Llama 3.1-405B 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
Radeon Instinct MI325X
256 GB · Q4_K_M · 11.3 tok/s
Fastest card
B300
19.3 tok/s · 288 GB
Which GPUs can run Llama 3.1-405B?
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.
4 cards match
Calculating| Needs | Quantisation | Fit | |||||
|---|---|---|---|---|---|---|---|
|
19.3
tok/s
16–23 |
B300 NVIDIA | 288 GB | 8,000 GB/s | Sep 2025 | 224.3 GB | Q4_K_M | Tight |
|
15.4
tok/s
9–25 · low confidence |
Radeon Instinct MI350X AMD | 288 GB | 8,190 GB/s | Jan 2025 | 224.3 GB | Q4_K_M | Tight |
|
15.4
tok/s
9–25 · low confidence |
Radeon Instinct MI355X AMD | 288 GB | 8,190 GB/s | Jan 2025 | 224.3 GB | Q4_K_M | Tight |
|
11.3
tok/s
7–18 · low confidence |
Radeon Instinct MI325X AMD | 256 GB | 6,000 GB/s | Oct 2024 | 224.3 GB | Q4_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
- Meta AI
- Organisation type
- Industry
- Country
- United States of America
- Published
- 23 July 2024
- Authors
- Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Alan Schelten, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston Zhang, Aurelien Rodriguez, Austen Gregerson, Ava Spataru, Baptiste Roziere, Bethany Biron, Binh Tang, Bobbie Chern, Charlotte Caucheteux, Ch…
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, Code generation, Mathematical reasoning
- Numerical format
- BF16
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
- 405B
- Training data
- 15,600,000,000,000 tokens
- Epochs
- 1
- Batch size
- 16,000,000
405B
15.6T 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
- 3.8 × 10²⁵ FLOP
- How it was established
- Reported,Operation counting
- Post-training compute
- 9.4 × 10²² FLOP
Stated in paper. Also, 6 * 405B * 15.6T training tokens = 3.8e25
Section 4 gives detail about the post-training process. They do 6 rounds of post-training, using the model from the previous iteration in each successive round. In each round, they fine-tune a copy of the language model into a reward model (RM) using preference data, then use the reward model to do rejection sampling on human annotation prompts. Next they do supervised fine-tuning (SFT) on the rejection sampled data along with some synthetic data (8.5k to 9k steps per round). Next, they do Direc…
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
- 16,384
- Wall-clock time
- 2,142 hours (89.3 days)
- Hardware utilisation
- MFU 40.4%
- Power draw
- 22.6 MW
- Compute cost
- $52,885,434
Trained on 30.84M GPU hours (https://huggingface.co/blog/llama31) and used "up to 16K H100 GPU[s]" so training took at least 30.84M / 16k = 1927.5 hours or ~80 days. Section 3.3.4 gives reliability details over a 54 day period during training, for which they had "higher than 90% effective training time" 1927.5 / 0.9 = 2142 hours Probably, full training time is somewhat longer, since it sounds like there were periods where not all 16k H100s were running.
MFU ranges between 0.38 and 0.43 depending on the specific parallelism used; I assume the geometric mean: sqrt(0.38 * 0.43) = 0.4042
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
- Open (restricted use)
- Hugging Face
- meta-llama
Llama 3.1 model license: https://huggingface.co/meta-llama/Meta-Llama-3.1-8B/blob/main/LICENSE must seek separate license if over 700m monthly users, acceptable use restrictions training code here: https://github.com/meta-llama/llama-recipes/blob/main/src/llama_recipes/utils/train_utils.py#L70
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
- Why it is tracked
- SOTA improvement,Training cost
- Record confidence
- Confident
High training compute, exceeds 4o and Claude 3.5 on some benchmarks: https://ai.meta.com/blog/meta-llama-3-1/
Sources
Where this record came from and when it was last checked.
- Reference
- The Llama 3 Herd of Models
- Last updated
- 28 November 2025
The extremes
The ten fastest GPUs that run Llama 3.1-405B
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 Llama 3.1-405B
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
Radeon Instinct MI325X
Memory needed
224.3 GB
Fastest
19.3 tok/s
Llama 3.1-405B reaches a parameter count of 405B. 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: 4.
At the low end it is handled by Radeon Instinct MI325X, with a memory capacity of 256 GB, running it at a compression of Q4_K_M and producing around 11.3 tokens per second.
At the other end sits B300, generating roughly 19.3 tokens per second on the strength of a memory bandwidth of 8,000 GB/s.
About this model
Llama 3.1-405B was published by Meta AI, in the country recorded as United States of America, during July 2024. 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, Code generation, Mathematical reasoning.
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. On Hugging Face it is published under the organisation meta-llama.
How fast it runs, and why
The median result is around 15.4 tokens per second. Exceeding reading speed outright: 4 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.
Its attention layout is on file, so the memory figures are computed exactly rather than approximated.
Training and provenance
Training it took a computation budget of roughly 3.8 × 10²⁵ 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.
It was trained on a corpus of about 15,600,000,000,000 tokens of text.
Its inclusion criterion: sOTA improvement,Training cost.
Step by step
How to choose a GPU for Llama 3.1-405B
The table above has already assessed every card we hold specifications for against this model. Getting to your answer takes six steps.
-
01
Read the memory figure first
Every card here has been checked against Llama 3.1-405B, needing around 224.3 GB at a compression of Q4_K_M. That figure, not the headline performance of a card, is what decides whether it runs.
-
02
Set the context length you will work at
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 3.1-405B.
-
03
Set a quality floor
Compression is what makes a model fit smaller cards, at some cost in accuracy, reaching a compression of Q4_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
Ranking by tokens per second follows memory bandwidth rather than core counts, for Llama 3.1-405B. It will not match a gaming ordering, because generation is bound by memory bandwidth. The card topping the list is B300, at 19.3 tok/s.
-
05
Read the fit column last
Tight means it loads and works with no room to raise the context later, in the case of Llama 3.1-405B. 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
Check the card from the other side
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 3.1-405B.
Answers
Llama 3.1-405B — common questions
Llama 3.1-405B— would two GPUs run it faster?
A second card roughly doubles the memory available but not the generation rate. The number already able to run it alone: 4. So a second card is rarely the answer here.
Llama 3.1-405B— why does the quantisation differ between cards?
A larger card holds a more accurate copy. The number of compression levels used across the cards that run it: 1. Bigger cards get the more accurate version, and the quality floor above pins the comparison to one level.
Llama 3.1-405B— 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: 16–23 tok/s on B300. The same model and card vary by thirty to fifty per cent depending on the inference software and its version.
Llama 3.1-405B— what GPU do I need to run it?
The smallest card in our catalogue that holds it is Radeon Instinct MI325X, with a memory capacity of 256 GB. It runs the model at a compression of Q4_K_M using about 224.3 GB, and produces roughly 11.3 tokens per second. The number of cards able to run it in total: 4.
Llama 3.1-405B— how fast is it on a GPU?
It depends on the card. The quickest we calculate is B300, at about 19.3 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: 4.
Llama 3.1-405B— how much VRAM does it need?
It needs about 224.3 GB at a compression of Q4_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 3.1-405B— 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 3.1-405B— how many parameters does it have?
It has a parameter count of 405B. 405B. 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 3.1-405B— who created it?
It was published by Meta AI, based in United States of America, an organisation categorised as industry.
Llama 3.1-405B— when was it released?
It was published in July 2024. Capability per parameter has improved considerably since, so a newer model of the same size is often the better use of the same hardware.
Llama 3.1-405B— 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, Code generation, Mathematical reasoning. Models frequently carry more than one of each, and the tags describe purpose rather than capability limits.
Llama 3.1-405B— where can I download it?
Its weights are published on Hugging Face, under the organisation meta-llama. We do not host model files — this site calculates what hardware is needed to run them.
Llama 3.1-405B— how much compute was used to train it?
Training consumed around 3.8 × 10²⁵ 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.
Llama 3.1-405B— can I run it if it does not fit in my GPU?
It can be split between the card and system memory, but it generates painfully slowly that way. The nearest miss we calculate falls short by 51.5 GB. Every figure here assumes the whole model is resident on the card.
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.