TensorFlowJS fails reliably on Intel GPU running Windows, otherwise succeeds

[neurolabusc]

This issue is easy to replicate from the web browser by opening the niivue-brainchop live demo and choosing Extract the brain (High acc, slow) from the drop down menu.

Note that the web page crashes for machines running Intel GPUs on the Windows operating system. In contrast, the model succeeds using Intel GPUs on MacOS and Linux operating systems. Likewise, the model succeeds with Apple, NVidia and AMD GPUs regardless of operating system.

While the live demo is the easiest way to replicate this, you can also replicate the problem with a hot reloadable source code that allows you to edit JavaScript and immediately see the results.

git clone git@github.com:niivue/niivue-brainchop.git
cd niivue-brainchop
npm install
npm run dev

You can automatically run one of the failing models (model 10) by changing lines 200 and 201 to read:

     modelSelect.selectedIndex = 10
     modelSelect.onchange()

System information

• We have tested this on numerous machines, and the results are consistent across many. Note that the Intel-based Macs that succeed are older than the Intel-based Windows computers that fail. An example of one of our Windows machines that fails is a Xiaomi Book Pro 16 2022 i5-1240P 16G 512GB SSD 4K OLED Touch Screen Notebook Window 11.

[gyagp]

This looks like an Intel specific issue, so we'll investigate this.