Llama Nemotron Ultra 253B TPS calculator
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 we hold specifications for
Smallest card that fits
H200 NVL
141 GB · Q3_K_M · 22.1 tok/s
Fastest card
B200
30.9 tok/s · 180 GB
Which GPUs can run Llama Nemotron Ultra 253B?
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.
9 cards match
Calculating| Needs | Quantisation | Fit | |||||
|---|---|---|---|---|---|---|---|
|
30.9
tok/s
19–49 · low confidence |
B200 NVIDIA | 180 GB | 8,000 GB/s | Jan 2024 | 153.8 GB | Q4_K_M | Tight |
|
22.1
tok/s
13–35 · low confidence |
H200 NVL NVIDIA | 141 GB | 4,890 GB/s | Nov 2024 | 124.3 GB | Q3_K_M | Tight |
|
22.1
tok/s
13–35 · low confidence |
H200 SXM 141 GB NVIDIA | 141 GB | 4,890 GB/s | Nov 2024 | 124.3 GB | Q3_K_M | Tight |
|
19.5
tok/s
12–31 · low confidence |
B300 NVIDIA | 288 GB | 8,000 GB/s | Sep 2025 | 212.7 GB | Q6_K | Comfortable |
|
16.1
tok/s
10–26 · low confidence |
Radeon Instinct MI300X AMD | 192 GB | 5,325 GB/s | Dec 2023 | 153.8 GB | Q4_K_M | Tight |
|
16.1
tok/s
10–26 · low confidence |
Radeon Instinct MI308X AMD | 192 GB | 5,325 GB/s | Dec 2023 | 153.8 GB | Q4_K_M | Tight |
|
15.5
tok/s
9–25 · low confidence |
Radeon Instinct MI350X AMD | 288 GB | 8,190 GB/s | Jan 2025 | 212.7 GB | Q6_K | Comfortable |
|
15.5
tok/s
9–25 · low confidence |
Radeon Instinct MI355X AMD | 288 GB | 8,190 GB/s | Jan 2025 | 212.7 GB | Q6_K | Comfortable |
|
11.4
tok/s
7–18 · low confidence |
Radeon Instinct MI325X AMD | 256 GB | 6,000 GB/s | Oct 2024 | 212.7 GB | Q6_K | 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
- 18 March 2025
- Authors
- Akhiad Bercovich, Itay Levy, Izik Golan, Mohammad Dabbah, Ran El-Yaniv, Omri Puny, Ido Galil, Zach Moshe, Tomer Ronen, Najeeb Nabwani, Ido Shahaf, Oren Tropp, Ehud Karpas, Ran Zilberstein, Jiaqi Zeng, Soumye Singhal, Alexander Bukharin, Yian Zhang, Tugrul Konuk, Gerald Shen, Ameya Sunil Mahabaleshwarkar, Bilal Kartal, Yoshi Suhara, Olivier Delalleau, Zijia Chen, Zhilin Wang, David Mosallanezhad, A…
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, Quantitative reasoning, Code generation, Neural Architecture Search - NAS
- Base model
- Llama 3.1-405B
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
- 253B
- Training data
- 603,000,000,000 tokens
253B "Dense decoder-only Transformer model Network Architecture: Llama-3.1-405B-Instruct, customized through Neural Architecture Search (NAS) **This model was developed based on Llama-3.1-405B-Instruct ** This model has 253B model parameters."
KD + Continued Training: "LN-Ultra is first trained with knowledge distillation for 65B tokens using the same distillation dataset, followed by 88B tokens of continued training on the Nemotron-H phase 4 pretraining dataset (NVIDIA et al., 2025)." (from the paper) Reasoning training data (SFT): for Super model (from the blog) "60B tokens of synthetic data (representing 4M of the 30M generated samples)" -> the entire dataset is ~450B tokens (Ultra model is likely to be trained on the entire dat…
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
- 3.9 × 10²⁵ FLOP
- How it was established
- Operation counting
- Fine-tuning compute
- 1.1 × 10²⁴ FLOP
*Training compute figure is total lifetime (Meta + Nvidia) compute, not incremental Nvidia training compute* Total training compute: 3.8e+25 FLOP (base model) + 1.11e+24 FLOP (fine-tuning) = 3.9e25 FLOP See calculation in the finetune compute notes.
Knowledge Distillation + Continued pre-training + SFT: 6 FLOP / parameter / token * 253000000000 parameters * 6033000000000 tokens [see dataset size notes] = 9.15354e+23 FLOP RL: "the whole training takes approximately 140k H100 hours" 989400000000000 FLOP / sec / GPU [bf16] * 140000 GPU-hours * 3600 sec / hour * 0.4 [assumed utilization] = 1.9946304e+23 FLOP Total: 9.15354e+23 FLOP + 1.9946304e+23 FLOP = 1.114817e+24 FLOP
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 (restricted use)
- Training code
- Unreleased
- Hugging Face
- nvidia
GOVERNING TERMS: Your use of this model is governed by the NVIDIA Open Model License. Additional Information: Llama 3.3 Community License Agreement. Built with Llama. https://huggingface.co/nvidia/Llama-3_1-Nemotron-Ultra-253B-v1
How it is classified
Labels the source dataset applies when tracking notable models, and how confident it is in the entry.
- Frontier model
- Yes
- Likely above 10²³ FLOP
- Yes
- Record confidence
- Likely
Sources
Where this record came from and when it was last checked.
- Reference
- Ultra is 253B distilled from Llama 3.1 405B for maximum agentic accuracy on multi-GPU data center servers.
- Last updated
- 21 July 2026
The extremes
The ten fastest GPUs that run Llama Nemotron Ultra 253B
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.
- 01 B200 180 GB · 8,000 GB/s · Q4_K_M 30.9 tok/s
- 02 H200 NVL 141 GB · 4,890 GB/s · Q3_K_M 22.1 tok/s
- 03 H200 SXM 141 GB 141 GB · 4,890 GB/s · Q3_K_M 22.1 tok/s
- 04 B300 288 GB · 8,000 GB/s · Q6_K 19.5 tok/s
- 05 Radeon Instinct MI300X 192 GB · 5,325 GB/s · Q4_K_M 16.1 tok/s
- 06 Radeon Instinct MI308X 192 GB · 5,325 GB/s · Q4_K_M 16.1 tok/s
- 07 Radeon Instinct MI350X 288 GB · 8,190 GB/s · Q6_K 15.5 tok/s
- 08 Radeon Instinct MI355X 288 GB · 8,190 GB/s · Q6_K 15.5 tok/s
- 09 Radeon Instinct MI325X 256 GB · 6,000 GB/s · Q6_K 11.4 tok/s
The smallest GPUs that still run Llama Nemotron Ultra 253B
The cheapest route in, by memory capacity. A tight fit runs the model but leaves nothing spare for a longer conversation.
- 01 H200 NVL 141 GB · needs 124.3 GB · Q3_K_M · tight 22.1 tok/s
- 02 H200 SXM 141 GB 141 GB · needs 124.3 GB · Q3_K_M · tight 22.1 tok/s
- 03 B200 180 GB · needs 153.8 GB · Q4_K_M · tight 30.9 tok/s
- 04 Radeon Instinct MI300X 192 GB · needs 153.8 GB · Q4_K_M · tight 16.1 tok/s
- 05 Radeon Instinct MI308X 192 GB · needs 153.8 GB · Q4_K_M · tight 16.1 tok/s
- 06 Radeon Instinct MI325X 256 GB · needs 212.7 GB · Q6_K · tight 11.4 tok/s
- 07 B300 288 GB · needs 212.7 GB · Q6_K · comfortable 19.5 tok/s
- 08 Radeon Instinct MI350X 288 GB · needs 212.7 GB · Q6_K · comfortable 15.5 tok/s
- 09 Radeon Instinct MI355X 288 GB · needs 212.7 GB · Q6_K · comfortable 15.5 tok/s
What the numbers mean
What you need to run it
Minimum card
H200 NVL
Memory needed
124.3 GB
Fastest
30.9 tok/s
Llama Nemotron Ultra 253B reaches a parameter count of 253B. That is beyond what any single graphics card holds. Running it means either splitting it across several cards or renting hardware built for the job, and every card able to hold it alone is a datacentre part. The number that can: 9.
The least hardware that works is H200 NVL, with a memory capacity of 141 GB, running it at a compression of Q3_K_M and producing around 22.1 tokens per second.
At the other end sits B200, generating roughly 30.9 tokens per second on the strength of a memory bandwidth of 8,000 GB/s.
Where it came from
Llama Nemotron Ultra 253B was published by NVIDIA, in the country recorded as United States of America, during March 2025. The category the publisher falls under is industry.
It works in the domain of Language, and is recorded as performing the task of language modeling/generation, Question answering, Quantitative reasoning, Code generation, Neural Architecture Search - NAS.
Rather than being trained from scratch, it is derived from Llama 3.1-405B. Most models at this scale are adapted from an existing base rather than built from nothing.
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.
Understanding the speeds
Across every card that can run it, the middle of the range sits at 16.1 tokens per second. Producing text faster than most people read it: 9 of them.
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.
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
The training run consumed about 3.9 × 10²⁵ FLOP. That figure measures what producing the model cost, and has no bearing on how fast it answers.
The training set ran to roughly 603,000,000,000 tokens of text.
Step by step
How to choose a GPU for Llama Nemotron Ultra 253B
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
Every card here has been checked against Llama Nemotron Ultra 253B, needing around 124.3 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
The conversation occupies memory too, and grows as it goes. Set the slider to the length you expect, because at long context a card that handles short questions easily can be dropped by Llama Nemotron Ultra 253B.
-
03
Set a quality floor
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
Compare tokens per second, not specifications
The speed ordering is effectively an ordering by memory bandwidth, for Llama Nemotron Ultra 253B. It will not match a gaming ordering, because generation is bound by memory bandwidth. The card topping the list is B200, at 30.9 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 Llama Nemotron Ultra 253B. 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
See what else that card runs
Following a card through to its own page shows every other model it can hold, which is the question that follows once you have settled on Llama Nemotron Ultra 253B.
Answers
Llama Nemotron Ultra 253B — common questions
Llama Nemotron Ultra 253B— what GPU do I need to run it?
The smallest card in our catalogue that holds it is H200 NVL, with a memory capacity of 141 GB. It runs the model at a compression of Q3_K_M using about 124.3 GB, and produces roughly 22.1 tokens per second. The number of cards able to run it in total: 9.
Llama Nemotron Ultra 253B— how fast is it on a GPU?
It depends on the card. The quickest we calculate is B200, at about 30.9 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: 9.
Llama Nemotron Ultra 253B— how much VRAM does it need?
It needs about 124.3 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.
Llama Nemotron Ultra 253B— 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.
Llama Nemotron Ultra 253B— how many parameters does it have?
It has a parameter count of 253B. 253B "Dense decoder-only Transformer model Network Architecture: Llama-3.1-405B-Instruct, customized through Neural Architecture Search (NAS) **This model was developed based on Llama-3.1-405B-Instruct ** This model has 253B 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.
Llama Nemotron Ultra 253B— who created it?
It was published by NVIDIA, based in United States of America, an organisation categorised as industry.
Llama Nemotron Ultra 253B— when was it released?
It was published in March 2025.
Llama Nemotron Ultra 253B— 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, Quantitative reasoning, Code generation, Neural Architecture Search - NAS. Models frequently carry more than one of each, and the tags describe purpose rather than capability limits.
Llama Nemotron Ultra 253B— 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.
Llama Nemotron Ultra 253B— how much compute was used to train it?
Training consumed around 3.9 × 10²⁵ 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.
Llama Nemotron Ultra 253B— 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 38.6 GB. Every figure here assumes the whole model is resident on the card.
Llama Nemotron Ultra 253B— would two GPUs run it faster?
Capacity adds across cards; throughput does not. The number of cards already holding it on their own: 9. So a second card is rarely the answer here.
Llama Nemotron Ultra 253B— 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: 3. Bigger cards get the more accurate version, and the quality floor above pins the comparison to one level.
Llama Nemotron Ultra 253B— 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: 19–49 tok/s on B200. The same model and card vary by thirty to fifty per cent depending on the inference software and its version.
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.