Though dedicated silicon like Google's TPU will always post higher _numbers_, GPUs have the ecosystem, the developers, the libraries, the language (languages? CUDA is probably the only one that matters)...
That's a lot to catch up to for all the competitors listed in the article.
Yes, yes, TensorFlow is cross-platform. History has shown that a cross-platform library or standard only serves to raise the barrier to entry for small startups. The cross-platform library/standard has never defined the cutting-edge or next-gen market makers. See: Posix, PDF, U2F, OpenCL/OpenGL, Intel HDA.
And yes, there are exceptions to the rule: USB, PCIe (and SATA, SDXC, etc.), WiFi.
Though dedicated silicon like Google's TPU will always post higher _numbers_, GPUs have the ecosystem, the developers, the libraries, the language (languages? CUDA is probably the only one that matters)...
That's a lot to catch up to for all the competitors listed in the article.
Yes, yes, TensorFlow is cross-platform. History has shown that a cross-platform library or standard only serves to raise the barrier to entry for small startups. The cross-platform library/standard has never defined the cutting-edge or next-gen market makers. See: Posix, PDF, U2F, OpenCL/OpenGL, Intel HDA.
And yes, there are exceptions to the rule: USB, PCIe (and SATA, SDXC, etc.), WiFi.