How Does STM32 Handle Mixed Precision Weight Transfers in AI Models?
Hi, I am working on AI for embedded systems and I am very curious to understand something more hardware-related. I am currently researching mixed precision models, which use different precisions for the weights. My question is about how these weights are moved within the microcontroller. If I understand correctly, each RAM register is 32 bits wide, meaning that with 8-bit representations, I can store 4 weights in one register.
My question is, when these weights are moved, does the microcontroller (STM32) transfer each bit one by one, or are all the bits moved together? I am asking this to understand the energy consumption, as I want to determine if it is based on the number of registers moved or the number of bits moved. This understanding is crucial since moving bits is one of the most power-consuming operations when running a neural network.
Solution:Jump to solution
@wafa_ath Microcontrollers like the
STM32 typically have 32-bit wide registers. This means that they can handle 32 bits of data in a single operation, when moving data internally the microcontroller generally transfers the entire 32-bit register in a single operation.2 Replies
@wafa_ath The STM32 microcontroller moves data in 32-bit chunks, not one bit at a time. So, energy consumption depends on how many memory accesses you make, not the individual bits moved.
Solution
@wafa_ath Microcontrollers like the
STM32 typically have 32-bit wide registers. This means that they can handle 32 bits of data in a single operation, when moving data internally the microcontroller generally transfers the entire 32-bit register in a single operation.