Unified-IO (XL) TPS calculator

Open weights Allen Institute for AI,University of Washington 2.9B parameters June 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

818 cards that can run it

818 cards we hold specifications for

Smallest card that fits

Tesla C1080

4 GB · Q6_K · 18.3 tok/s

Fastest card

B200

1,158 tok/s · 180 GB

Which GPUs can run Unified-IO (XL)?

Set the inputs, read the answer

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

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

818 cards match

Calculating
Needs Quantisation Fit
1,158 tok/s

695–1,853 · low confidence

B200 NVIDIA 180 GB 8,000 GB/s Jan 2024 3.8 GB Q8_0 Comfortable
1,158 tok/s

695–1,853 · low confidence

B300 NVIDIA 288 GB 8,000 GB/s Sep 2025 3.8 GB Q8_0 Comfortable
925 tok/s

555–1,480 · low confidence

Radeon Instinct MI350X AMD 288 GB 8,190 GB/s Jan 2025 3.8 GB Q8_0 Comfortable
925 tok/s

555–1,480 · low confidence

Radeon Instinct MI355X AMD 288 GB 8,190 GB/s Jan 2025 3.8 GB Q8_0 Comfortable
740 tok/s

444–1,184 · low confidence

Radeon Instinct MI300 AMD 128 GB 6,550 GB/s Jan 2023 3.8 GB Q8_0 Comfortable
708 tok/s

425–1,133 · low confidence

H200 NVL NVIDIA 141 GB 4,890 GB/s Nov 2024 3.8 GB Q8_0 Comfortable
708 tok/s

425–1,133 · low confidence

H200 SXM 141 GB NVIDIA 141 GB 4,890 GB/s Nov 2024 3.8 GB Q8_0 Comfortable
678 tok/s

407–1,084 · low confidence

Radeon Instinct MI325X AMD 256 GB 6,000 GB/s Oct 2024 3.8 GB Q8_0 Comfortable
601 tok/s

361–962 · low confidence

Radeon Instinct MI300A AMD 128 GB 5,325 GB/s Dec 2023 3.8 GB Q8_0 Comfortable
601 tok/s

361–962 · low confidence

Radeon Instinct MI300X AMD 192 GB 5,325 GB/s Dec 2023 3.8 GB Q8_0 Comfortable
601 tok/s

361–962 · low confidence

Radeon Instinct MI308X AMD 192 GB 5,325 GB/s Dec 2023 3.8 GB Q8_0 Comfortable
571 tok/s

342–913 · low confidence

H100 NVL 94 GB NVIDIA 94 GB 3,940 GB/s Mar 2023 3.8 GB Q8_0 Comfortable
487 tok/s

292–778 · low confidence

H100 PCIe 96 GB NVIDIA 96 GB 3,360 GB/s Mar 2023 3.8 GB Q8_0 Comfortable
487 tok/s

292–778 · low confidence

H100 SXM5 80 GB NVIDIA 80 GB 3,360 GB/s Oct 2022 3.8 GB Q8_0 Comfortable
487 tok/s

292–778 · low confidence

H100 SXM5 94 GB NVIDIA 94 GB 3,360 GB/s Mar 2023 3.8 GB Q8_0 Comfortable
487 tok/s

292–778 · low confidence

H100 SXM5 96 GB NVIDIA 96 GB 3,360 GB/s Mar 2023 3.8 GB Q8_0 Comfortable
487 tok/s

292–778 · low confidence

H800 SXM5 NVIDIA 80 GB 3,360 GB/s Mar 2023 3.8 GB Q8_0 Comfortable
370 tok/s

222–593 · low confidence

Radeon Instinct MI250 AMD 128 GB 3,280 GB/s Nov 2021 3.8 GB Q8_0 Comfortable
370 tok/s

222–593 · low confidence

Radeon Instinct MI250X AMD 128 GB 3,280 GB/s Nov 2021 3.8 GB Q8_0 Comfortable
309 tok/s

185–494 · low confidence

Data Center GPU Max 1550 Intel 128 GB 3,280 GB/s Jan 2023 3.8 GB Q8_0 Comfortable
302 tok/s

181–483 · low confidence

Data Center GPU Max Subsystem Intel 128 GB 3,210 GB/s Jan 2023 3.8 GB Q8_0 Comfortable
295 tok/s

177–473 · low confidence

A100 SXM4 80 GB NVIDIA 80 GB 2,040 GB/s Nov 2020 3.8 GB Q8_0 Comfortable
295 tok/s

177–473 · low confidence

A100X NVIDIA 80 GB 2,040 GB/s Jun 2021 3.8 GB Q8_0 Comfortable
295 tok/s

177–473 · low confidence

A800 SXM4 80 GB NVIDIA 80 GB 2,040 GB/s Aug 2022 3.8 GB Q8_0 Comfortable
295 tok/s

177–473 · low confidence

H100 CNX NVIDIA 80 GB 2,040 GB/s Mar 2023 3.8 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
Allen Institute for AI,University of Washington
Organisation type
Research collective,Academia
Country
United States of America
Published
17 June 2022
Authors
Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, Aniruddha Kembhavi

What it does

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

Domain
Multimodal, Vision, Language
Task
Object detection, Language modeling/generation, Image generation, Visual question answering, Image classification, Image captioning, Text classification, Text summarization, Question answering

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
2.9B

2952M, Table 2

Training data
74,880,000,000 tokens
Epochs
7.88

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
3.5 × 10²¹ FLOP

1M steps, batch size 1024. Sequence length may be 128-256: "We use a maximum of 256 and 128 text tokens for inputs and outputs respectively, and a maximum length of 576 (i.e. 24 × 24 patch encoding from a 384 × 384 image) for image inputs and 256 (i.e. 16 × 16 latent codes from a 256 × 256 image) for image outputs 6 * 1 million * 1024 * 128 * 2.9 billion = 2.3e21 6 * 1 million * 1024 * 256 * 2.9 billion = 4.6e21 average is 3.5e21 No hardware details.

How it was established
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

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

Apache 2.0, looks like model and inference code: https://github.com/allenai/unified-io-inference

How it is classified

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

Record confidence
Speculative
Citations
513

Sources

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

Reference
Unified-IO: A Unified Model for Vision, Language, and Multi-Modal Tasks
Last updated
25 May 2026

The extremes

The ten fastest GPUs that run Unified-IO (XL)

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.

  1. 01 B300 288 GB · 8,000 GB/s · Q8_0 1,158 tok/s
  2. 02 B200 180 GB · 8,000 GB/s · Q8_0 1,158 tok/s
  3. 03 Radeon Instinct MI350X 288 GB · 8,190 GB/s · Q8_0 925 tok/s
  4. 04 Radeon Instinct MI355X 288 GB · 8,190 GB/s · Q8_0 925 tok/s
  5. 05 Radeon Instinct MI300 128 GB · 6,550 GB/s · Q8_0 740 tok/s
  6. 06 H200 NVL 141 GB · 4,890 GB/s · Q8_0 708 tok/s
  7. 07 H200 SXM 141 GB 141 GB · 4,890 GB/s · Q8_0 708 tok/s
  8. 08 Radeon Instinct MI325X 256 GB · 6,000 GB/s · Q8_0 678 tok/s
  9. 09 Radeon Instinct MI300A 128 GB · 5,325 GB/s · Q8_0 601 tok/s
  10. 10 Radeon Instinct MI300X 192 GB · 5,325 GB/s · Q8_0 601 tok/s

The smallest GPUs that still run Unified-IO (XL)

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

  1. 01 GeForce RTX 4010 4 GB · needs 3.1 GB · Q6_K · tight 20.2 tok/s
  2. 02 RTX A400 4 GB · needs 3.1 GB · Q6_K · tight 20.2 tok/s
  3. 03 RTX 500 Mobile Ada Generation 4 GB · needs 3.1 GB · Q6_K · tight 26.9 tok/s
  4. 04 GeForce RTX 3050 A Mobile 4 GB · needs 3.1 GB · Q6_K · tight 40.4 tok/s
  5. 05 Jetson Orin Nano 4 GB 4 GB · needs 3.1 GB · Q6_K · tight 7.2 tok/s
  6. 06 Radeon RX 6450M 4 GB · needs 3.1 GB · Q6_K · tight 21.0 tok/s
  7. 07 Radeon RX 6550M 4 GB · needs 3.1 GB · Q6_K · tight 23.6 tok/s
  8. 08 Radeon RX 6550S 4 GB · needs 3.1 GB · Q6_K · tight 21.0 tok/s
  9. 09 Arc A310 4 GB · needs 3.1 GB · Q6_K · tight 17.0 tok/s
  10. 10 Arc Pro A30M 4 GB · needs 3.1 GB · Q6_K · tight 17.5 tok/s

What the numbers mean

Hardware requirements in practice

Minimum card

Tesla C1080

Memory needed

3.1 GB

Fastest

1,158 tok/s

Unified-IO (XL) is small enough at 2.9B parameters that hardware is rarely the obstacle — 818 of the cards we track can run it, including cards several years old.

The entry point is the Tesla C1080: 4 GB of memory, Q6_K compression, roughly 18.3 tokens per second.

Top of the range is the B200, at roughly 1,158 tokens per second thanks to 8,000 GB/s of bandwidth.

Background

Unified-IO (XL) was published by Allen Institute for AI,University of Washington, in United States of America, in June 2022. It comes out of research collective,Academia.

It works in Multimodal, Vision, Language, and is recorded as doing object detection, Language modeling/generation, Image generation, Visual question answering, Image classification, Image captioning, Text classification, Text summarization, Question answering.

The weights being open is what puts this page in the calculator rather than only in the catalogue: it is a model you can actually hold.

Reading the throughput figures

Half the cards that hold it manage more than 36.8 tokens per second, and 783 exceed reading speed outright.

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.

Without the attention layout on record, the memory column is an approximation. It is close enough to choose hardware by, and least reliable at long context.

What went into building it

Producing it required around 3.5 × 10²¹ FLOP of arithmetic, on Google TPU v4, which is a statement about the training budget rather than about inference.

It was trained on about 74,880,000,000 tokens of text.

Step by step

How to choose a GPU for Unified-IO (XL)

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

  1. 01

    Check what it needs before anything else

    The table lists every card that can hold Unified-IO (XL) — around 3.1 GB at Q6_K. That figure, not the card's headline performance, is what decides whether it runs.

  2. 02

    Decide how long your conversations run

    Set the context to what you will actually use. The cache grows with the conversation, and it is the usual reason Unified-IO (XL) stops fitting a card that seemed fine.

  3. 03

    Set a quality floor

    The quantisation column varies by card, because a bigger card holds a more accurate copy of Unified-IO (XL) — Q6_K on the smallest card that fits. Set a floor to hold the comparison at one level.

  4. 04

    Rank by throughput rather than spec sheet

    Ranking by tokens per second for Unified-IO (XL) follows memory bandwidth, not core counts, which is why the B200 tops it at 1,158 tok/s.

  5. 05

    Look at the headroom, not just the fit

    The fit column separates cards that just manage Unified-IO (XL) from those with room to spare. Buy for the second if the context might grow.

  6. 06

    Open the card you have settled on

    Every card name links to its own page, which runs the same calculation across the whole model catalogue. Worth a look before buying for Unified-IO (XL) alone — a card is usually bought for more than one model.

Answers

Unified-IO (XL) — common questions

01

Can I run Unified-IO (XL) on a 24 GB GPU?

Yes. A GeForce RTX 5090 D V2 with 24 GB runs it at Q8_0, using about 3.8 GB and generating roughly 194 tokens per second — a comfortable fit.

02

Is Unified-IO (XL) open source?

Its weights are published, so Unified-IO (XL) 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.

03

How many parameters does Unified-IO (XL) have?

Unified-IO (XL) has 2.9B parameters. 2952M, Table 2. 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.

04

Who created Unified-IO (XL)?

Unified-IO (XL) was published by Allen Institute for AI,University of Washington, based in United States of America, categorised as research collective,Academia.

05

When was Unified-IO (XL) released?

Unified-IO (XL) was published in June 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.

06

What is Unified-IO (XL) used for?

Unified-IO (XL) works in Multimodal, Vision, Language, and is recorded as handling object detection, Language modeling/generation, Image generation, Visual question answering, Image classification, Image captioning, Text classification, Text summarization, Question answering. These are the areas it was designed around; they describe intent rather than a hard boundary.

07

Where can I download Unified-IO (XL)?

The weights for Unified-IO (XL) are published, though we do not hold a repository link for it. This site calculates hardware requirements rather than hosting model files.

08

How much compute was used to train Unified-IO (XL)?

Around 3.5 × 10²¹ FLOP, on 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.

09

Can I run Unified-IO (XL) 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 Unified-IO (XL) is rarely worth using. Every figure here assumes the whole model is on the card.

10

Would two GPUs run Unified-IO (XL) faster?

A second card roughly doubles the memory available but not the generation rate. With 818 cards already able to run Unified-IO (XL) alone, the case for pairing is weak.

11

Why does the quantisation differ between cards for Unified-IO (XL)?

A larger card holds a more accurate copy. Across the cards that run Unified-IO (XL), 2 compression levels are used; the floor control above pins it to one.

12

How accurate are these Unified-IO (XL) speed estimates?

They are calculated from specifications rather than measured, and each carries a range — 695–1,853 tok/s on the B200, for instance. The same model and card vary by thirty to fifty per cent depending on the inference software and its version.

13

What GPU do I need to run Unified-IO (XL)?

The smallest card in our catalogue that holds Unified-IO (XL) is the Tesla C1080, with 4 GB of memory. It runs the model at Q6_K using about 3.1 GB, and produces roughly 18.3 tokens per second. 818 cards in total can run it.

14

How fast is Unified-IO (XL) on a GPU?

It depends on the card. The quickest we calculate is a B200 at about 1,158 tokens per second; the slowest that still runs it manages considerably less. Reading speed is around ten tokens per second, and 783 of the cards that can run Unified-IO (XL) clear that.

15

How much VRAM does Unified-IO (XL) need?

About 3.1 GB at Q6_K compression, 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.

16

Can I run Unified-IO (XL) on a 8 GB GPU?

Yes. A CMP 170HX 8 GB with 8 GB runs it at Q8_0, using about 3.8 GB and generating roughly 216 tokens per second — a comfortable fit.

17

Can I run Unified-IO (XL) on a 12 GB GPU?

Yes. A GeForce RTX 3080 Ti with 12 GB runs it at Q8_0, using about 3.8 GB and generating roughly 132 tokens per second — a comfortable fit.

18

Can I run Unified-IO (XL) on a 16 GB GPU?

Yes. A Tesla V100 SXM2 16 GB with 16 GB runs it at Q8_0, using about 3.8 GB and generating roughly 164 tokens per second — a comfortable fit.

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