Looking at the TPU architecture was fascinating for an old-timer like me. I looked into implementing backpropagation on a Connection Machine CM2 back in the lat

Looking at the TPU architecture was fascinating for an old-timer like me. I looked into implementing backpropagation on a Connection Machine CM2 back in the late 1980s. I got to the point of implementing the training algorithm in C* (an SIMD variant of C). However, I ultimately decided in favor of using a shared memory multi-processor. The main problem was that time on my unversity's CM2 was very hard to come by.