. Layers that do not fit sit in system memory and run at a fraction of the speed, so a mostly-offloaded model is rarely worth using. The nearest miss we calculate falls short by 2.9 GB. Every figure here assumes the whole model is resident on the card.

UL2— would two GPUs run it faster?

A second card roughly doubles the memory available but not the generation rate. The number already able to run it alone: 293. So a second card is rarely the answer here.

UL2— why does the quantisation differ between cards?

Each card is shown running the least-compressed copy it can hold. The number of distinct compression levels across the cards that fit it: 4. Bigger cards get the more accurate version, and the quality floor above pins the comparison to one level.

UL2— how accurate are these speed estimates?

Every figure is derived from memory bandwidth and model size, not benchmarked, which is why each is published as a range rather than a single number. One example: 102–271 tok/s on B200. The same model and card vary by thirty to fifty per cent depending on the inference software and its version.

UL2— what GPU do I need to run it?

The smallest card in our catalogue that holds it is Quadro K6000, with a memory capacity of 12 GB. It runs the model at a compression of Q3_K_M using about 10.5 GB, and produces roughly 14.0 tokens per second. The number of cards able to run it in total: 293.

UL2— how fast is it on a GPU?

It depends on the card. The quickest we calculate is B200, at about 169 tokens per second; the slowest that still runs it manages considerably less. Reading speed is around ten tokens per second, and the number of cards clearing that: 248.

UL2— how much VRAM does it need?

It needs about 10.5 GB at a compression of Q3_K_M, which is what the smallest card that runs it uses. Less compression needs more: the figures in the memory column above are recalculated for each card, because each one holds the least-compressed version it can.

Source

Original publication

Record last updated 25 May 2026

The other direction

Looking at it from the other side?

This page starts from the model. If you already own a card and want to know everything it will run, start from the hardware instead.

All AI models