Performance of llama.cpp on Nvidia CUDA #15013
Replies: 43 comments 30 replies
This comment has been hidden.
This comment has been hidden.
This comment has been hidden.
This comment has been hidden.
This comment has been hidden.
This comment has been hidden.
This comment has been hidden.
This comment has been hidden.
This comment has been hidden.
This comment has been hidden.
-
|
Here's the results for my devices. Not sure how to get a "cuda info string" though. CUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)
CUDA Scoreboard for Llama 2 7B, Q4_0 (with FA)
|
Beta Was this translation helpful? Give feedback.
-
|
While technically not directly related, there may also be value in comparing AMD ROCM build here too, as ROCM acts a replacement (sometimes a directly compatible layer) for most CUDA calls. I admit risk of confusion for Nvidia users in the thread if this path is taken. |
Beta Was this translation helpful? Give feedback.
-
|
As I know you cannot run ROCm on Nvidia GPU. If you would like to see compared results check Vulkan thread. You can find there results for Vulkan/CUDA and Vulkan/ROCm. UPD: Created ROCm discussion. |
Beta Was this translation helpful? Give feedback.
-
|
Device 0: NVIDIA GeForce RTX 3090 Ti, compute capability 8.6, VMM: yes
build: 9c35706 (6060) Device 0: NVIDIA GeForce RTX 3080, compute capability 8.6, VMM: yes
build: 9c35706 (6060) |
Beta Was this translation helpful? Give feedback.
-
|
Device 0: NVIDIA GeForce RTX 4070 Ti SUPER, compute capability 8.9, VMM: yes
build: 9c35706 (647) |
Beta Was this translation helpful? Give feedback.
-
|
@olegshulyakov One more benchmark for RTX 4080: Device 0: NVIDIA GeForce RTX 4080, compute capability 8.9, VMM: yes
build: 20638e4 (2) |
Beta Was this translation helpful? Give feedback.
-
|
@Ristovski why so slow? Have you undervolted it? It pretty as RTX 3080, I expected somewhere between RTX 3090 and 3080 Ti =( |
Beta Was this translation helpful? Give feedback.
-
Hmm indeed, I didn't give much thought to the score at first. It should be stock but not completely sure as that is one of our work machines. I didn't have much time to investigate today, will check again tomorrow! |
Beta Was this translation helpful? Give feedback.
-
|
Device 0: 3090. Power limit to 250w
build: 9c35706 (6060) Device 2: 5090. Power limit to 400w
build: 9c35706 (6060) |
Beta Was this translation helpful? Give feedback.
-
|
Can you please launch them without a limit on full power? |
Beta Was this translation helpful? Give feedback.
-
|
Sure, results with defaults power limits: 3090 at 390W
build: 9c35706 (6060) 5090 at 600W
build: 9c35706 (6060) |
Beta Was this translation helpful? Give feedback.
-
|
Device 0: NVIDIA GeForce GTX 1080 Ti, compute capability 6.1, VMM: yes
Device 0: NVIDIA GeForce GTX 1080 Ti, compute capability 6.1, VMM: yes
build: 9c35706 (6060) |
Beta Was this translation helpful? Give feedback.
-
|
@olegshulyakov To help users quickly understand the approximate largest models that can run on each GPU, I suggest adding a VRAM column next to the GPU name on the main scoreboard. Example:
|
Beta Was this translation helpful? Give feedback.
-
|
Made it a little bit better 🙂 |
Beta Was this translation helpful? Give feedback.
-
|
Device 0: NVIDIA GeForce RTX 2060 SUPER, compute capability 7.5, VMM: yes
build: 5c0eb5e (6075) |
Beta Was this translation helpful? Give feedback.
-
|
@ggerganov Can you please add "performance" label? |
Beta Was this translation helpful? Give feedback.
-
|
@olegshulyakov I see you grabbed some of my numbers from the Vulkan thread. However, I flooded that post with a bunch of data that probably came across as noise. While you quoted my correct numbers for Non-FA, the FA results you grabbed were actually when run on two GPUs instead of one. To make things easier, here are the numbers from a single card: RTX 5060 Ti 16 GB
And here's another GPU for the collection: RTX 4060 Ti 8 GB
|
Beta Was this translation helpful? Give feedback.
-
|
Nice 64GB VRAM setup you got there!
We all be here showing off our GPU collections 😅 |
Beta Was this translation helpful? Give feedback.
-
|
Thanks. It isn't the fastest setup around, especially when working with 70B+ models, but it is completely usable for inference. There are also some benefits I like about these particular cards (Gigabyte Windforce):
|
Beta Was this translation helpful? Give feedback.
-
|
Device 0: NVIDIA GeForce RTX 2080 Ti, compute capability 7.5, VMM: yes
build: 9c35706 (6060) |
Beta Was this translation helpful? Give feedback.
-
|
Yeah also saw numbers for my 4090 taken from the Vulkan thread. Re-ran CUDA results so you can get the latest FA and non-FA results from same build: FA:
Non-FA:
nvidia-dkms 575.64.03-1 ❯ nvcc --version |
Beta Was this translation helpful? Give feedback.
-
|
NVIDIA P106-100 I ran two times, took the best on 2 different build
build: 5fd160b (6106)
build: 860a9e4 (5688) Sadly, nvidia was not supporting this device for the vulkan driver |
Beta Was this translation helpful? Give feedback.
-
|
I just bricked my gtx 1070 Ti :( so i would not be able to reproduce the result with newer build |
Beta Was this translation helpful? Give feedback.
-
|
@pebaryan I've taken the last build one. |
Beta Was this translation helpful? Give feedback.
-
|
Would like to participate with a slightly exotic one from my cute server cube.. :-) (RTX 2000 Ada, 16GB, 75W) I did two runs:
gml_cuda_init: GGML_CUDA_FORCE_MMQ: no
build: 756cfea (6105)
ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
build: 1d72c84 (6109) Seems to make no big difference... ^^ |
Beta Was this translation helpful? Give feedback.
-
|
I finally got my hands on similar card as before (NP106) but with display output NVIDIA GTX 1060
build: 5fd160b (6106) |
Beta Was this translation helpful? Give feedback.
-
|
just realized i didn't use the latest build, not that difference though
build: 79c1160 (6123) |
Beta Was this translation helpful? Give feedback.
-
|
NVIDIA K80 recompile using visual studio 2022 and cuda toolkit 11.4
build: 32732f2 (6248) it seems in the background, the layer is split into two devices, first 2G is on Device 0 and the rest in the second Device. The utilization at some point is 75% and 25% |
Beta Was this translation helpful? Give feedback.
-
|
Re-run with |
Beta Was this translation helpful? Give feedback.
-
|
I have found possible inefficiencies in CUDA code: many Attached file shows the output of llama.cpp compiled with |
Beta Was this translation helpful? Give feedback.
-
|
ilintar@LinuksowaJaskinia:/devel/models$ llama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1
build: 9ef5369 (6256) Pretty much in line with @slaren results. |
Beta Was this translation helpful? Give feedback.
-
|
Device 0: Tesla P100-PCIE-16GB, compute capability 6.0, VMM: yes
build: 9ef53690 (6256) |
Beta Was this translation helpful? Give feedback.
-
build: 79c1160 (6123)
build: 79c1160 (6123) |
Beta Was this translation helpful? Give feedback.
-
|
I tried but I have 4 3090s and it only used 1 of them the other 3 were at 0% |
Beta Was this translation helpful? Give feedback.
-
|
NVIDIA Quadro P1000 - 4 GB GDDR5 - 128 bit
build: 1e74897 (1) You will not be able to run this benchmark if you are running GNOME or anything else that uses GPU memory, boot into text mode to run it. With that said the GPU will run smaller models (< 2.5 GB) while using the graphical environment fine, for example IBM Granite 3.3 2B, Qwen3 1.7B, etc. |
Beta Was this translation helpful? Give feedback.
-
|
4090 -pl 300 (300 watts, stock clocks) on official container:latest
4090 stock 450watts
% difference:
|
Beta Was this translation helpful? Give feedback.
-
|
RTX 5080 (16GB) Ubuntu 22.04.1 (Linux 6.8.0) Driver Version: 575.64.03 CUDA Version: 12.9
build: 8a4280c (6307) |
Beta Was this translation helpful? Give feedback.
-
|
RTX A6000 (48 GB / GDDR6 / 384-bit) Linux 5.15.0 ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
build: 4795c91 (6342) |
Beta Was this translation helpful? Give feedback.
-
|
Quadro T1000 (TU117 4GB / GDDR6 / 128-bit) , HP Z2 mini G5 I couldn't load 512 tokens with 4GB of VRAM. I tested pp256 instead. Oracle Linux 9.6
With pt13762104's GGML_CUDA_NO_TURING_MMA modification
I'm running offline batch inference on this old workstation. Llama.cpp and Qwen3-4B-Thinking-2507 IQ4 work well. Many thanks to the llama.cpp team! |
Beta Was this translation helpful? Give feedback.
-
|
A100 80GB PCIe ./build/bin/llama-bench -m llama-2-7b.Q4_0.gguf -fa 0,1 -p 512,1024,2048,4096,8192,16384,32768 -n 128,256,512,1024,2048,4096
build: 5143fa8 (6392) |
Beta Was this translation helpful? Give feedback.
-
|
A100 80GB SXM4 ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
build: 5143fa8 (6392) |
Beta Was this translation helpful? Give feedback.
-
|
H100 80GB PCIe ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
build: 5143fa8 (6392) |
Beta Was this translation helpful? Give feedback.
-
|
H100 80GB SXM5 ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
build: 5143fa8 (6392) |
Beta Was this translation helpful? Give feedback.
-
|
After seeing @Hedede's numbers with the H100, I had to try the RTX 6000 Pro Blackwell on the latest llama.cpp version to compare. Just barely manages to edge it out. :) With fa: Without fa: |
Beta Was this translation helpful? Give feedback.
and others
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
-
This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.
We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.
Instructions
Either run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using
-sm none -mg YOUR_GPU_NUMBERunless the model is too big to fit in VRAM.Share your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.
If multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!
CUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)
CUDA Scoreboard for Llama 2 7B, Q4_0 (with FA)
More detailed test
The main idea of this test is to show a decrease in performance with increasing size.
Beta Was this translation helpful? Give feedback.
All reactions