SciBERT TPS calculator

Open weights Allen Institute for AI 110M parameters March 2019

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 · 335 tok/s

Fastest card

B200

30,802 tok/s · 180 GB

Which GPUs can run SciBERT?

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
30,802 tok/s 18,481–49,283 · low confidence	B200 NVIDIA	180 GB	8,000 GB/s	Jan 2024	0.8 GB	Q8_0	Comfortable
30,802 tok/s 18,481–49,283 · low confidence	B300 NVIDIA	288 GB	8,000 GB/s	Sep 2025	0.8 GB	Q8_0	Comfortable
24,596 tok/s 14,758–39,354 · low confidence	Radeon Instinct MI350X AMD	288 GB	8,190 GB/s	Jan 2025	0.8 GB	Q8_0	Comfortable
24,596 tok/s 14,758–39,354 · low confidence	Radeon Instinct MI355X AMD	288 GB	8,190 GB/s	Jan 2025	0.8 GB	Q8_0	Comfortable
19,671 tok/s 11,803–31,474 · low confidence	Radeon Instinct MI300 AMD	128 GB	6,550 GB/s	Jan 2023	0.8 GB	Q8_0	Comfortable
18,828 tok/s 11,297–30,124 · low confidence	H200 NVL NVIDIA	141 GB	4,890 GB/s	Nov 2024	0.8 GB	Q8_0	Comfortable
18,828 tok/s 11,297–30,124 · low confidence	H200 SXM 141 GB NVIDIA	141 GB	4,890 GB/s	Nov 2024	0.8 GB	Q8_0	Comfortable
18,019 tok/s 10,812–28,831 · low confidence	Radeon Instinct MI325X AMD	256 GB	6,000 GB/s	Oct 2024	0.8 GB	Q8_0	Comfortable
15,992 tok/s 9,595–25,587 · low confidence	Radeon Instinct MI300A AMD	128 GB	5,325 GB/s	Dec 2023	0.8 GB	Q8_0	Comfortable
15,992 tok/s 9,595–25,587 · low confidence	Radeon Instinct MI300X AMD	192 GB	5,325 GB/s	Dec 2023	0.8 GB	Q8_0	Comfortable
15,992 tok/s 9,595–25,587 · low confidence	Radeon Instinct MI308X AMD	192 GB	5,325 GB/s	Dec 2023	0.8 GB	Q8_0	Comfortable
15,170 tok/s 9,102–24,272 · low confidence	H100 NVL 94 GB NVIDIA	94 GB	3,940 GB/s	Mar 2023	0.8 GB	Q8_0	Comfortable
12,937 tok/s 7,762–20,699 · low confidence	H100 PCIe 96 GB NVIDIA	96 GB	3,360 GB/s	Mar 2023	0.8 GB	Q8_0	Comfortable
12,937 tok/s 7,762–20,699 · low confidence	H100 SXM5 80 GB NVIDIA	80 GB	3,360 GB/s	Oct 2022	0.8 GB	Q8_0	Comfortable
12,937 tok/s 7,762–20,699 · low confidence	H100 SXM5 94 GB NVIDIA	94 GB	3,360 GB/s	Mar 2023	0.8 GB	Q8_0	Comfortable
12,937 tok/s 7,762–20,699 · low confidence	H100 SXM5 96 GB NVIDIA	96 GB	3,360 GB/s	Mar 2023	0.8 GB	Q8_0	Comfortable
12,937 tok/s 7,762–20,699 · low confidence	H800 SXM5 NVIDIA	80 GB	3,360 GB/s	Mar 2023	0.8 GB	Q8_0	Comfortable
9,851 tok/s 5,910–15,761 · low confidence	Radeon Instinct MI250 AMD	128 GB	3,280 GB/s	Nov 2021	0.8 GB	Q8_0	Comfortable
9,851 tok/s 5,910–15,761 · low confidence	Radeon Instinct MI250X AMD	128 GB	3,280 GB/s	Nov 2021	0.8 GB	Q8_0	Comfortable
8,209 tok/s 4,925–13,134 · low confidence	Data Center GPU Max 1550 Intel	128 GB	3,280 GB/s	Jan 2023	0.8 GB	Q8_0	Comfortable
8,034 tok/s 4,820–12,854 · low confidence	Data Center GPU Max Subsystem Intel	128 GB	3,210 GB/s	Jan 2023	0.8 GB	Q8_0	Comfortable
7,855 tok/s 4,713–12,567 · low confidence	A100 SXM4 80 GB NVIDIA	80 GB	2,040 GB/s	Nov 2020	0.8 GB	Q8_0	Comfortable
7,855 tok/s 4,713–12,567 · low confidence	A100X NVIDIA	80 GB	2,040 GB/s	Jun 2021	0.8 GB	Q8_0	Comfortable
7,855 tok/s 4,713–12,567 · low confidence	A800 SXM4 80 GB NVIDIA	80 GB	2,040 GB/s	Aug 2022	0.8 GB	Q8_0	Comfortable
7,855 tok/s 4,713–12,567 · low confidence	H100 CNX NVIDIA	80 GB	2,040 GB/s	Mar 2023	0.8 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
Organisation type: Research collective
Country: United States of America
Published: 26 March 2019
Authors: Iz Beltagy, Kyle Lo, Arman Cohan

What it does

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

Domain: Language
Task: Relation extraction, Sentiment classification, Text classification, Named entity recognition (NER)

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: 110M
Training data: 3,170,000,000 tokens

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: 8.9 × 10¹⁹ FLOP
How it was established: Hardware

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 v3
Chips used: 4
Chip-hours: 672
Wall-clock time: 168 hours (7 days)
Power draw: 3.7 kW
Compute cost: $247

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: Open source

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: Highly cited,SOTA improvement
Record confidence: Confident
Citations: 3,705

Sources

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

Reference: SciBERT: A Pretrained Language Model for Scientific Text
Last updated: 25 May 2026

The extremes

The ten fastest GPUs that run SciBERT

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 SciBERT

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 it takes to run this model

Minimum card

Tesla C1080

Memory needed

0.8 GB

Fastest

30,802 tok/s

SciBERT is small enough at 110M parameters that hardware is rarely the obstacle — 818 of the cards we track can run it, including cards several years old.

The smallest card that holds it is the Tesla C1080 with 4 GB, running it at Q8_0 and producing around 335 tokens per second.

Top of the range is the B200, at roughly 30,802 tokens per second thanks to 8,000 GB/s of bandwidth.

What this model is

SciBERT was published by Allen Institute for AI, in United States of America, in March 2019. It comes out of research collective.

It works in Language, and is recorded as doing relation extraction, Sentiment classification, Text classification, Named entity recognition (NER).

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.

What decides the speed

The median result is around 864.9 tokens per second; 818 cards produce text faster than most people read it.

Every weight participates in every token here, so bandwidth is the whole story: the ranking below is effectively a ranking of memory throughput.

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.

Training and provenance

Training it took roughly 8.9 × 10¹⁹ FLOP of computation, on Google TPU v3 — a measure of what producing the model cost, not of how fast it answers.

It was trained on about 3,170,000,000 tokens of text.

Its inclusion criterion is highly cited,SOTA improvement.

Step by step

How to choose a GPU for SciBERT

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

01

Check what it needs before anything else

Look at what SciBERT actually needs — around 0.8 GB at Q8_0. No amount of processing power compensates for a card that cannot hold it.
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 SciBERT can slip off a card that handles short questions easily.
03

Choose how far you will compress it

The quantisation column varies by card, because a bigger card holds a more accurate copy of SciBERT — Q8_0 on the smallest card that fits. Set a floor to hold the comparison at one level.
04

Compare tokens per second, not specifications

Sort by speed to see how cards rank for SciBERT. It will not match a gaming ordering — generation is bound by memory bandwidth, which is why the B200 tops it at 30,802 tok/s.
05

Read the fit column last

Tight means SciBERT loads and works, with no room to raise the context later. Comfortable means you can. The difference matters more than a few tokens per second.
06

Check the card from the other side

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

Answers

SciBERT — common questions

Can I run SciBERT on a 24 GB GPU?

Yes. A GeForce RTX 5090 D V2 with 24 GB runs it at Q8_0, using about 0.8 GB and generating roughly 5,159 tokens per second — a comfortable fit.

Is SciBERT open source?

Its weights are published, so SciBERT 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 SciBERT have?

SciBERT has 110M parameters. 110M size of bert base from https://huggingface.co/google-bert/bert-base-uncased relevant citation: "We use the original BERT code to train SCIBERT on our corpus with the same con- figuration and size as BERT-Base. We train 4 different versions of SCIBERT: (i) cased or un- cased and (ii) BASEVOCAB or SCIVOCAB. The two models that use BASEVOCAB are finetuned from the corresponding BERT-Base models. The other two models that use the new SCIVOCAB are trained from scratch.". 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 SciBERT?

SciBERT was published by Allen Institute for AI, based in United States of America, categorised as research collective.

When was SciBERT released?

SciBERT was published in March 2019. Capability per parameter has improved considerably since, so a newer model of the same size is often the better use of the same hardware.

What is SciBERT used for?

SciBERT works in Language, and is recorded as handling relation extraction, Sentiment classification, Text classification, Named entity recognition (NER). Models frequently carry more than one of each, and the tags describe purpose rather than capability limits.

Where can I download SciBERT?

The weights for SciBERT 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 SciBERT?

Around 8.9 × 10¹⁹ FLOP, on Google TPU v3. 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 SciBERT if it does not fit in my GPU?

It can be split between the card and system memory, but SciBERT generates painfully slowly that way. Nothing on this page assumes offloading.

Would two GPUs run SciBERT faster?

Two cards buy memory rather than speed. That matters for SciBERT only if one card cannot hold it — 818 can, so a second adds little.

Why does the quantisation differ between cards for SciBERT?

A larger card holds a more accurate copy. Across the cards that run SciBERT, 1 compression levels are used; the floor control above pins it to one.

How accurate are these SciBERT 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 18,481–49,283 tok/s on the B200 rather than a single number.

What GPU do I need to run SciBERT?

The smallest card in our catalogue that holds SciBERT is the Tesla C1080, with 4 GB of memory. It runs the model at Q8_0 using about 0.8 GB, and produces roughly 335 tokens per second. 818 cards in total can run it.

How fast is SciBERT on a GPU?

It depends on the card. The quickest we calculate is a B200 at about 30,802 tokens per second; the slowest that still runs it manages considerably less. Reading speed is around ten tokens per second, and 818 of the cards that can run SciBERT clear that.

How much VRAM does SciBERT need?

About 0.8 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 SciBERT on a 8 GB GPU?

Yes. A CMP 170HX 8 GB with 8 GB runs it at Q8_0, using about 0.8 GB and generating roughly 5,737 tokens per second — a comfortable fit.

Can I run SciBERT on a 12 GB GPU?

Yes. A GeForce RTX 3080 Ti with 12 GB runs it at Q8_0, using about 0.8 GB and generating roughly 3,513 tokens per second — a comfortable fit.

Can I run SciBERT on a 16 GB GPU?

Yes. A Tesla V100 SXM2 16 GB with 16 GB runs it at Q8_0, using about 0.8 GB and generating roughly 4,351 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