[alexey-salmin]

It definitely had its downsides, but holding TOP1 for 6 lists (3 years straight) was an achievement. Biggest issues weren't in engineering IMO.

GPUs also have a weird programming model and yet here we are. I think in the end what mattered the most was the strategic failure to address the low-end market with Phi. When the right time came everyone did CUDA because everyone already had a GPU -- basically the same reason why x86 won the server market against SPARC decades ago.

In the meanwhile came the 2015 US export ban, then loss of interest by the management right before matrix multiplication stopped being an HPC problem and came into every segment in the form of ML.

Based on what we know now probably the best strategy was to bet everything on Intel Graphics and leverage the widespread of built-in graphics while it was dominated by Intel. From there it was possible to eat Nvidia's lunch in hi-end and HPC too. However in 2010 it wasn't certain at all. No one was talking about AI, the buzzword of the time was "big data" which relied on conventional computing methods. Deep learning revolution didn't happen yet, even GBDTs weren't a thing, ML was about linear regression. MPP architectures were confined to physics simulation and 3d graphics (which is a physics simulation of a sort).