My old PC
Intel® Core™ i5-6500 CPU @ 3.20GHz × 4, MM40G, NVIDIA GeForce GTX 1650
Ubuntu 22.04.3 LTS
Notice
Incompatible versions
TensorFlow, CUDA, cuDNN, and TensorRT all need to be compatible. Make sure that I am using versions of these libraries that work together.
I could find this information in the compatibility matrix on the NVIDIA website.
procedure
0. Update NVIDIA-HDMI display driver
sudo ubuntu-drivers autoinstall
1. GTX driver download from NVIDIA website
2. Get Tensorflow and other lib-configuration information by answering questions from the NVIDIA website.
3. Install CUDA from NVIDIA's Base Installer.
4. Set PATH and LD_LIBRARY_PATH manually.
5. Run "pip install tensorflow[and-cuda
6. Install TensorRT, but it is not installed due to some dependency issues.
7. Memory is not enough at all.
sudo fallocate -l 128G /swapfile
caution
1. recommends that everybody switches tensorflow.keras in place of plain keras.
troble issue ( Insufficient operation )
1. WARNING:tensorflow:Layer lstm will not use cuDNN kernels since it doesn't meet the criteria. It will use a generic GPU kernel as fallback when running on GPU.
2. external/local_tsl/tsl/framework/bfc_allocator.cc:485] Allocator (GPU_0_bfc) ran out of memory trying to allocate 1.89GiB (rounded to 2028265984)requested by op_EagerConst
my LSTM operation
tensorflow/core/common_runtime/gpu/gpu_device.cc:1929] Created device /job:localhost/replica:0/task:0/device:GPU:0 with 2754 MB memory: -> device: 0, name: NVIDIA GeForce GTX 1650, pci bus id: 0000:01:00.0, compute capability: 7.5
machine | CPU | GPU | settting | comment | 1epoc time |
Aiit-8 | 8 x Xeon_E-2334 | non | non | 18min | |
Hp-desktop | 4 x i5_6500 | GTX1650 | non | general CPU operation | 30min |
Hp-desktop | 4 x i5_6500 | GTX1650 | CuDA | Insufficient GPU operation | 10min |