Windows8下Anaconda环境MKL线程数受限的技术咨询

I’ve dealt with similar MKL thread capping issues on older Windows versions, so let’s walk through some targeted fixes beyond the steps you’ve already tried:

1. Rule Out Windows 8 Power Plan Throttling #

Windows 8’s default power plans often restrict CPU usage to save battery, which can indirectly cap MKL’s thread count. Here’s how to fix it:

• Head to Control Panel > Hardware and Sound > Power Options
• Switch to the High Performance plan (this ensures your CPU runs at full capacity)
• Open Task Manager’s Performance tab to confirm all 4 cores are active and not stuck at low utilization

2. Create a Custom MKL Config File (Anaconda-specific) #

Anaconda sometimes overrides MKL settings behind the scenes. A dedicated config file can force it to respect your thread count:

• Navigate to your Anaconda install directory, then go to Library\etc\mkl
• Create a new text file named mkl.conf (if it doesn’t exist already)
• Paste these lines into the file:

  MKL_NUM_THREADS=4
  MKL_DYNAMIC=FALSE
  

• Save the file, close all your Python/Anaconda sessions, then restart and test with this code:

  import mkl
  mkl.set_num_threads(4)
  print(mkl.get_max_threads())  # Should now return 4
  

3. Sync PyTorch Thread Settings with MKL #

PyTorch has its own thread management that can clash with MKL. Explicitly set both to match:

import torch
import mkl

# Force MKL to use 4 threads
mkl.set_num_threads(4)
# Set PyTorch's intra-operation threads to match MKL
torch.set_num_threads(4)
# Optional: Disable inter-operation threads if you don't need multi-task parallelism
torch.set_num_interop_threads(1)

# Verify settings
print(f"MKL max threads: {mkl.get_max_threads()}")
print(f"PyTorch active threads: {torch.get_num_threads()}")

4. Downgrade MKL for Windows 8 + Python 3.6 Compatibility #

Python 3.6 is end-of-life, and newer MKL versions (2020+) have known compatibility quirks with Windows 8. Let’s roll back to a stable version:

• Open your Anaconda Prompt and run:

  conda list mkl
  

• If your MKL version is 2020 or newer, downgrade it:

  conda install mkl=2019.4 -y
  

• Reinstall a compatible PyTorch version for Python 3.6:

  conda install pytorch=1.7.1 torchvision=0.8.2 cpuonly -c pytorch -y
  

5. Check System-wide Processor Restrictions #

Windows 8 might have hidden group policies limiting CPU cores/threads:

• Press Win + R, type gpedit.msc and hit Enter to open Local Group Policy Editor
• Navigate to Computer Configuration > Administrative Templates > System > Processor Scheduling
• Double-click Limit maximum number of processors and set it to Not Configured (or explicitly set it to 4)
• Restart your PC to apply changes

内容的提问来源于stack exchange，提问作者Aviv Rubinstein