ViT-G (model soup) TPS calculator

Open weights University of Washington,Columbia University,Google,Meta AI,Tel Aviv University 1.8B parameters March 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 · Q8_0 · 20.0 tok/s

Fastest card

B200

1,838 tok/s · 180 GB

Which GPUs can run ViT-G (model soup)?

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

818 cards match

Calculating

					Needs	Quantisation	Fit
1,838 tok/s 1,103–2,942 · low confidence	B200 NVIDIA	180 GB	8,000 GB/s	Jan 2024	2.7 GB	Q8_0	Comfortable
1,838 tok/s 1,103–2,942 · low confidence	B300 NVIDIA	288 GB	8,000 GB/s	Sep 2025	2.7 GB	Q8_0	Comfortable
1,468 tok/s 881–2,349 · low confidence	Radeon Instinct MI350X AMD	288 GB	8,190 GB/s	Jan 2025	2.7 GB	Q8_0	Comfortable
1,468 tok/s 881–2,349 · low confidence	Radeon Instinct MI355X AMD	288 GB	8,190 GB/s	Jan 2025	2.7 GB	Q8_0	Comfortable
1,174 tok/s 704–1,879 · low confidence	Radeon Instinct MI300 AMD	128 GB	6,550 GB/s	Jan 2023	2.7 GB	Q8_0	Comfortable
1,124 tok/s 674–1,798 · low confidence	H200 NVL NVIDIA	141 GB	4,890 GB/s	Nov 2024	2.7 GB	Q8_0	Comfortable
1,124 tok/s 674–1,798 · low confidence	H200 SXM 141 GB NVIDIA	141 GB	4,890 GB/s	Nov 2024	2.7 GB	Q8_0	Comfortable
1,075 tok/s 645–1,721 · low confidence	Radeon Instinct MI325X AMD	256 GB	6,000 GB/s	Oct 2024	2.7 GB	Q8_0	Comfortable
954 tok/s 573–1,527 · low confidence	Radeon Instinct MI300A AMD	128 GB	5,325 GB/s	Dec 2023	2.7 GB	Q8_0	Comfortable
954 tok/s 573–1,527 · low confidence	Radeon Instinct MI300X AMD	192 GB	5,325 GB/s	Dec 2023	2.7 GB	Q8_0	Comfortable
954 tok/s 573–1,527 · low confidence	Radeon Instinct MI308X AMD	192 GB	5,325 GB/s	Dec 2023	2.7 GB	Q8_0	Comfortable
905 tok/s 543–1,449 · low confidence	H100 NVL 94 GB NVIDIA	94 GB	3,940 GB/s	Mar 2023	2.7 GB	Q8_0	Comfortable
772 tok/s 463–1,235 · low confidence	H100 PCIe 96 GB NVIDIA	96 GB	3,360 GB/s	Mar 2023	2.7 GB	Q8_0	Comfortable
772 tok/s 463–1,235 · low confidence	H100 SXM5 80 GB NVIDIA	80 GB	3,360 GB/s	Oct 2022	2.7 GB	Q8_0	Comfortable
772 tok/s 463–1,235 · low confidence	H100 SXM5 94 GB NVIDIA	94 GB	3,360 GB/s	Mar 2023	2.7 GB	Q8_0	Comfortable
772 tok/s 463–1,235 · low confidence	H100 SXM5 96 GB NVIDIA	96 GB	3,360 GB/s	Mar 2023	2.7 GB	Q8_0	Comfortable
772 tok/s 463–1,235 · low confidence	H800 SXM5 NVIDIA	80 GB	3,360 GB/s	Mar 2023	2.7 GB	Q8_0	Comfortable
588 tok/s 353–941 · low confidence	Radeon Instinct MI250 AMD	128 GB	3,280 GB/s	Nov 2021	2.7 GB	Q8_0	Comfortable
588 tok/s 353–941 · low confidence	Radeon Instinct MI250X AMD	128 GB	3,280 GB/s	Nov 2021	2.7 GB	Q8_0	Comfortable
490 tok/s 294–784 · low confidence	Data Center GPU Max 1550 Intel	128 GB	3,280 GB/s	Jan 2023	2.7 GB	Q8_0	Comfortable
479 tok/s 288–767 · low confidence	Data Center GPU Max Subsystem Intel	128 GB	3,210 GB/s	Jan 2023	2.7 GB	Q8_0	Comfortable
469 tok/s 281–750 · low confidence	A100 SXM4 80 GB NVIDIA	80 GB	2,040 GB/s	Nov 2020	2.7 GB	Q8_0	Comfortable
469 tok/s 281–750 · low confidence	A100X NVIDIA	80 GB	2,040 GB/s	Jun 2021	2.7 GB	Q8_0	Comfortable
469 tok/s 281–750 · low confidence	A800 SXM4 80 GB NVIDIA	80 GB	2,040 GB/s	Aug 2022	2.7 GB	Q8_0	Comfortable
469 tok/s 281–750 · low confidence	H100 CNX NVIDIA	80 GB	2,040 GB/s	Mar 2023	2.7 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: University of Washington,Columbia University,Google,Meta AI,Tel Aviv University
Organisation type: Academia,Academia,Industry,Industry,Academia
Country: United States of America, Israel
Published: 10 March 2022
Authors: Mitchell Wortsman, Gabriel Ilharco, Samir Yitzhak Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S. Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, Ludwig Schmidt

What it does

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

Domain: Vision
Task: Image classification

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: 1.8B
Training data: 1,752,320,000 tokens
Epochs: 8

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: 3.4 × 10²¹ FLOP
How it was established: Operation counting

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 (non-commercial)
Training code: Unreleased

How it is classified

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

Why it is tracked: SOTA improvement
Record confidence: Confident
Citations: 1,499

Sources

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

Reference: Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time
Last updated: 25 May 2026

The extremes

The ten fastest GPUs that run ViT-G (model soup)

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 ViT-G (model soup)

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

Hardware requirements in practice

Minimum card

Tesla C1080

Memory needed

2.7 GB

Fastest

1,838 tok/s

ViT-G (model soup) is small enough at 1.8B 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, Q8_0 compression, roughly 20.0 tokens per second.

The quickest result comes from a B200 at around 1,838 tokens per second — its 8,000 GB/s of bandwidth is what buys that.

About this model

ViT-G (model soup) was published by University of Washington,Columbia University,Google,Meta AI,Tel Aviv University, in United States of America, in March 2022. academia,Academia,Industry,Industry,Academia is the category the publisher falls under.

It works in Vision, and is recorded as doing image classification.

Because its weights were released, nothing about running it depends on a provider staying available — it is yours once downloaded.

How fast it runs, and why

Half the cards that hold it manage more than 51.6 tokens per second, and 792 exceed reading speed outright.

It is a dense model, so every parameter is read for every token produced. That makes speed track memory bandwidth almost exactly — a card with twice the bandwidth generates roughly twice as fast.

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.

What went into building it

Training it took roughly 3.4 × 10²¹ FLOP of computation — a measure of what producing the model cost, not of how fast it answers.

Around 1,752,320,000 tokens went into training it.

Its inclusion criterion is sOTA improvement.

Step by step

How to choose a GPU for ViT-G (model soup)

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

01

Start from the memory column

The table lists every card that can hold ViT-G (model soup) — around 2.7 GB at Q8_0. That figure, not the card's headline performance, is what decides whether it runs.
02

Match the context to your actual use

The conversation occupies memory too, and grows as it goes. Set the slider to the length you expect: at long context ViT-G (model soup) can slip off a card that handles short questions easily.
03

Set a quality floor

Compression is what makes ViT-G (model soup) fit smaller cards, at some cost in accuracy — Q8_0 on the smallest card that fits. A minimum quality removes the ones that go too far.
04

Rank by throughput rather than spec sheet

Ranking by tokens per second for ViT-G (model soup) follows memory bandwidth, not core counts, which is why the B200 tops it at 1,838 tok/s.
05

Check the fit verdict before buying

A tight fit runs ViT-G (model soup) but leaves nothing spare for a longer conversation; comfortable has headroom. If you expect to grow the context, buy for comfortable.
06

See what else that card runs

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

Answers

ViT-G (model soup) — common questions

What is ViT-G (model soup) used for?

ViT-G (model soup) works in Vision, and is recorded as handling image classification. A model can carry several of each, so these are the areas it was built for rather than a limit on what it will attempt.

Where can I download ViT-G (model soup)?

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

How much compute was used to train ViT-G (model soup)?

Around 3.4 × 10²¹ FLOP. 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.

Can I run ViT-G (model soup) 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 ViT-G (model soup) is rarely worth using. Every figure here assumes the whole model is on the card.

Would two GPUs run ViT-G (model soup) faster?

Two cards buy memory rather than speed. That matters for ViT-G (model soup) only if one card cannot hold it — 818 can, so a second adds little.

Why does the quantisation differ between cards for ViT-G (model soup)?

Because capacity varies, so does how hard ViT-G (model soup) has to be squeezed — 1 distinct levels appear in the table above. Set a minimum quality to compare at one.

How accurate are these ViT-G (model soup) speed estimates?

Every figure is derived from memory bandwidth and model size, not benchmarked. That is why each is published as a range such as 1,103–2,942 tok/s on the B200 rather than a single number.

What GPU do I need to run ViT-G (model soup)?

The smallest card in our catalogue that holds ViT-G (model soup) is the Tesla C1080, with 4 GB of memory. It runs the model at Q8_0 using about 2.7 GB, and produces roughly 20.0 tokens per second. 818 cards in total can run it.

How fast is ViT-G (model soup) on a GPU?

It depends on the card. The quickest we calculate is a B200 at about 1,838 tokens per second; the slowest that still runs it manages considerably less. Reading speed is around ten tokens per second, and 792 of the cards that can run ViT-G (model soup) clear that.

How much VRAM does ViT-G (model soup) need?

About 2.7 GB at Q8_0 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.

Can I run ViT-G (model soup) on a 8 GB GPU?

Yes. A CMP 170HX 8 GB with 8 GB runs it at Q8_0, using about 2.7 GB and generating roughly 342 tokens per second — a comfortable fit.

Can I run ViT-G (model soup) on a 12 GB GPU?

Yes. A GeForce RTX 3080 Ti with 12 GB runs it at Q8_0, using about 2.7 GB and generating roughly 210 tokens per second — a comfortable fit.

Can I run ViT-G (model soup) on a 16 GB GPU?

Yes. A Tesla V100 SXM2 16 GB with 16 GB runs it at Q8_0, using about 2.7 GB and generating roughly 260 tokens per second — a comfortable fit.

Can I run ViT-G (model soup) on a 24 GB GPU?

Yes. A GeForce RTX 5090 D V2 with 24 GB runs it at Q8_0, using about 2.7 GB and generating roughly 308 tokens per second — a comfortable fit.

Is ViT-G (model soup) open source?

Its weights are published, so ViT-G (model soup) 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.

How many parameters does ViT-G (model soup) have?

ViT-G (model soup) has 1.8B parameters. This is from the original ViT-G paper. 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.

Who created ViT-G (model soup)?

ViT-G (model soup) was published by University of Washington,Columbia University,Google,Meta AI,Tel Aviv University, based in United States of America, categorised as academia,Academia,Industry,Industry,Academia.

When was ViT-G (model soup) released?

ViT-G (model soup) was published in March 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.

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