Skip to content

INSTALL_CUDA_11.8_STEP_BY_STEP.md

Latest commit

 

History

History
468 lines (334 loc) · 11.1 KB

INSTALL_CUDA_11.8_STEP_BY_STEP.md

File metadata and controls

468 lines (334 loc) · 11.1 KB
Raw

CUDA 11.8 Installation - Step-by-Step Guide

Goal: Install CUDA 11.8 + cuDNN 8.6 untuk enable TensorFlow GPU di Eaglearn

Current Status:

  • ✅ GPU: RTX 3050 Laptop (4GB VRAM)
  • ✅ Driver: 581.29 (latest)
  • ⚠️ CUDA: 13.0 (need 11.8 for TensorFlow)

Estimated Time: 45 minutes

📥 STEP 1: Download Files (15 minutes)

A. CUDA 11.8.0 Toolkit

Link: https://developer.nvidia.com/cuda-11-8-0-download-archive

Steps:

  1. Click link di atas
  2. Select:
    • Operating System: Windows
    • Architecture: x86_64
    • Version: 10 atau 11
    • Installer Type: exe (network) ← Recommended (smaller)
  3. Click Download
  4. File: cuda_11.8.0_522.06_windows_network.exe (~3 MB)
  5. Save to: D:\Downloads\ (atau folder downloads kamu)

Alternative (jika internet lambat):

  • Pilih exe (local) = ~3 GB (full offline installer)

B. cuDNN 8.6.0 for CUDA 11.x

Link: https://developer.nvidia.com/cudnn

⚠️ PENTING: Butuh NVIDIA Developer Account (gratis)

Steps:

  1. Click link di atas
  2. Click Download cuDNN
  3. Login atau Create Account (gratis, 2 menit)
  4. Accept License Agreement
  5. Click Download cuDNN v8.6.0 (October 3rd, 2022), for CUDA 11.x
  6. Click Local Installer for Windows (Zip)
  7. File: cudnn-windows-x86_64-8.6.0.163_cuda11-archive.zip (~600 MB)
  8. Save to: D:\Downloads\

Wait for both downloads to complete before continuing!

🗑️ STEP 2: Uninstall CUDA 13.0 (5 minutes)

⚠️ Note: CUDA runtime 13.0 yang terdeteksi di nvidia-smi adalah driver runtime, bukan toolkit. Kita perlu check apakah CUDA Toolkit 13.0 terinstall.

Check if CUDA Toolkit Installed:

# Open Command Prompt (Win+R, ketik "cmd")
nvcc --version

If you see:

nvcc: NVIDIA (R) Cuda compiler driver
...release 13.0...

→ CUDA Toolkit 13.0 IS installed, need to uninstall

If you see:

'nvcc' is not recognized...

→ No CUDA Toolkit installed, skip to STEP 3 ✅

If CUDA 13.0 Toolkit Found, Uninstall:

Method A: Control Panel (Recommended)

  1. Open: Control Panel → Programs and Features
  2. Search for "CUDA" in list
  3. Uninstall ALL of these (if present):
    • ❌ NVIDIA CUDA Runtime 13.0
    • ❌ NVIDIA CUDA Development 13.0
    • ❌ NVIDIA CUDA Documentation 13.0
    • ❌ NVIDIA CUDA Samples 13.0
    • ❌ NVIDIA CUDA Visual Studio Integration 13.0
    • ✅ Keep: NVIDIA Graphics Driver (jangan uninstall!)
  4. Click each → Uninstall → Follow prompts
  5. Restart computer (recommended)

Method B: Manual Cleanup (Optional, if uninstall fails)

# Open Command Prompt as Administrator (Win+X → Command Prompt (Admin))
rmdir /s /q "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.0"

📦 STEP 3: Install CUDA 11.8 (10 minutes)

Prerequisites: Both files downloaded

Installation Steps:

  1. Run installer:

    • Navigate to D:\Downloads\
    • Right-click → Run as Administrator
    • File: cuda_11.8.0_522.06_windows_network.exe

  2. NVIDIA Installer will open:

    • Click OK to extract
    • Wait for extraction (~1 minute)

  3. License Agreement:

    • Read (or skip 😄)
    • Click Agree and Continue

  4. Installation Options:

    • CRITICAL: Select Custom (Advanced)
    • Click Next

  5. Component Selection (IMPORTANT!):

    UNCHECK these (already have newer versions):

    • ❌ Driver components (your 581.29 is newer!)
    • ❌ NVIDIA GeForce Experience (already installed)
    • ❌ PhysX (not needed)
    • ❌ NVIDIA Nsight (not needed)

    CHECK these (needed for TensorFlow):

    • ✅ CUDA Toolkit 11.8
    • ✅ CUDA Samples 11.8 (optional, useful for testing)
    • ✅ CUDA Documentation 11.8 (optional)
    • ✅ CUDA Visual Studio Integration (if you use Visual Studio)

  6. Installation Location:

    • Default is fine: C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8
    • Click Next

  7. Install:

    • Click Next
    • Wait for installation (~5-10 minutes)
    • ☕ Take a coffee break!

  8. Completion:

    • Click Close
    • DO NOT restart yet (wait after cuDNN install)

📦 STEP 4: Install cuDNN 8.6 (5 minutes)

Prerequisites: CUDA 11.8 installed, cuDNN zip downloaded

Installation Steps:

  1. Extract cuDNN zip:

    • Navigate to D:\Downloads\
    • Right-click cudnn-windows-x86_64-8.6.0.163_cuda11-archive.zip
    • Click Extract All...
    • Extract to: D:\Downloads\cudnn\
    • Click Extract

  2. Verify extracted folder structure:

    D:\Downloads\cudnn\cudnn-windows-x86_64-8.6.0.163_cuda11-archive\
├── bin\
│   ├── cudnn64_8.dll
│   ├── cudnn_ops_infer64_8.dll
│   └── cudnn_cnn_infer64_8.dll
├── include\
│   └── cudnn.h
└── lib\
    └── x64\
        └── cudnn.lib

  3. Copy files to CUDA directory:

    Option A: Manual Copy (Easiest)

    Open two File Explorer windows:

    Window 1 (Source):

    • D:\Downloads\cudnn\cudnn-windows-x86_64-8.6.0.163_cuda11-archive\

    Window 2 (Destination):

    • C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8\

    Copy these folders (drag & drop):

    • Drag bin contents → merge with CUDA bin folder
    • Drag include contents → merge with CUDA include folder
    • Drag lib\x64 contents → merge with CUDA lib\x64 folder

    Click Replace if prompted

    Option B: Command Line (Advanced)

    Open Command Prompt as Administrator:

    cd D:\Downloads\cudnn\cudnn-windows-x86_64-8.6.0.163_cuda11-archive

xcopy bin\*.* "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8\bin\" /Y
xcopy include\*.* "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8\include\" /Y
xcopy lib\x64\*.* "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8\lib\x64\" /Y

  4. Verify cuDNN files copied:

    Check these files exist:

    • C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8\bin\cudnn64_8.dll ✅
    • C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8\include\cudnn.h ✅
    • C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8\lib\x64\cudnn.lib ✅

🔄 STEP 5: Restart Computer (IMPORTANT!)

Why restart?

  • Load new CUDA drivers
  • Update system PATH
  • Initialize GPU with new toolkit

Steps:

  1. Save all work
  2. Close all applications
  3. Restart computer
  4. Come back after restart! ←

✅ STEP 6: Verify Installation (5 minutes)

After restart, open NEW Command Prompt:

A. Check CUDA Version

nvcc --version

Expected Output:

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2022 NVIDIA Corporation
Built on Wed_Sep_21_10:41:10_Pacific_Daylight_Time_2022
Cuda compilation tools, release 11.8, V11.8.89
Build cuda_11.8.r11.8/compiler.31833905_0

✅ If you see "release 11.8" → SUCCESS! ❌ If you see error or different version → Something wrong

B. Check nvidia-smi

nvidia-smi

Expected Output:

CUDA Version: 11.8 (or higher like 13.0 is OK - this is driver version)

✅ GPU should still show → SUCCESS!

C. Check Environment Variables

echo %CUDA_PATH%

Expected:

C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8

✅ Path contains "v11.8" → SUCCESS!

D. Check PATH Variable

echo %PATH%

Expected (should contain):

...CUDA\v11.8\bin;...CUDA\v11.8\libnvvp;...

✅ If you see "CUDA\v11.8" paths → SUCCESS!

🧪 STEP 7: Test TensorFlow GPU (5 minutes)

Navigate to Eaglearn project:

cd D:\Eaglearn-Project
python test_gpu.py

Expected Output (SUCCESS):

============================================================
TensorFlow GPU Detection Test
============================================================

📦 TensorFlow Version: 2.15.0

🔍 Searching for GPU devices...
✅ GPU DETECTED: 1 device(s) found!

GPU 0:
  Name: /physical_device:GPU:0
  Type: GPU

✅ GPU memory growth enabled successfully!
✅ GPU computation successful!

📚 Library Versions:
  CUDA: 11.8
  cuDNN: 8.6

============================================================
✅ GPU IS READY FOR EAGLEARN!
============================================================

✅ If you see this → CUDA 11.8 installation SUCCESS! 🎉

🚀 STEP 8: Test Eaglearn App (5 minutes)

cd D:\Eaglearn-Project
python app.py

Watch startup logs for:

🚀 TensorFlow GPU detected: 1 device(s)
✅ GPU memory growth enabled for: /physical_device:GPU:0
🚀 Using RetinaFace backend (TensorFlow GPU accelerated)
🔧 Backend: retinaface | TensorFlow GPU: True
🔧 Confidence Threshold: 0.20

✅ If you see GPU detected → Eaglearn GPU WORKING! 🎉

Open browser: http://localhost:8080

Test emotion detection performance!

🎯 Expected Performance After Installation

Metric Before (CPU) After (GPU) Improvement
Backend SSD RetinaFace Better accuracy
Accuracy ~85% ~95% +10%
FPS 10-15 20-25 +67% speed
Latency ~100ms ~50ms 2x faster
Confidence 0.25 0.20 More sensitive
GPU Usage 0% 30-50% Utilized

🐛 Troubleshooting

Problem: "nvcc: command not found"

Solution:

  1. Restart Command Prompt (or computer)
  2. Check PATH environment variable
  3. Manually add to PATH if needed:
    • System Properties → Environment Variables
    • Edit "Path" variable
    • Add: C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8\bin

Problem: "Could not load dynamic library 'cudnn64_8.dll'"

Solution:

  1. Verify cuDNN files in CUDA\v11.8\bin\
  2. Check file: cudnn64_8.dll exists
  3. Restart computer
  4. If still fails, re-copy cuDNN files

Problem: TensorFlow still not detecting GPU

Solution:

  1. Reinstall TensorFlow: 
    pip uninstall tensorflow tf-keras
pip install tensorflow==2.15.0 tf-keras==2.15.0
  2. Restart Command Prompt
  3. Run python test_gpu.py again

Problem: Installation fails with error

Solution:

  1. Disable antivirus temporarily
  2. Run installer as Administrator
  3. Check disk space (need ~5GB free on C:)
  4. Close NVIDIA processes in Task Manager
  5. Try again

📞 Need Help?

Check logs:

  • TensorFlow: python test_gpu.py
  • nvidia-smi: Shows GPU status
  • nvcc --version: Shows CUDA version

Reference docs:

  • Full guide: docs/CUDA_INSTALLATION_GUIDE.md
  • GPU optimization: docs/GPU_OPTIMIZATION.md

✅ Installation Checklist

Before continuing, verify ALL these:

  • [ ] CUDA 11.8 downloaded
  • [ ] cuDNN 8.6 downloaded
  • [ ] CUDA 13.0 uninstalled (if was installed)
  • [ ] CUDA 11.8 installed (Custom, without driver)
  • [ ] cuDNN files copied to CUDA directory
  • [ ] Computer restarted
  • [ ] nvcc --version shows 11.8
  • [ ] test_gpu.py detects GPU
  • [ ] Eaglearn app runs with GPU

All checked? Congratulations! 🎉

Ready to start? Let's do this! 🚀

Current Step: Download CUDA 11.8 from link above!