WSL 3 is not a clean break from its predecessors so much as an architectural pivot. WSL 1 relied on syscall translation to run Linux binaries, while WSL 2 switched to a lightweight managed virtual machine (VM) that runs a real Linux kernel. WSL 3 retains the familiar wsl shell experience. Additionally, you'll be able to run Linux containers under WSL 3.

The important change in WSL 3 is how Linux processes communicate with the underlying hardware, with an emphasis on minimizing the layers between the Linux user space and Windows devices.

Microsoft promised that WSL 3 will bring you "GPU and NPU without the performance tax." Under WSL 2, Microsoft added GPU support and improved graphics integration, but those features still operated through the VM. The resulting context-switching and translation overhead meant you had to deal with slowdowns when you needed GPUs and NPUs. That approach was OK for many developer workflows, but it seriously limited your performance when working with AI workloads, even if your PC came with dedicated NPUs.

WSL 3 improves performance by implementing a new paravirtualized VM approach. Paravirtualization provides both a VM and access to the native hardware. According to Microsoft, "This means Linux containers will now run directly on Windows, with no additional configuration required." This approach sounds good to me.