Nemotron-H 56B TPS calculator

Open weights NVIDIA 56B parameters April 2025

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

92 cards that can run it

818 cards we hold specifications for

Smallest card that fits

FirePro S9170

32 GB · Q3_K_M · 5.1 tok/s

Fastest card

B200

60.5 tok/s · 180 GB

Which GPUs can run Nemotron-H 56B?

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

92 cards match

Calculating

					Needs	Quantisation	Fit
60.5 tok/s 36–97 · low confidence	B200 NVIDIA	180 GB	8,000 GB/s	Jan 2024	60.7 GB	Q8_0	Comfortable
60.5 tok/s 36–97 · low confidence	B300 NVIDIA	288 GB	8,000 GB/s	Sep 2025	60.7 GB	Q8_0	Comfortable
48.3 tok/s 29–77 · low confidence	Radeon Instinct MI350X AMD	288 GB	8,190 GB/s	Jan 2025	60.7 GB	Q8_0	Comfortable
48.3 tok/s 29–77 · low confidence	Radeon Instinct MI355X AMD	288 GB	8,190 GB/s	Jan 2025	60.7 GB	Q8_0	Comfortable
38.6 tok/s 23–62 · low confidence	Radeon Instinct MI300 AMD	128 GB	6,550 GB/s	Jan 2023	60.7 GB	Q8_0	Comfortable
38.2 tok/s 23–61 · low confidence	DRIVE A100 PROD NVIDIA	32 GB	1,870 GB/s	May 2020	28.1 GB	Q3_K_M	Tight
38.2 tok/s 23–61 · low confidence	GRID A100A NVIDIA	32 GB	1,870 GB/s	May 2020	28.1 GB	Q3_K_M	Tight
37.0 tok/s 22–59 · low confidence	H200 NVL NVIDIA	141 GB	4,890 GB/s	Nov 2024	60.7 GB	Q8_0	Comfortable
37.0 tok/s 22–59 · low confidence	H200 SXM 141 GB NVIDIA	141 GB	4,890 GB/s	Nov 2024	60.7 GB	Q8_0	Comfortable
36.5 tok/s 22–58 · low confidence	GeForce RTX 5090 NVIDIA	32 GB	1,790 GB/s	Jan 2025	28.1 GB	Q3_K_M	Tight
36.5 tok/s 22–58 · low confidence	GeForce RTX 5090 D NVIDIA	32 GB	1,790 GB/s	Jan 2025	28.1 GB	Q3_K_M	Tight
35.4 tok/s 21–57 · low confidence	Radeon Instinct MI325X AMD	256 GB	6,000 GB/s	Oct 2024	60.7 GB	Q8_0	Comfortable
31.4 tok/s 19–50 · low confidence	Radeon Instinct MI300A AMD	128 GB	5,325 GB/s	Dec 2023	60.7 GB	Q8_0	Comfortable
31.4 tok/s 19–50 · low confidence	Radeon Instinct MI300X AMD	192 GB	5,325 GB/s	Dec 2023	60.7 GB	Q8_0	Comfortable
31.4 tok/s 19–50 · low confidence	Radeon Instinct MI308X AMD	192 GB	5,325 GB/s	Dec 2023	60.7 GB	Q8_0	Comfortable
29.8 tok/s 18–48 · low confidence	H100 NVL 94 GB NVIDIA	94 GB	3,940 GB/s	Mar 2023	60.7 GB	Q8_0	Comfortable
27.2 tok/s 16–44 · low confidence	A100 PCIe 40 GB NVIDIA	40 GB	1,560 GB/s	Jun 2020	34.6 GB	Q4_K_M	Tight
27.2 tok/s 16–44 · low confidence	A100 SXM4 40 GB NVIDIA	40 GB	1,560 GB/s	May 2020	34.6 GB	Q4_K_M	Tight
27.2 tok/s 16–44 · low confidence	A800 PCIe 40 GB NVIDIA	40 GB	1,560 GB/s	Nov 2022	34.6 GB	Q4_K_M	Tight
25.4 tok/s 15–41 · low confidence	H100 PCIe 96 GB NVIDIA	96 GB	3,360 GB/s	Mar 2023	60.7 GB	Q8_0	Comfortable
25.4 tok/s 15–41 · low confidence	H100 SXM5 80 GB NVIDIA	80 GB	3,360 GB/s	Oct 2022	60.7 GB	Q8_0	Tight
25.4 tok/s 15–41 · low confidence	H100 SXM5 94 GB NVIDIA	94 GB	3,360 GB/s	Mar 2023	60.7 GB	Q8_0	Comfortable
25.4 tok/s 15–41 · low confidence	H100 SXM5 96 GB NVIDIA	96 GB	3,360 GB/s	Mar 2023	60.7 GB	Q8_0	Comfortable
25.4 tok/s 15–41 · low confidence	H800 SXM5 NVIDIA	80 GB	3,360 GB/s	Mar 2023	60.7 GB	Q8_0	Tight
25.3 tok/s 15–40 · low confidence	GRID A100B NVIDIA	48 GB	1,870 GB/s	May 2020	41.1 GB	Q5_K_M	Tight

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: NVIDIA
Organisation type: Industry
Country: United States of America
Published: 14 April 2025
Authors: NVIDIA: Aaron Blakeman, Aarti Basant, Abhinav Khattar, Adithya Renduchintala, Akhiad Bercovich, Aleksander Ficek, Alexis Bjorlin, Ali Taghibakhshi, Amala Sanjay Deshmukh, Ameya Sunil Mahabaleshwarkar, Andrew Tao, Anna Shors, Ashwath Aithal, Ashwin Poojary, Ayush Dattagupta, Balaram Buddharaju, Bobby Chen, Boris Ginsburg, Boxin Wang, Brandon Norick, Brian Butterfield, Bryan Catanzaro, Carlo del Mun…

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, Translation, Quantitative reasoning, Code generation
Numerical format: FP8

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: 56B
Training data: tokens
Epochs: 1
Batch size: 6,291,456

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: 6.7 × 10²⁴ FLOP
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: NVIDIA H100 SXM5 80GB
Chips used: 6,144
Power draw: 8.4 MW

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
Hugging Face: nvidia

How it is classified

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

Likely above 10²³ FLOP: Yes
Record confidence: Confident

Sources

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

Reference: Nemotron-H: A Family of Accurate and Efficient Hybrid Mamba-Transformer Models
Last updated: 28 November 2025

The extremes

The ten fastest GPUs that run Nemotron-H 56B

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 Nemotron-H 56B

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

The hardware side

Minimum card

FirePro S9170

Memory needed

28.1 GB

Fastest

60.5 tok/s

Nemotron-H 56B reaches a parameter count of 56B. 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: 92.

At the low end it is handled by FirePro S9170, with a memory capacity of 32 GB, running it at a compression of Q3_K_M and producing around 5.1 tokens per second.

The fastest we calculate for it is B200, generating roughly 60.5 tokens per second on the strength of a memory bandwidth of 8,000 GB/s.

Background

Nemotron-H 56B was published by NVIDIA, in the country recorded as United States of America, during April 2025. The publishing organisation is categorised as industry.

It works in the domain of Language, and is recorded as performing the task of language modeling/generation, Question answering, Translation, Quantitative reasoning, Code generation.

The weights are published, so it can be downloaded and run on your own hardware indefinitely, offline, with no account attached. On Hugging Face it is published under the organisation nvidia.

Reading the throughput figures

Half the cards that hold it manage more than 15.4 tokens per second. Clearing the ten tokens per second that roughly matches reading speed: 72 of them.

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

Memory here is estimated from size rather than computed from the architecture, which is not recorded for this model — the numbers are indicative rather than exact.

Training and provenance

Producing it required arithmetic totalling around 6.7 × 10²⁴ FLOP, on hardware recorded as NVIDIA H100 SXM5 80GB. That figure measures what producing the model cost, and has no bearing on how fast it answers.

Step by step

How to choose a GPU for Nemotron-H 56B

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

Start from what it actually needs, which is the requirement of Nemotron-H 56B, needing around 28.1 GB at a compression of Q3_K_M. Capacity is the gate — a card either holds it or it does not.
02

Decide how long your conversations run

Longer conversations cost memory on top of what the weights need. Move the slider to your real working length before trusting any row for Nemotron-H 56B.
03

Choose how far you will compress it

Compression is what makes a model fit smaller cards, at some cost in accuracy, 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

Compare tokens per second, not specifications

Sort by speed to see how cards rank for Nemotron-H 56B. It will not match a gaming ordering, because generation is bound by memory bandwidth. The card topping the list is B200, at 60.5 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 Nemotron-H 56B. 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

Open the card you have settled on

Each card page repeats this sweep for every model we hold, answering what else the hardware is good for beyond Nemotron-H 56B.

Answers

Nemotron-H 56B — common questions

Nemotron-H 56B— 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 9.4 GB. Every figure here assumes the whole model is resident on the card.

Nemotron-H 56B— would two GPUs run it faster?

Two cards buy memory rather than speed, which matters only if one card cannot hold it. The number that can: 92. So a second card is rarely the answer here.

Nemotron-H 56B— why does the quantisation differ between cards?

Because capacity varies, so does how hard it has to be squeezed. The number of distinct levels in the table above: 5. Bigger cards get the more accurate version, and the quality floor above pins the comparison to one level.

Nemotron-H 56B— 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: 36–97 tok/s on B200. The same model and card vary by thirty to fifty per cent depending on the inference software and its version.

Nemotron-H 56B— what GPU do I need to run it?

The smallest card in our catalogue that holds it is FirePro S9170, with a memory capacity of 32 GB. It runs the model at a compression of Q3_K_M using about 28.1 GB, and produces roughly 5.1 tokens per second. The number of cards able to run it in total: 92.

Nemotron-H 56B— how fast is it on a GPU?

It depends on the card. The quickest we calculate is B200, at about 60.5 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: 72.

Nemotron-H 56B— how much VRAM does it need?

It needs about 28.1 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.

Nemotron-H 56B— 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.

Nemotron-H 56B— how many parameters does it have?

It has a parameter count of 56B. Model Architecture Architecture Type: Hybrid Mamba-Transformer Network Architecture: Nemotron-H This model has 56B model parameters. 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.

Nemotron-H 56B— who created it?

It was published by NVIDIA, based in United States of America, an organisation categorised as industry.

Nemotron-H 56B— when was it released?

It was published in April 2025.

Nemotron-H 56B— 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, Translation, Quantitative reasoning, Code generation. 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.

Nemotron-H 56B— where can I download it?

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

Nemotron-H 56B— how much compute was used to train it?

Training consumed around 6.7 × 10²⁴ FLOP, on hardware recorded as NVIDIA H100 SXM5 80GB. 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.

Source

Original publication

Record last updated 28 November 2025

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