Performance of llama.cpp on Apple Silicon A-series

[ggerganov]

Summary

🟥 - benchmark data missing
🟨 - benchmark data partial
✅ - benchmark data available

• PP means "prompt processing" (bs = 512), TG means "text-generation" (bs = 1)

TinyLlama 1.1B

	CPU Cores	GPU Cores	F16 PP [t/s]	F16 TG [t/s]	Q8_0 PP [t/s]	Q8_0 TG [t/s]	Q4_0 PP [t/s]	Q4_0 TG [t/s]
✅ A14 1	2+4	4	251.98	10.26	250.54	24.11	242.37	39.21
🟥 A15 2	2+3	5
✅ A15 2	2+4	4	428.48	13.63	411.16	24.12	405.30	39.03
✅ A15 2	2+4	5	531.03	13.66	494.18	23.84	496.49	39.09
✅ A16 3	2+4	5	565.68	20.06	511.30	34.30	505.52	54.24
✅ A17 4	2+4	6	683.95	20.23	637.14	35.60	646.06	56.86
✅ A18 5	2+4	6	682.50	23.84	655.27	42.11	647.54	69.77
✅ A19 6	2+4	5	1443.33	29.98	1640.39	53.92	1811.19	87.33
🟥 A19 6	2+4	6

Phi-2 2.7B

	CPU Cores	GPU Cores	Q8_0 PP [t/s]	Q8_0 TG [t/s]	Q4_0 PP [t/s]	Q4_0 TG [t/s]
✅ A14 1	2+4	4	X	X	51.39	8.52
🟥 A15 2	2+3	5
🟥 A15 2	2+4	4
✅ A15 2	2+4	5	X	X	120.47	16.73
✅ A16 3	2+4	5	119.58	14.06	121.64	23.31
✅ A17 4	2+4	6	158.03	14.74	157.33	24.71
✅ A18 5	2+4	6	180.04	16.93	-	-
✅ A19 6	2+4	5	597.27	21.42	690.04	36.01
🟥 A19 6	2+4	6

Mistral 7B

	CPU Cores	GPU Cores	Q4_0 PP [t/s]	Q4_0 TG [t/s]
✅ A14 1	2+4	4	X	X
🟥 A15 2	2+3	5
🟥 A15 2	2+4	4
✅ A15 2	2+4	5	X	X
🟥 A16 3	2+4	5
✅ A17 4	2+4	6	80.55	9.01
✅ A18 5	2+4	6	94.43	11.64
✅ A19 6	2+4	5	258.84	12.74
🟥 A19 6	2+4	6

---

Description

This is a collection of short llama.cpp benchmarks on various Apple Silicon hardware. It can be useful to compare the performance that llama.cpp achieves across the A-series chips. Similar collection for the M-series is available here: #4167

Comparison of A-series chips

Instructions

• Clone the project

  git clone https://github.com/ggerganov/llama.cpp
  git checkout 0e18b2e

• Open the examples/llama.swiftui with Xcode
• Enable Release build
  
• Deploy on your iPhone / iPad
• Stop Xcode and run the app from the device. This is important because the performance when running through Xcode is significantly slower
• Download the models and run the "Bench" for each one
  
• Running the "Bench" a second time can give more accurate results
• Copy the results in the comments below, adding information about the device

---

iPhone 13 mini ✅

model	size	params	backend	test	t/s
llama 1B Q8_0	1.09 GiB	1.10 B	Metal	pp 512	411.16 ± 6.22
llama 1B Q8_0	1.09 GiB	1.10 B	Metal	tg 128	24.12 ± 0.04
llama 1B Q4_0	0.59 GiB	1.10 B	Metal	pp 512	405.30 ± 7.26
llama 1B Q4_0	0.59 GiB	1.10 B	Metal	tg 128	39.03 ± 0.08

Footnotes

1. https://en.wikipedia.org/wiki/Apple_A14 ↩ ↩² ↩³

2. https://en.wikipedia.org/wiki/Apple_A15 ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹

3. https://en.wikipedia.org/wiki/Apple_A16 ↩ ↩² ↩³

4. https://en.wikipedia.org/wiki/Apple_A17 ↩ ↩² ↩³

5. https://en.wikipedia.org/wiki/Apple_A18 ↩ ↩² ↩³

6. https://en.wikipedia.org/wiki/Apple_A19 ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

[jhen0409]

iPhone 15 Pro (A17 Pro) ✅

model	size	params	backend	test	t/s
llama 1B F16	2.05 GiB	1.10 B	Metal	pp 512	683.95 ± 8.24
llama 1B F16	2.05 GiB	1.10 B	Metal	tg 128	20.23 ± 0.08
llama 1B Q8_0	1.09 GiB	1.10 B	Metal	pp 512	637.14 ± 18.73
llama 1B Q8_0	1.09 GiB	1.10 B	Metal	tg 128	35.60 ± 0.25
llama 1B Q4_0	0.59 GiB	1.10 B	Metal	pp 512	646.06 ± 17.17
llama 1B Q4_0	0.59 GiB	1.10 B	Metal	tg 128	56.86 ± 0.17
phi2 3B Q8_0	2.75 GiB	2.78 B	Metal	pp 512	158.03 ± 14.03
phi2 3B Q8_0	2.75 GiB	2.78 B	Metal	tg 128	14.74 ± 0.07
phi2 3B Q4_0	1.49 GiB	2.78 B	Metal	pp 512	157.33 ± 14.25
phi2 3B Q4_0	1.49 GiB	2.78 B	Metal	tg 128	24.71 ± 0.04
llama 7B Q4_0	3.83 GiB	7.24 B	Metal	pp 512	80.55 ± 21.88
llama 7B Q4_0	3.83 GiB	7.24 B	Metal	tg 128	9.01 ± 0.50

[ggerganov]

The models above are now available in the app

[jhen0409]

Added phi-2 and f16 tinyllama results. Also checked the prev results are not changed in commit 0e18b2e.

[Dampfinchen]

Text Generation speed using Mistral is more than useable on newer iPhones it seems. Prompt processing is very slow however, even when using Metal. I wonder if this is a compute or bandwidth limitation.

[shouryan01]

What's the minimum t/s for it to be usable? For example, is 9 t/s usable for Mistral 7b?

[rhematt]

It's a UX design principle. For a model to be usable, the benchmark I've been working towards is 400ms for the first token and subsequent tokens. If a token isn't being processed to the user at least every 400ms, then the model won't be usable. Minimum usability should, therefore, be about 3t/s...

[ymcui]

iPhone 15 Pro Max (A17 Pro) ✅

model	size	params	backend	test	t/s
llama 1B F16	2.05 GiB	1.10 B	Metal	pp 512	652.70 ± 18.14
llama 1B F16	2.05 GiB	1.10 B	Metal	tg 128	19.82 ± 0.30
llama 1B Q8_0	1.09 GiB	1.10 B	Metal	pp 512	662.89 ± 11.28
llama 1B Q8_0	1.09 GiB	1.10 B	Metal	tg 128	34.95 ± 0.08
llama 1B Q4_0	0.59 GiB	1.10 B	Metal	pp 512	645.78 ± 9.16
llama 1B Q4_0	0.59 GiB	1.10 B	Metal	tg 128	54.95 ± 0.14

Tested under iOS 17.3 Developer beta 1 (21D5026f)

[ggerganov]

Thanks - the iPhone 15 Pro Max should be using the same A17 chip as iPhone 15 Pro, correct? At least this is what I get from wikipedia and the numbers seem to mostly match the one from @jhen0409 above

[ymcui]

Yes. Both iPhone 15 Pro and Pro Max use A17. Concretely, it is named with A17 Pro by Apple. See here.

[ymcui]

@ggerganov Just updated F16 results. The model is taken from https://huggingface.co/SergiusFlavius/TinyLlama-1.1B-1T-OpenOrca-GGUF/blob/main/tinyllama-1.1b-1t-openorca.F16.gguf

[ymcui]

iPhone 12 mini (A14) ✅

tinyllama:

model	size	params	backend	test	t/s
llama 1B F16	2.05 GiB	1.10 B	Metal	pp 512	251.98 ± 5.15
llama 1B F16	2.05 GiB	1.10 B	Metal	tg 128	10.26 ± 4.23
llama 1B Q8_0	1.09 GiB	1.10 B	Metal	pp 512	250.54 ± 0.95
llama 1B Q8_0	1.09 GiB	1.10 B	Metal	tg 128	24.11 ± 0.02
llama 1B Q4_0	0.59 GiB	1.10 B	Metal	pp 512	242.37 ± 0.81
llama 1B Q4_0	0.59 GiB	1.10 B	Metal	tg 128	39.21 ± 0.25

phi-2:

model	size	params	backend	test	t/s
phi2 3B Q8_0	2.75 GiB	2.78 B	Metal	pp 512	CRASHED
phi2 3B Q8_0	2.75 GiB	2.78 B	Metal	tg 128	CRASHED
phi2 3B Q4_0	1.49 GiB	2.78 B	Metal	pp 512	51.39 ± 11.89
phi2 3B Q4_0	1.49 GiB	2.78 B	Metal	tg 128	8.52 ± 2.78

1. phi-2 3B Q8_0 (2.75 GiB) cannot be loaded. The phone gets restarted when loading it.
2. I won't test Mistral-7B-Q4_0 (3.8 GiB) on iPhone 12 mini either, because it's too large to fit in memory (4 GiB). iPhone 12 Pro & Pro Max may have a chance to run it, as they have 6 GB RAM.

Tested under iOS 17.1.2 (21B101)

[ymcui]

Add phi-2 3B Q4_0 results. Others can't be loaded.

[ymcui]

Some additional info with memory and relevant devices.

	CPU Cores	GPU Cores	Memory [GB]	Devices
A14	2+4	4	4-6	iPhone 12 (all variants), iPad Air (4th gen), iPad (10th gen)
A15	2+3	5	4	Apple TV 4K (3rd gen)
A15	2+4	4	4	iPhone SE (3rd gen), iPhone 13 & Mini
A15	2+4	5	4-6	iPad Mini (6th gen), iPhone 13 Pro & Pro Max, iPhone 14 & Plus
A16	2+4	5	6	iPhone 14 Pro & Pro Max, iPhone 15 & Plus
A17 Pro	2+4	6	8	iPhone 15 Pro & Pro Max

[ymcui]

Apple TV 4K (3rd gen) seems to be only one that has A15 (5CPU + 5GPU).
I checked my Apple TV 4K, and unfortunately, it is 2nd gen (A12) 😂

[ggerganov]

Damn, first LLM on a TV 😄

[nikolay-kapustin]

if you really need this, i can build and run these tests for aTV 3gen )

[ggerganov]

Don't really need it, but it might be a cool achievement. LLM on a watch has already been demonstrated: https://twitter.com/shxf0072/status/1736713832045982040, but I don't think that is the case for LLM on a TV

[nikolay-kapustin]

iPhone 13 Pro (A15) ✅

model	size	params	backend	test	t/s
llama 1B Q4_0	0.59 GiB	1.10 B	Metal	pp 512	496.49 ± 3.82
llama 1B Q4_0	0.59 GiB	1.10 B	Metal	tg 128	39.09 ± 0.12
llama 1B Q8_0	1.09 GiB	1.10 B	Metal	pp 512	494.18 ± 4.93
llama 1B Q8_0	1.09 GiB	1.10 B	Metal	tg 128	23.84 ± 0.05
llama 1B F16	2.05 GiB	1.10 B	Metal	pp 512	531.03 ± 5.96
llama 1B F16	2.05 GiB	1.10 B	Metal	tg 128	13.66 ± 0.02
phi2 3B Q4_0	1.49 GiB	2.78 B	Metal	pp 512	120.47 ± 1.44
phi2 3B Q4_0	1.49 GiB	2.78 B	Metal	tg 128	16.73 ± 0.02

also a model phi-2 3B Q8_0 cannot be loaded.
and Mistral-7B-Q4_0 (3.8 GiB) not to fit in memory

[Krish120003]

iPhone 14 Pro (A16) ✅

model	size	params	backend	test	t/s
llama 1B Q4_0	0.59 GiB	1.10 B	Metal	pp 512	505.52 ± 0.58
llama 1B Q4_0	0.59 GiB	1.10 B	Metal	tg 128	54.24 ± 0.04
llama 1B Q8_0	1.09 GiB	1.10 B	Metal	pp 512	511.30 ± 1.00
llama 1B Q8_0	1.09 GiB	1.10 B	Metal	tg 128	34.30 ± 0.13
llama 1B F16	2.05 GiB	1.10 B	Metal	pp 512	565.68 ± 0.21
llama 1B F16	2.05 GiB	1.10 B	Metal	tg 128	20.06 ± 0.04
phi2 3B Q4_0	1.49 GiB	2.78 B	Metal	pp 512	121.64 ± 0.01
phi2 3B Q4_0	1.49 GiB	2.78 B	Metal	tg 128	23.31 ± 0.05
phi2 3B Q8_0	2.75 GiB	2.78 B	Metal	pp 512	119.58 ± 0.05
phi2 3B Q8_0	2.75 GiB	2.78 B	Metal	tg 128	14.06 ± 0.14

[ggerganov]

Thanks! I guess Mistral 7B does not fit on this device?

[Krish120003]

@ggerganov It does, I can load it and run it by sending a message, but the Bench button keeps aborting indicating heat up time being too long.

[Pablo-Merino]

iPhone 12 (A14) 🟨

model	size	params	backend	test	t/s
llama 1B Q4_0	0.59 GiB	1.10 B	Metal	pp 512	227.46 ± 14.55
llama 1B Q4_0	0.59 GiB	1.10 B	Metal	tg 128	37.89 ± 0.27
llama 1B Q8_0	1.09 GiB	1.10 B	Metal	pp 512	224.22 ± 22.57
llama 1B Q8_0	1.09 GiB	1.10 B	Metal	tg 128	23.23 ± 0.12

• llama 1B F16 has a long heat up time, so it aborted (28,16 seconds)
• phi2 3B Q4_0 has a long heat up time, so it aborted (9,44 seconds)
• phi2 Q8_0 crashes the app
• Not testing the Mistral model since it's too large for the device's RAM (3.8GB model vs 4GB device)

[anchorbob]

can anyone tell me what does the output metric (t/s) mean? tokens per second or what?

[XiongjieDai]

It's tokens per second.

[anchorbob]

thanks for confirmation

[xsailor511]

Can anyone tell me about llama 1b download link? I can't find it on HF or not sure which is.

[lawyinking]

https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v0.3-GGUF/tree/main

[cosmo3769]

Hi, I was trying to load starcoderbase-3b-GGUF. It is not getting loaded in iphone 15 pro simulator. It is stuck with Loading model....
When investigating, I encountered one warning message: Publishing changes from background threads is not allowed; make sure to publish values from the main thread (via operators like receive(on:)) on model updates. What could be the cause of this? Thank you.

[cosmo3769]

The above model got loaded on my device (iphone 13, ios 17.3). But when I try to send a message or benchmark it, I get Heat up time is too long message. The same message I am getting in iphone 15 (simulator), iphone 15 pro (simulator), and iphone 15 pro max (simulator) as well. The size of the model is 2.05 GB. @ggerganov

[beebopkim]

iPhone SE (3rd Generation), A15 2+4 CPU, 4 GPU, 4 GB of RAM ✅

model	size	params	backend	test	t/s
llama 1B F16	2.05 GiB	1.10 B	Metal	pp 512	428.48 ± 1.24
llama 1B F16	2.05 GiB	1.10 B	Metal	tg 128	13.63 ± 0.03
llama 1B Q8_0	1.09 GiB	1.10 B	Metal	pp 512	141.71 ± 59.55
llama 1B Q8_0	1.09 GiB	1.10 B	Metal	tg 128	15.03 ± 0.87
llama 1B Q4_0	0.59 GiB	1.10 B	Metal	pp 512	152.40 ± 53.64
llama 1B Q4_0	0.59 GiB	1.10 B	Metal	tg 128	18.37 ± 1.67
phi2 3B Q8_0	2.75 GiB	2.78 B	Metal	pp 512	Model loaded but benchmark failed because llama.swift was killed
phi2 3B Q8_0	2.75 GiB	2.78 B	Metal	tg 128	Model loaded but benchmark failed because llama.swift was killed
phi2 3B Q4_0	1.49 GiB	2.78 B	Metal	pp 512	29.73 ± 5.09
phi2 3B Q4_0	1.49 GiB	2.78 B	Metal	tg 128	7.08 ± 2.08
llama 7B Q4_0	3.83 GiB	7.24 B	Metal	pp 512	Looks like model was loaded but benchmark failed with Heat up time is too long
llama 7B Q4_0	3.83 GiB	7.24 B	Metal	tg 128	Looks like model was loaded but benchmark failed with Heat up time is too long

[aneeshmb02]

What data/prompts are used for this?

[eakarsu]

I have run llamma.cpp on ios device (iphone) described here. But models are giving garbage response. what am I doing wrong?

[kinchahoy]

Would it be possible to update these instructions for a recent version of XCode? I get a simple error that I can't quite figure out:
"/Users/USER/inference/llama.cpp/air-lld:1:1 81 duplicated symbols for target 'air64_v23-apple-ios14.0.0-simulator"

[Beingpax]

Same error as kinchahoy. Getting the same error.

[sienaiwun]

Hi, I was trying to deploy llama.swiftUI on IOS 15.
I found "
:1:9: note: in file included from :1:
#import "llama.h"
^
/Users/naiwenxie/Documents/androidBUld/llama.cpp/Sources/llama/llama.h:3:10: error: 'llama.h' file not found with include; use "quotes" instead
#include <llama.h>
^
/Users/naiwenxie/Documents/androidBUld/llama.cpp/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift:2:8: error: could not build Objective-C module 'llama'
import llama
^
/Users/naiwenxie/Documents/androidBUld/llama.cpp/examples/llama.swiftui/:1:9: In file included from :1:
" this error.
Does someone have any idea about that?

[poom3d]

Did you checkout code from 0e18b2e?
Did you enable release build? IIRC, I got some errors if not on release build.

[sienaiwun]

Thanks I'm on the latest master branch as of February 2, and the issue still occurs. I opened the project directly with Xcode, configured the signing settings, and the issue appeared—all settings are at their defaults. Are there any additional steps or guidance?

[sienaiwun]

I was able to run it, thank you. The instructions for running it are here: #4508

[NegiAbhijeet]

why master branch's example/llama.swiftui not supporting simulator, like Building for 'iOS-simulator', but linking in dylib (/opt/homebrew/Cellar/llama.cpp/4733/lib/libggml.dylib) built for 'macOS'

i want to use qwen2 which is not supported in 0e18b2e branch, is there any solution

[matthewnaples]

why is it that we need to use the "Release" build?

Thanks

[HaohongLin1]

Hi, I'm new to the llama cpp. The llama.swiftui works well in my local machine by using the default downloadable models, but it failed to start after I copied the quatanized meta llama 3.2 1B, which is created by using HuggingFace https://huggingface.co/spaces/ggml-org/gguf-my-repo, to /examples/llama.swiftui/llama.swiftui/Resources/models.

It showed errors from console:

error loading model: create_tensor: tensor 'output.weight' not found
llama_load_model_from_file: failed to load model
Could not load model at /Users/xxx/Library/Developer/CoreSimulator/Devices/5EEF15B1-B91D-422B-AC6D-27039A036350/data/Containers/Bundle/Application/ACA68B26-D133-49E7-9341-D8E5F32CB63E/llama.swiftui.app/models/llama-3.2-1b-q4_k_m.gguf
Error while loading the model: The operation couldn’t be completed. (llama_swiftui.LlamaError error 0.)

Would you please point out what went wrong?

[tuan88it]

@HaohongLin1 : Please guide me through the steps to build the LLAMA.cpp project. I can’t run LLAMA.swiftUI because I keep getting the error “llama.h not found.”

[HaohongLin1]

@tuan88it

OS: MacOS Sequoia 15.1.1

For the LLAMA.swiftUI, I only follow Instructions in this page, I don't need any extra steps to build it in XCode.

[guanyuhong]

@HaohongLin1 : Please guide me through the steps to build the LLAMA.cpp project. I can’t run LLAMA.swiftUI because I keep getting the error “llama.h not found.”

git checkout 0e18b2e I can run the code , but I still face the error "error loading model: create_tensor: tensor 'output.weight' not found" 😭

[poom3d]

After checkout I run the code just fine for both simulator and on device. Did you tap download link below first? (It takes a while to download some model).

[poom3d]

iPhone 16 pro, A18 Pro 2+4CPU, 6GPU, 8 GB of RAM ✅

The first thing to do after I got my new phone is to run llama benchmark 👍

model	size	params	backend	test	t/s
llama 1B Q4_0	0.59 GiB	1.10 B	Metal	pp 512	647.54 土 2.09
llama 1B Q4_0	0.59 GiB	1.10 B	Metal	tg 128	69.77 土 0.12
llama 1B Q8_0	1.09 GiB	1.10 B	Metal	pp 512	655.27 土 1.90
llama 1B Q8_0	1.09 GiB	1.10 B	Metal	tg 128	42.11 土 0.03
llama 1B F16	2.05 GiB	1.10 B	Metal	pp 512	682.50 土 1.19
llama 1B F16	2.05 GiB	1.10 B	Metal	tg 128	23.84 土 0.04
phi2 3B Q8_0	2.75 GiB	2.78 B	Metal	pp 512	180.04 土 4.55
phi2 3B Q8_0	2.75 GiB	2.78 B	Metal	tg 128	16.93 土 0.02
llama 7B Q4_0	3.83 GiB	7.24 B	Metal	pp 512	94.43 土 13.98
llama 7B Q4_0	3.83 GiB	7.24 B	Metal	tg 128	11.64 土 1.07

[donohara]

Thanks for your work on this! I'm a definite noob but love the fact that I can study the code and read all the comments on this repo. Exciting that the attitude is 'let's see what we can do, and how to make it smaller/faster/better!

[sienaiwun]

I’d like to ask how you built the iOS target. Does the memory information show up when running the benchmark? When I tested on Android, I found that the memory usage (PSS value) was more than twice the size of the model.

[poom3d]

I just followed the instruction on this thread. For memory, it's a spec of the phone I'm running with.
The result os copied from running the benchmarks.

[jhen0409]

iPhone Air (A19 Pro, 5c GPU, 12GB RAM) ✅

Commit: ece0f5c

model	size	params	backend	test	t/s
llama 1B Q4_0	0.59 GiB	1.10 B	Metal	pp 512	1811.19 ± 4.26
llama 1B Q4_0	0.59 GiB	1.10 B	Metal	tg 128	87.33 ± 2.03
llama 1B Q8_0	1.09 GiB	1.10 B	Metal	pp 512	1640.39 ± 17.86
llama 1B Q8_0	1.09 GiB	1.10 B	Metal	tg 128	53.92 ± 0.15
llama 1B F16	2.05 GiB	1.10 B	Metal	pp 512	1443.33 ± 24.19
llama 1B F16	2.05 GiB	1.10 B	Metal	tg 128	29.98 ± 0.18
phi2 3B Q4_0	1.49 GiB	2.78 B	Metal	pp 512	690.04 ± 6.13
phi2 3B Q4_0	1.49 GiB	2.78 B	Metal	tg 128	36.01 ± 0.07
phi2 3B Q8_0	2.75 GiB	2.78 B	Metal	pp 512	597.27 ± 26.71
phi2 3B Q8_0	2.75 GiB	2.78 B	Metal	tg 128	21.42 ± 0.05
llama 7B Q4_0	3.83 GiB	7.24 B	Metal	pp 512	258.84 ± 23.36
llama 7B Q4_0	3.83 GiB	7.24 B	Metal	tg 128	12.74 ± 2.74
llama 7B Q3_K - Medium	3.28 GiB	7.24 B	Metal	pp 512	183.59 ± 16.45
llama 7B Q3_K - Medium	3.28 GiB	7.24 B	Metal	tg 128	14.93 ± 1.24

[jhen0409]

For has_tensor = false (Disabled Metal 4 API usage):

Click to expand

model	size	params	backend	test	t/s
llama 1B Q4_0	0.59 GiB	1.10 B	Metal	pp 512	819.90 ± 9.79
llama 1B Q4_0	0.59 GiB	1.10 B	Metal	tg 128	88.96 ± 0.13
llama 1B Q8_0	1.09 GiB	1.10 B	Metal	pp 512	820.27 ± 5.62
llama 1B Q8_0	1.09 GiB	1.10 B	Metal	tg 128	54.78 ± 0.15
llama 1B F16	2.05 GiB	1.10 B	Metal	pp 512	860.57 ± 6.20
llama 1B F16	2.05 GiB	1.10 B	Metal	tg 128	30.33 ± 0.02
phi2 3B Q4_0	1.49 GiB	2.78 B	Metal	pp 512	305.28 ± 12.79
phi2 3B Q4_0	1.49 GiB	2.78 B	Metal	tg 128	36.42 ± 0.13
phi2 3B Q8_0	2.75 GiB	2.78 B	Metal	pp 512	291.93 ± 27.23
phi2 3B Q8_0	2.75 GiB	2.78 B	Metal	tg 128	19.53 ± 3.40
llama 7B Q4_0	3.83 GiB	7.24 B	Metal	pp 512	100.20 ± 30.94
llama 7B Q4_0	3.83 GiB	7.24 B	Metal	tg 128	12.12 ± 2.78
llama 7B Q3_K - Medium	3.28 GiB	7.24 B	Metal	pp 512	87.45 ± 24.82
llama 7B Q3_K - Medium	3.28 GiB	7.24 B	Metal	tg 128	10.43 ± 4.31

[ggerganov]

Thanks - updated the data.

Would be great to get also get benches for the 6-core A19 Pro version.

[nikolay-kapustin]

Ping me here at the end of day i will try

[avesa95]

Can somebody explain, step by step, how to build and deploy a LLaMA model—starting with iPhone 12 and up to the latest devices?
I’m very curious how these models can be integrated into production apps: maintaining context, and using an optimized KV cache.

[rsergiuistoc]

+1

[andreea-irina]

+1

[rebecabasca20]

+1

[ggerganov]

There is an example in this very discussion - does it not work for you?