This page has some details of Tenstorrent's current chips [1]. Looks like a manycore simd design focused on tensor ops from a brief look. Apparently they also have some sort or compression scheme to boost memory bandwidth.
I don't notice anything in particular that stands out vs. the many other AI chips people are making, at first glance. But I'm far from an expert. There are several other technical videos on their YouTube channel as well: https://www.youtube.com/channel/UC7041p6DlAh0r4_Fnlk10pQ
I’m pretty surprised no-one has actually exposed the actor model for parallelising neural networks, it seems it would work quite well and allow you to have a layer per node (or actually many split configurations). Maybe data locality would be an issue with actor based approaches. They seem to be solving this at a lower level but with less knowledge of the actual parallelism in software.
I wonder what they are referring to. Are they accelerating what SHAP's GradientExplainer [1] does? (namely: crafting inputs at a specific layer, propagating forward to see the influence on class prediction, and sort of backpropagating to pixels) Or is it about something more related to Judea Pearl's work on causality?
[1] https://github.com/slundberg/shap#deep-learning-example-with...