For small projects, the NVIDIA Jetson Nano is interesting as is the Coral Dev Board (Google product). These both support full Linux as well as specialized cores for all the math required. But the number of cores is fairly small - 128 in the case of the Nano.
For larger projects, I'm not aware of any company even coming close to the NVIDIA graphics cards with over 10,000 CUDA cores. That's a LOT of parallel computing.
The new $1,500 RTX 3090 has 10,496 cores, for 36 teraflops. We went to the Moon and back using machines capable of 2-3 megaflops!
You can do a lot of machine learning with 36 teraflops.
https://www.engadget.com/nvidia-rtx-3090-3080-3070-cuda-core-int32-fp32-210059544.htmlTensor cores are the new big deal - worth searching Google...