How good RTX 3090 is for Machine Learning, AI and Video Rendering tasks full review and comparison with RTX 3060

0 links_and_resources.txt

video link : https://www.youtube.com/watch?v=lgP1LNnaUaQ

ffmpeg with cuda support : https://github.com/GyanD/codexffmpeg/releases
check ffmpeg cuda support : ffmpeg -hwaccels
ffmpeg -version

used whisper commit : b5851c6c40e753606765ac45b85b298e3ae9e00d

used automatic1111 web ui commit : 22bcc7be428c94e9408f589966c2040187245d81

used dreambooth extension commit : 926ae204ef5de17efca2059c334b6098492a0641

whisper "C:\rtx 3090 review\whisper_test.mp3" --model large-v1 --language en --initial_prompt "Welcome to the Software Engineering Courses channel." --best_of 10 --beam_size 10 --output_dir "C:\rtx 3090 review\result" --device cuda:1

cpu-z : https://www.cpuid.com/softwares/cpu-z.html

gpu-z : https://www.techpowerup.com/gpuz/

core temp : https://www.alcpu.com/CoreTemp/

nvidia driver download : https://www.nvidia.com/Download/index.aspx?lang=en-us

uninstall older torch for any case before install: pip uninstall torch

pytorch 2 : pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

pytorch 1.13 : pip3 install torch==1.13 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu117

xformers compile wheels : https://github.com/facebookresearch/xformers/actions/workflows/wheels.yml

xformers for pytorch 1.13 and python 3.10 : pip install https://huggingface.co/MonsterMMORPG/SECourses/resolve/main/xformers-0.0.18.dev494-cp310-cp310-win_amd64.whl

latest cuda download (if requires login make a new account it is free) : https://developer.nvidia.com/cuda-downloads

latest cudnn download (if requires login make a new account it is free) : https://developer.nvidia.com/rdp/cudnn-download

set CUDA_VISIBLE_DEVICES=1

for full scripts check below files

1 cudaVersion.py

import torch
from torch.utils.cpp_extension import CUDA_HOME

print(f"CUDA Version: {torch.version.cuda}")
print(f"Torch Version: {torch.__version__}")

cudnn_version = torch.backends.cudnn.version()
print(f"The version of cuDNN DLL file used by PyTorch is {cudnn_version}.")

2 pciExpressSpeed.py

# pip install py3nvml

import py3nvml.py3nvml as nvml

def get_gpu_info():
    nvml.nvmlInit()
    device_count = nvml.nvmlDeviceGetCount()

    for i in range(device_count):
        device_handle = nvml.nvmlDeviceGetHandleByIndex(i)
        pci_info = nvml.nvmlDeviceGetPciInfo(device_handle)

        # Get GPU name
        gpu_name = nvml.nvmlDeviceGetName(device_handle)

        # Get GPU memory amount
        memory_info = nvml.nvmlDeviceGetMemoryInfo(device_handle)
        memory_amount = memory_info.total / (1024**2)  # Convert bytes to MiB

        # Get currently active PCIe Generation
        try:
            pcie_gen = nvml.nvmlDeviceGetCurrPcieLinkGeneration(device_handle)
        except nvml.NVMLError as error:
            print(f"Error getting currently active PCIe generation for device {i}: {error}")
            pcie_gen = "Unknown"

        # Get currently active PCIe Link Width
        try:
            pcie_width = nvml.nvmlDeviceGetCurrPcieLinkWidth(device_handle)
        except nvml.NVMLError as error:
            print(f"Error getting currently active PCIe link width for device {i}: {error}")
            pcie_width = "Unknown"

        print(f"GPU {i}:")
        print(f"  Name: {gpu_name}")
        print(f"  Memory Amount: {memory_amount:.2f} MiB")
        print(f"  PCI Bus ID: {pci_info.busId.decode('utf-8')}")
        print(f"  Currently Active PCIe Generation: {pcie_gen}")
        print(f"  Currently Active PCIe Link Width: {pcie_width} lanes")

    nvml.nvmlShutdown()

if __name__ == "__main__":
    get_gpu_info()

3 ffmpeg command and explanation

set CUDA_VISIBLE_DEVICES=1   // we are using this to disable first gpu in that particular cmd window so that ffmpeg uses second card
ffmpeg -hwaccel cuda -hwaccel_output_format cuda -i "C:\\rtx 3090 review\\encode_test.mp4" -c:v hevc_nvenc -preset slow -rc vbr -b:v 100m -rc-lookahead 60 -maxrate 200m -bufsize 300m -vf "scale_cuda=w=7680:h=4320" -c:a copy "C:\\rtx 3090 review\\encode_test_upscaled.mp4"

This is a command for FFmpeg, a popular open-source software suite for handling multimedia files. The command is used to transcode a video file, and it utilizes the NVIDIA hardware acceleration features for improved performance. Here's a breakdown of the parameters:

-hwaccel cuda: This parameter enables hardware acceleration using NVIDIA's CUDA technology. This offloads video processing tasks from the CPU to the GPU, resulting in faster processing.

-hwaccel_output_format cuda: This specifies the output format for the hardware-accelerated decoding to be CUDA.

-i "C:\\rtx 3090 review\\encode_test.mp4": This indicates the input file, which is a video file located in the "C:\rtx 3090 review" folder and named "encode_test.mp4".

-c:v hevc_nvenc: This sets the video codec to use NVIDIA's hardware-accelerated HEVC (H.265) encoder, hevc_nvenc.

-preset slow: This sets the encoding preset for the HEVC encoder to "slow". This provides a better compression ratio at the expense of encoding speed. Other available presets include: ultrafast, superfast, veryfast, faster, fast, medium, and veryslow.

-rc vbr: This sets the rate control mode to variable bitrate (VBR), which allows the bitrate to vary depending on the complexity of the video frames. This can result in better overall video quality compared to constant bitrate (CBR).

-b:v 100m: This sets the target video bitrate to 100 Mbps (100m).

-rc-lookahead 64: This sets the number of frames to be analyzed before encoding, allowing the encoder to make better decisions on bitrate allocation. A higher value can improve quality but may increase encoding time.

-maxrate 200m: This sets the maximum video bitrate to 200 Mbps (200m).

-bufsize 300m: This sets the buffer size for rate control to 300 Mbps (300m).

-vf "scale_cuda=w=7680:h=4320": This applies a video filter (-vf) to scale the video resolution to 7680x4320 using the CUDA-based scaler.

-c:a copy: This copies the audio stream from the input file to the output file without re-encoding it.

"C:\\rtx 3090 review\\encode_test_upscaled.mp4": This specifies the output file, which will be saved in the "C:\rtx 3090 review" folder and named "encode_test_upscaled.mp4".

Other options you may consider:

-threads: This parameter sets the number of threads used for encoding. For example, -threads 8 would use 8 threads.

-crf: This parameter sets the Constant Rate Factor (CRF), which controls the trade-off between quality and file size. Lower values yield better quality but larger files.

-c:v libx264 or -c:v libx265: These options use the software-based H.264 and H.265 encoders, respectively. These may be slower than hardware-accelerated encoders but can provide better quality in some cases.

-c:a aac: This sets the audio codec to AAC for encoding the audio stream. You can also set the audio bitrate with -b:a, for example, -b:a 128k for 128 kbps.

-ss and -t: These options can be used to trim the video. -ss sets the start time, and -t sets the duration. For example, `-ss 00:01:00 -t 00:02:00` would trim the video to start at 1 minute and last for 2 minutes.

-vf "fps=30": This applies a video filter to change the output frame rate. In this example, the output video will have a frame rate of 30 frames per second.

-af "volume=2": This applies an audio filter to adjust the volume. In this example, the output audio will be twice as loud as the input.

-vf "transpose=1": This applies a video filter to rotate the video by 90 degrees clockwise.

-vf "crop=w=1920:h=1080:x=0:y=0": This applies a video filter to crop the video frame to the specified width, height, and position (x and y).

-map 0:v:0 -map 0:a:0: These options can be used to select specific video and audio streams from the input file. In this example, the first video stream (0:v:0) and first audio stream (0:a:0) are selected.

These are just a few examples of the many options available in FFmpeg. For a comprehensive list of options and their explanations, refer to the official FFmpeg documentation: https://ffmpeg.org/ffmpeg.html.