FunSearch TPS calculator

Open weights Google DeepMind 15B parameters December 2023

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

306 cards that can run it

818 cards we hold specifications for

Smallest card that fits

P102-101

10 GB · IQ4_XS · 18.9 tok/s

Fastest card

B200

226 tok/s · 180 GB

Which GPUs can run FunSearch?

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

306 cards match

Calculating

					Needs	Quantisation	Fit
226 tok/s 136–361 · low confidence	B200 NVIDIA	180 GB	8,000 GB/s	Jan 2024	16.8 GB	Q8_0	Comfortable
226 tok/s 136–361 · low confidence	B300 NVIDIA	288 GB	8,000 GB/s	Sep 2025	16.8 GB	Q8_0	Comfortable
180 tok/s 108–289 · low confidence	Radeon Instinct MI350X AMD	288 GB	8,190 GB/s	Jan 2025	16.8 GB	Q8_0	Comfortable
180 tok/s 108–289 · low confidence	Radeon Instinct MI355X AMD	288 GB	8,190 GB/s	Jan 2025	16.8 GB	Q8_0	Comfortable
144 tok/s 87–231 · low confidence	Radeon Instinct MI300 AMD	128 GB	6,550 GB/s	Jan 2023	16.8 GB	Q8_0	Comfortable
138 tok/s 83–221 · low confidence	H200 NVL NVIDIA	141 GB	4,890 GB/s	Nov 2024	16.8 GB	Q8_0	Comfortable
138 tok/s 83–221 · low confidence	H200 SXM 141 GB NVIDIA	141 GB	4,890 GB/s	Nov 2024	16.8 GB	Q8_0	Comfortable
132 tok/s 79–211 · low confidence	Radeon Instinct MI325X AMD	256 GB	6,000 GB/s	Oct 2024	16.8 GB	Q8_0	Comfortable
117 tok/s 70–188 · low confidence	Radeon Instinct MI300A AMD	128 GB	5,325 GB/s	Dec 2023	16.8 GB	Q8_0	Comfortable
117 tok/s 70–188 · low confidence	Radeon Instinct MI300X AMD	192 GB	5,325 GB/s	Dec 2023	16.8 GB	Q8_0	Comfortable
117 tok/s 70–188 · low confidence	Radeon Instinct MI308X AMD	192 GB	5,325 GB/s	Dec 2023	16.8 GB	Q8_0	Comfortable
111 tok/s 67–178 · low confidence	H100 NVL 94 GB NVIDIA	94 GB	3,940 GB/s	Mar 2023	16.8 GB	Q8_0	Comfortable
108 tok/s 65–173 · low confidence	CMP 170HX 10 GB NVIDIA	10 GB	1,560 GB/s	Sep 2021	8.9 GB	IQ4_XS	Tight
94.9 tok/s 57–152 · low confidence	H100 PCIe 96 GB NVIDIA	96 GB	3,360 GB/s	Mar 2023	16.8 GB	Q8_0	Comfortable
94.9 tok/s 57–152 · low confidence	H100 SXM5 80 GB NVIDIA	80 GB	3,360 GB/s	Oct 2022	16.8 GB	Q8_0	Comfortable
94.9 tok/s 57–152 · low confidence	H100 SXM5 94 GB NVIDIA	94 GB	3,360 GB/s	Mar 2023	16.8 GB	Q8_0	Comfortable
94.9 tok/s 57–152 · low confidence	H100 SXM5 96 GB NVIDIA	96 GB	3,360 GB/s	Mar 2023	16.8 GB	Q8_0	Comfortable
94.9 tok/s 57–152 · low confidence	H800 SXM5 NVIDIA	80 GB	3,360 GB/s	Mar 2023	16.8 GB	Q8_0	Comfortable
72.2 tok/s 43–116 · low confidence	Radeon Instinct MI250 AMD	128 GB	3,280 GB/s	Nov 2021	16.8 GB	Q8_0	Comfortable
72.2 tok/s 43–116 · low confidence	Radeon Instinct MI250X AMD	128 GB	3,280 GB/s	Nov 2021	16.8 GB	Q8_0	Comfortable
60.2 tok/s 36–96 · low confidence	Data Center GPU Max 1550 Intel	128 GB	3,280 GB/s	Jan 2023	16.8 GB	Q8_0	Comfortable
59.5 tok/s 36–95 · low confidence	GeForce RTX 3080 12 GB NVIDIA	12 GB	912 GB/s	Jan 2022	9.8 GB	Q4_K_M	Tight
59.5 tok/s 36–95 · low confidence	GeForce RTX 3080 Ti NVIDIA	12 GB	912 GB/s	May 2021	9.8 GB	Q4_K_M	Tight
58.9 tok/s 35–94 · low confidence	Data Center GPU Max Subsystem Intel	128 GB	3,210 GB/s	Jan 2023	16.8 GB	Q8_0	Comfortable
57.6 tok/s 35–92 · low confidence	A100 SXM4 80 GB NVIDIA	80 GB	2,040 GB/s	Nov 2020	16.8 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: Google DeepMind
Organisation type: Industry
Country: United States of America
Published: 14 December 2023
Authors: Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Matej Balog, M. Pawan Kumar, Emilien Dupont, Francisco J. R. Ruiz, Jordan S. Ellenberg, Pengming Wang, Omar Fawzi, Pushmeet Kohli, Alhussein Fawzi

What it does

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

Domain: Language, Search
Task: Code generation
Base model: PaLM 2

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: 15B
Training data: 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: 3.9 × 10²³ 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.

Wall-clock time: 48 hours

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: Unreleased

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
Why it is tracked: SOTA improvement,Historical significance
Record confidence: Speculative
Citations: 593

Sources

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

Reference: Mathematical discoveries from program search with large language models
Last updated: 1 January 2026

The extremes

The ten fastest GPUs that run FunSearch

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 FunSearch

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

P102-101

Memory needed

8.9 GB

Fastest

226 tok/s

FunSearch reaches a parameter count of 15B. That is small enough that hardware is rarely the obstacle, including on cards several years old. The number of cards we track that can run it: 306.

The entry point is P102-101, with a memory capacity of 10 GB, running it at a compression of IQ4_XS and producing around 18.9 tokens per second.

At the other end sits B200, generating roughly 226 tokens per second on the strength of a memory bandwidth of 8,000 GB/s.

Background

FunSearch was published by Google DeepMind, in the country recorded as United States of America, during December 2023. It comes out of an organisation categorised as industry.

It works in the domain of Language, Search, and is recorded as performing the task of code generation.

Rather than being trained from scratch, it is derived from PaLM 2. That is the usual way a specialised model is produced.

The weights are published, so it can be downloaded and run on your own hardware indefinitely, offline, with no account attached.

Reading the throughput figures

The median result is around 21.1 tokens per second. Producing text faster than most people read it: 266 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 × 10²³ FLOP. That figure measures what producing the model cost, and has no bearing on how fast it answers.

The reason it appears in this catalogue at all: sOTA improvement,Historical significance.

Step by step

How to choose a GPU for FunSearch

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 FunSearch, needing around 8.9 GB at a compression of IQ4_XS. 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 FunSearch.
03

Set a quality floor

The quantisation column varies by card, because a bigger card holds a more accurate copy, reaching a compression of IQ4_XS 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 FunSearch. It will not match a gaming ordering, because generation is bound by memory bandwidth. The card topping the list is B200, at 226 tok/s.
05

Read the fit column last

The fit column separates cards that just manage it from those with room to spare, in the case of FunSearch. 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

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

Answers

FunSearch — common questions

FunSearch— how much compute was used to train it?

Training consumed around 3.9 × 10²³ 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.

FunSearch— can I run it if it does not fit in my GPU?

Only by offloading, which is usually a false economy: the part held in system memory drags the whole thing down. The nearest miss we calculate falls short by 2.6 GB. Every figure here assumes the whole model is resident on the card.

FunSearch— 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: 306. So a second card is rarely the answer here.

FunSearch— 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: 4. Bigger cards get the more accurate version, and the quality floor above pins the comparison to one level.

FunSearch— 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: 136–361 tok/s on B200. The same model and card vary by thirty to fifty per cent depending on the inference software and its version.

FunSearch— what GPU do I need to run it?

The smallest card in our catalogue that holds it is P102-101, with a memory capacity of 10 GB. It runs the model at a compression of IQ4_XS using about 8.9 GB, and produces roughly 18.9 tokens per second. The number of cards able to run it in total: 306.

FunSearch— how fast is it on a GPU?

It depends on the card. The quickest we calculate is B200, at about 226 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: 266.

FunSearch— how much VRAM does it need?

It needs about 8.9 GB at a compression of IQ4_XS, 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.

FunSearch— can I run it on a GPU holding 12 GB?

Yes. The card GeForce RTX 3080 Ti, holding 12 GB, runs it at a compression of Q4_K_M, using about 9.8 GB and generating roughly 59.5 tokens per second. The fit is tight.

FunSearch— can I run it on a GPU holding 16 GB?

Yes. The card Tesla V100 SXM2 16 GB, holding 16 GB, runs it at a compression of Q6_K, using about 13.3 GB and generating roughly 46.4 tokens per second. The fit is tight.

FunSearch— can I run it on a GPU holding 24 GB?

Yes. The card GeForce RTX 5090 D V2, holding 24 GB, runs it at a compression of Q8_0, using about 16.8 GB and generating roughly 37.8 tokens per second. The fit is comfortable.

FunSearch— 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.

FunSearch— how many parameters does it have?

It has a parameter count of 15B. From the section called "Pretrained LLM": "We use Codey, an LLM built on top of the PaLM2 model family... Because FunSearch relies on sampling from an LLM extensively, an important performance-defining tradeoff is between the quality of the samples and the inference speed of the LLM. In practice, we have chosen to work with a fast-inference model (rather than slower-inference, higher-quality)" Unclear which PaLM2 model was used (of Gecko, Otter, Bison, and Unicorn); above quote indicates it was perhaps Otter or Bison, but not Unicorn. Exact parameter counts are not publicly disclosed for any of these models. In comparisons where FunSearch uses StarCoder-15B, Codey is an improvement but not obviously of an entirely different model class. I report the 15B parameters from StarCoder-15B, used as an open-source comparison. 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.

FunSearch— who created it?

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

FunSearch— when was it released?

It was published in December 2023. Capability per parameter has improved considerably since, so a newer model of the same size is often the better use of the same hardware.

FunSearch— what is it used for?

It works in the domain of Language, Search, and is recorded as handling the task of 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.

FunSearch— where can I download it?

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

Source

Original publication

Record last updated 1 January 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