SPHINX (Llama 2 13B) TPS calculator

Open weights Shanghai AI Lab,Chinese University of Hong Kong (CUHK),ShanghaiTech University 19.9B parameters November 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

293 cards that can run it

818 cards we hold specifications for

Smallest card that fits

Quadro K6000

12 GB · Q3_K_M · 14.1 tok/s

Fastest card

B200

170 tok/s · 180 GB

Which GPUs can run SPHINX (Llama 2 13B)?

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

293 cards match

Calculating

					Needs	Quantisation	Fit
170 tok/s 102–272 · low confidence	B200 NVIDIA	180 GB	8,000 GB/s	Jan 2024	22.0 GB	Q8_0	Comfortable
170 tok/s 102–272 · low confidence	B300 NVIDIA	288 GB	8,000 GB/s	Sep 2025	22.0 GB	Q8_0	Comfortable
136 tok/s 82–218 · low confidence	Radeon Instinct MI350X AMD	288 GB	8,190 GB/s	Jan 2025	22.0 GB	Q8_0	Comfortable
136 tok/s 82–218 · low confidence	Radeon Instinct MI355X AMD	288 GB	8,190 GB/s	Jan 2025	22.0 GB	Q8_0	Comfortable
109 tok/s 65–174 · low confidence	Radeon Instinct MI300 AMD	128 GB	6,550 GB/s	Jan 2023	22.0 GB	Q8_0	Comfortable
104 tok/s 62–167 · low confidence	H200 NVL NVIDIA	141 GB	4,890 GB/s	Nov 2024	22.0 GB	Q8_0	Comfortable
104 tok/s 62–167 · low confidence	H200 SXM 141 GB NVIDIA	141 GB	4,890 GB/s	Nov 2024	22.0 GB	Q8_0	Comfortable
99.6 tok/s 60–159 · low confidence	Radeon Instinct MI325X AMD	256 GB	6,000 GB/s	Oct 2024	22.0 GB	Q8_0	Comfortable
88.4 tok/s 53–141 · low confidence	Radeon Instinct MI300A AMD	128 GB	5,325 GB/s	Dec 2023	22.0 GB	Q8_0	Comfortable
88.4 tok/s 53–141 · low confidence	Radeon Instinct MI300X AMD	192 GB	5,325 GB/s	Dec 2023	22.0 GB	Q8_0	Comfortable
88.4 tok/s 53–141 · low confidence	Radeon Instinct MI308X AMD	192 GB	5,325 GB/s	Dec 2023	22.0 GB	Q8_0	Comfortable
83.9 tok/s 50–134 · low confidence	H100 NVL 94 GB NVIDIA	94 GB	3,940 GB/s	Mar 2023	22.0 GB	Q8_0	Comfortable
71.5 tok/s 43–114 · low confidence	H100 PCIe 96 GB NVIDIA	96 GB	3,360 GB/s	Mar 2023	22.0 GB	Q8_0	Comfortable
71.5 tok/s 43–114 · low confidence	H100 SXM5 80 GB NVIDIA	80 GB	3,360 GB/s	Oct 2022	22.0 GB	Q8_0	Comfortable
71.5 tok/s 43–114 · low confidence	H100 SXM5 94 GB NVIDIA	94 GB	3,360 GB/s	Mar 2023	22.0 GB	Q8_0	Comfortable
71.5 tok/s 43–114 · low confidence	H100 SXM5 96 GB NVIDIA	96 GB	3,360 GB/s	Mar 2023	22.0 GB	Q8_0	Comfortable
71.5 tok/s 43–114 · low confidence	H800 SXM5 NVIDIA	80 GB	3,360 GB/s	Mar 2023	22.0 GB	Q8_0	Comfortable
55.5 tok/s 33–89 · low confidence	Tesla V100 SXM2 16 GB NVIDIA	16 GB	1,130 GB/s	Nov 2019	12.7 GB	Q4_K_M	Tight
54.5 tok/s 33–87 · low confidence	Radeon Instinct MI250 AMD	128 GB	3,280 GB/s	Nov 2021	22.0 GB	Q8_0	Comfortable
54.5 tok/s 33–87 · low confidence	Radeon Instinct MI250X AMD	128 GB	3,280 GB/s	Nov 2021	22.0 GB	Q8_0	Comfortable
52.4 tok/s 31–84 · low confidence	GeForce RTX 3080 12 GB NVIDIA	12 GB	912 GB/s	Jan 2022	10.4 GB	Q3_K_M	Tight
52.4 tok/s 31–84 · low confidence	GeForce RTX 3080 Ti NVIDIA	12 GB	912 GB/s	May 2021	10.4 GB	Q3_K_M	Tight
47.2 tok/s 28–75 · low confidence	GeForce RTX 5080 NVIDIA	16 GB	960 GB/s	Jan 2025	12.7 GB	Q4_K_M	Tight
45.4 tok/s 27–73 · low confidence	Data Center GPU Max 1550 Intel	128 GB	3,280 GB/s	Jan 2023	22.0 GB	Q8_0	Comfortable
44.4 tok/s 27–71 · low confidence	Data Center GPU Max Subsystem Intel	128 GB	3,210 GB/s	Jan 2023	22.0 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: Shanghai AI Lab,Chinese University of Hong Kong (CUHK),ShanghaiTech University
Organisation type: Academia,Academia,Academia
Country: China, Hong Kong
Published: 13 November 2023
Authors: Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, Jiaming Han, Siyuan Huang, Yichi Zhang, Xuming He, Hongsheng Li, Yu Qiao

What it does

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

Domain: Vision, Language, Multimodal
Task: Visual question answering, Image captioning
Base model: Llama 2-13B
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: 19.9B
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 × 10²² FLOP
How it was established: Hardware
Fine-tuning compute: 4 × 10²¹ FLOP

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 A100 SXM4 40 GB
Chips used: 32
Chip-hours: 9,280
Wall-clock time: 290 hours (12.1 days)
Power draw: 25.4 kW
Compute cost: $239,189

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: Alpha-VLLM

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: SOTA improvement
Record confidence: Likely

Sources

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

Reference: SPHINX: The Joint Mixing of Weights, Tasks, and Visual Embeddings for Multi-modal Large Language Models
Last updated: 11 February 2026

The extremes

The ten fastest GPUs that run SPHINX (Llama 2 13B)

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 SPHINX (Llama 2 13B)

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

Quadro K6000

Memory needed

10.4 GB

Fastest

170 tok/s

SPHINX (Llama 2 13B) reaches a parameter count of 19.9B. 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: 293.

The entry point is Quadro K6000, with a memory capacity of 12 GB, running it at a compression of Q3_K_M and producing around 14.1 tokens per second.

The fastest we calculate for it is B200, generating roughly 170 tokens per second on the strength of a memory bandwidth of 8,000 GB/s.

Where it came from

SPHINX (Llama 2 13B) was published by Shanghai AI Lab,Chinese University of Hong Kong (CUHK),ShanghaiTech University, in the country recorded as China, during November 2023. The publishing organisation is categorised as academia,Academia,Academia.

It works in the domain of Vision, Language, Multimodal, and is recorded as performing the task of visual question answering, Image captioning.

Rather than being trained from scratch, it is derived from Llama 2-13B. That is why it shares the base model's general shape and size.

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 Alpha-VLLM.

Understanding the speeds

The median result is around 20.7 tokens per second. Exceeding reading speed outright: 248 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.

What went into building it

Producing it required arithmetic totalling around 3 × 10²² FLOP, on hardware recorded as NVIDIA A100 SXM4 40 GB. 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.

Step by step

How to choose a GPU for SPHINX (Llama 2 13B)

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

01

Start from the memory column

Every card here has been checked against SPHINX (Llama 2 13B), needing around 10.4 GB at a compression of Q3_K_M. That figure, not the headline performance of a card, is what decides whether it runs.
02

Decide how long your conversations run

Set the context to what you will actually use. The cache grows with the conversation, and it is the usual reason a card that seemed fine stops fitting SPHINX (Llama 2 13B).
03

Set a quality floor

Each card runs the least-compressed copy it can hold, 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

Rank by throughput rather than spec sheet

Ranking by tokens per second follows memory bandwidth rather than core counts, for SPHINX (Llama 2 13B). It will not match a gaming ordering, because generation is bound by memory bandwidth. The card topping the list is B200, at 170 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 SPHINX (Llama 2 13B). 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 SPHINX (Llama 2 13B).

Answers

SPHINX (Llama 2 13B) — common questions

SPHINX (Llama 2 13B)— 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 Q3_K_M, using about 10.4 GB and generating roughly 52.4 tokens per second. The fit is tight.

SPHINX (Llama 2 13B)— 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 Q4_K_M, using about 12.7 GB and generating roughly 55.5 tokens per second. The fit is tight.

SPHINX (Llama 2 13B)— 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 Q6_K, using about 17.4 GB and generating roughly 41.4 tokens per second. The fit is comfortable.

SPHINX (Llama 2 13B)— 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.

SPHINX (Llama 2 13B)— how many parameters does it have?

It has a parameter count of 19.9B. SPHINX + Llama 2 13B SPHINX component involves four vision encoders: - CLIP - ViT - CLIP - ConvNeXt V2 (89M to 659M params, depending on size) - DinoV2 - ViT (22M to 1.14B params, depending on size) - Q-former (188M params) Also involves to projection networks Huggingface Hub model files appear to be 39.8GB. Assuming models are stored in fp16 there are 2 bytes per parameter, so 39.8 / 2 = 19.9B 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.

SPHINX (Llama 2 13B)— who created it?

It was published by Shanghai AI Lab,Chinese University of Hong Kong (CUHK),ShanghaiTech University, based in China, an organisation categorised as academia,Academia,Academia.

SPHINX (Llama 2 13B)— when was it released?

It was published in November 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.

SPHINX (Llama 2 13B)— what is it used for?

It works in the domain of Vision, Language, Multimodal, and is recorded as handling the task of visual question answering, Image captioning. Models frequently carry more than one of each, and the tags describe purpose rather than capability limits.

SPHINX (Llama 2 13B)— where can I download it?

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

SPHINX (Llama 2 13B)— how much compute was used to train it?

Training consumed around 3 × 10²² FLOP, on hardware recorded as NVIDIA A100 SXM4 40 GB. 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.

SPHINX (Llama 2 13B)— 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 2.8 GB. Every figure here assumes the whole model is resident on the card.

SPHINX (Llama 2 13B)— 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: 293. So a second card is rarely the answer here.

SPHINX (Llama 2 13B)— 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.

SPHINX (Llama 2 13B)— how accurate are these speed estimates?

These are estimates with real error bars, and any of them could reasonably land anywhere in its published range depending on which runtime you use. The fastest result here: 102–272 tok/s on B200. The same model and card vary by thirty to fifty per cent depending on the inference software and its version.

SPHINX (Llama 2 13B)— what GPU do I need to run it?

The smallest card in our catalogue that holds it is Quadro K6000, with a memory capacity of 12 GB. It runs the model at a compression of Q3_K_M using about 10.4 GB, and produces roughly 14.1 tokens per second. The number of cards able to run it in total: 293.

SPHINX (Llama 2 13B)— how fast is it on a GPU?

It depends on the card. The quickest we calculate is B200, at about 170 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: 248.

SPHINX (Llama 2 13B)— how much VRAM does it need?

It needs about 10.4 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.

Source

Original publication

Record last updated 11 February 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