GPUs for People Who Just Want Their Model to Run
You don't need to write CUDA to make good decisions about GPUs. You need to understand memory, bandwidth, and why batch size is the only knob that matters.
// foundations
Memory, bandwidth, batching and precision: enough hardware to make good decisions.
You don't need to write CUDA to make good decisions about GPUs. You need to understand memory, bandwidth, and why batch size is the only knob that matters.