How much performance does 'GeForce RTX 4070 SUPER' with 12GB of VRAM capacity demonstrate in image generation AI? A thorough comparison review with the previous generation Grabo
'GeForce RTX 4070 SUPER' released on January 17, 2024 is equipped with 12GB of VRAM, and can comfortably operate generation-based AI such as image generation AI. Since I was able to borrow '
GeForce RTX 4070 Family Graphics Card | NVIDIA
https://www.nvidia.com/ja-jp/geforce/graphics-cards/40-series/rtx-4070-family/
GeForce RTX 4070 SUPER is a performance-enhanced graphics board of GeForce RTX 4070, which consumes less power than the previous generation's higher-end model, GeForce RTX 3090, and has the same or higher processing performance. The following article provides a detailed summary of the performance comparison between GeForce RTX 4070 SUPER and GeForce RTX 3090, as well as heat and noise during high loads.
Thoroughly compare the processing performance of 'GeForce RTX 4070 SUPER' and 'GeForce RTX 3090' with various benchmarks & check all the heat distribution, fan noise, and power consumption A comprehensive review - GIGAZINE
This time we will introduce 'GeForce RTX 4070 SUPER Founders Edition (hereinafter referred to as RTX 4070 SUPER)', 'ASUS Dual GeForce RTX 3060 V2 OC Edition 12GB GDDR6 (hereinafter referred to as RTX 3060 12GB)', and 'GeForce RTX 3090 GD3090-24GEBST (hereinafter referred to as RTX 3090)'. Let's compare the generation speed when running the image generation AI. The VRAM capacity is 12GB for RTX 4070 SUPER, 12GB for RTX 3060 12GB, and 24GB for RTX 3090.
The PC configuration other than the graphics board is as follows. I am using the press version 511.15 of the GPU driver.
| CPU | AMD Ryzen 5 7600X BOX |
| motherboard | X670E Taichi Carrara |
| memory | Crucial CT16G48C40U5 PC5-38400 (DDR5-4800) 288pin DDR5 UDIMM 16GB |
| storage | WD_BLACK SN750 NVMe |
| OS | Windows 11 Pro 64bit |
◆Check the image generation speed with Stable Diffusion
We introduced NVIDIA's official acceleration extension ' TensorRT Extension for Stable Diffusion ' to the Stable Diffusion web UI (AUTOMATIC1111 version) and measured the generation speed when 100 images were generated using three types of graphic boards. .
First, I set the model to 'Stable Diffusion 1.5', the sampler to 'Euler a', the number of steps to 20, and the resolution to 512 x 512 pixels, and generated 100 images. The prompt is 'girl using laptop, cyberpunk, high quality, anime style' and the negative prompt is 'wrong, low quality.'
With Stable Diffusion, you can check the processing time from the console, so let's compare this time.
The image generation time of RTX 4070 SUPER, RTX 3060 12GB, and RTX 3090 is summarized in the table below. The RTX 4070 SUPER finished generating in half the time of the RTX 3060 12GB. It is also worth noting that it can generate images in a shorter time than the previous generation top model RTX 3090.
| GPU | Generation time |
|---|---|
| RTX 4070 SUPER | 1 minute 28 seconds |
| RTX 3060 12GB | 2 minutes 57 seconds |
| RTX 3090 | 1 minute 31 seconds |
Next, I changed the model to 'Stable Diffusion 2.1', set the sampler to 'Euler a', the number of steps to 20, and the resolution to 512 x 512 pixels, and generated 100 images.
The image generation time of RTX 4070 SUPER, RTX 3060 12GB, and RTX 3090 is summarized in the table below. Even after changing the model, the RTX 4070 SUPER was able to generate images the fastest.
| GPU | Generation time |
|---|---|
| RTX 4070 SUPER | 1 minute 32 seconds |
| RTX 3060 12GB | 3 minutes 25 seconds |
| RTX 3090 | 1 minute 47 seconds |
◆Measuring the generation rate per second with StreamDiffusion
Using `` StreamDiffusion ,'' which has become a hot topic for its ability to generate ultra-high-speed images, I tried to see how many images can be generated per second. Below is how StreamDiffusion's 'optimal-performance/single.py' is executed to generate an image on a PC equipped with RTX 4070 SUPER. You can see that images can be generated fairly quickly.
I tried running the explosive image generation AI 'StreamDiffusion' on RTX 4070 SUPER - YouTube
At the bottom of the screen where the image is displayed, the number of images generated per second (fps) is displayed. In the case of RTX 4070 SUPER, it was around 40fps.
Also, the average fps will be displayed on the console after image generation is complete. The average fps for RTX 4070 SUPER was 40.57132355074917.
The average fps of RTX 4070 SUPER, RTX 3060 12GB, and RTX 3090 are summarized in the table below. Although the RTX 4070 SUPER was able to generate images faster than the RTX 3060 12GB, the RTX 3090 was the fastest of the three graphics boards.
| GPU | average fps |
|---|---|
| RTX 4070 SUPER | 40.57132355074917 |
| RTX 3060 12GB | 19.826912350032618 |
| RTX 3090 | 45.25673838451412 |
Next, I tried running 'optimal-performance/multi.py', which performs batch processing optimized for RTX 4090. You can check the execution state in the movie below.
Running the explosive image generation AI 'StreamDiffusion' on RTX 4070 SUPER and generating more than 60 images per second - YouTube
Below is a table summarizing the average fps of RTX 4070 SUPER, RTX 3060 12GB, and RTX 3090. Although the RTX 4070 SUPER was not as good as the RTX 3090, it was able to generate images at an astonishing speed of over 66 images per second.
| GPU | average fps |
|---|---|
| RTX 4070 SUPER | 66.84877768893824 |
| RTX 3060 12GB | 16.57194484320879 |
| RTX 3090 | 73.78827724014783 |
◆Tell us what you would like us to review about “GeForce RTX 4070 SUPER”!
Please use the link below to tell us the points you would like us to try because you will actually be using it, such as ``Try to see what happens when you do this!'' ``Can you do this?'' ``What will happen in this situation?'' It's okay if the content overlaps with someone else's, but the more you have, the more helpful it will be for you to understand that this is what you're concerned about. And it will be reflected in your next review article creation!
• Discord | “Tell us what you would like us to review about “GeForce RTX 4070 SUPER”! ' | GIGAZINE
Related Posts:
