3 ms·
Almost every programming model is Turing complete therefore you can do just about anything with them. But unless OpenGL shader language has equivalent concepts
by varelse 5y ago
Almost every programming model is Turing complete therefore you can do just about anything with them. But unless OpenGL shader language has equivalent concepts of the warp and the syncthreads instruction, you end up having to do a global sync every time you want to do any synchronization and that means returning to the CPU and that's really slow in comparison and therefore very power inefficient as well.
- zamadatix 5y agoWhich is fine in this class of device (i.e. no access to proper compute shaders) as long as it's faster or more power efficient than doing it any other way on the device.
- varelse 5y agoThat's going to be very task dependent but probably okay for matrix multiplication and streaming kernels which is most of AI.