There are at least three technical reasons why none of these accelerators have worked: 1) systolic arrays without SIMT or general purpose multicore are not fle

There are at least three technical reasons why none of these accelerators have worked: 1) systolic arrays without SIMT or general purpose multicore are not flexible enough, creating a flawed foundation 2) everyone targeting tsmc is building the same circuits, leading to no differentiation and no motivation to move off an incumbent 3) the CUDA software stack is so big it creates porting friction So you have a flawed foundation that makes everything else harder together with no advantage and a huge porting effort No wonder why no one switches I’d like to see hardware startups take more risk and get off the beaten path in ways that enable fundamentally new algorithms When you put all of this against the backdrop of Intel management, what other outcome could there have been?