From reading (not viewing the videos), it appears to be an [artificial] neural net implemented in hardware. So, it's a novel, brain-like architecture for multi-core, programmed in the same way as neural nets (i.e. 'trained')

I have doubts about [artificial] neural nets, since they never really went anywhere, and I don't think scale or hardware implementation will automatically change that.

But playing with novel hardware can give novel perspectives, leading to new insights - and IBM is quite right that our present architectures seem to have stalled. That is, apart from GPGPUs.