TPU Architecture
Does Mojo utilize external TPU hardware such as the Coral USB accelerator? If not, are there future plans?
5 Replies
That hasn't been explicitly worked on yet no, the focus is on NVIDIA gpu's right now as that's the most common hardware. But Mojo sits on top of MLIR and so will be able to optimize for all types of exotic hardware.
Will it work with the hardware if it was compiled before the hardware was installed?
If you're writing raw Mojo you'll likely need to write different implementations for CPU/GPU/TPU, and then have a compile time flag or runtime check. If you're building a framework on top of the MIN graph building API, or just using the MAX engine for inference, that complexity will be taken care of for you.
Is the MAX engine written using Mojo?
All of the kernels are written using Mojo