Discussion: GPUs vs. FPGAs on ML/DL?
I'm the editor of TechTalks (and an ML practitioner). A while ago, I was pitched an idea about the failures of GPUs in machine learning systems. The key arguments are:
1- GPUs quickly break down under environmental factors
2- Have a very short life span
3- produce a lot of heat and require extra electricity to cool down
4- maintenance, repair, replacement is a nightmare
All of these factors make it difficult to use GPUs in use cases where AI is deployed at the edge (SDCs, surveillance cameras, smart farming, etc.)
Meanwhile, all of these problems are solved in FPGAs. They're rugged, they produce less heat, require less energy, and have a longer lifespan.
In general the reasoning is sound, but coming from an FPGA vendor, I took it with a grain of salt. Does anyone have experience in using FPGA in production use cases of ML/DL? How does it compare to GPUs in the above areas?
Thanks