Google's Tensor Chip Stomps Intel, Nvidia, And Amd

http://www.tomshardware.com/news/google-tpu-comparison-haswell-k80,34069.html

This actually comes as no surprise to anyone who knows the subject and merely illustrates how, rather than the technology driving the industry, it is increasingly driven by money, supply lines, and the military-industrial complex. Google's tensor chips are a straightforward engineering approach to saving time, energy, and money that they can leverage while the military and industries fight it out amongst themselves for control of the money, supply lines, and technology. To put this in perspective, others have already developed chips that are over a thousand times faster than anything on the market and make Google's chips look positively slow and inefficient with the theoretical upper limit being well over a million times faster and more efficient.

What Intel, AMD, and Nvidia are doing in the meantime is pounding out the next generation scalar non-Von Neumann architecture that will eventually replace virtually every computer on the planet. In the last seventeen years the industry has gone from single core to multicore processing and what the next basic architecture requires is combining the cpu, gpu, and a neuromorphic chip all on the same SoC, or system on a chip, where they can all share the same underlying architecture that allows them to share memory, processing, and bandwidth in the fastest and most efficient manner possible. No matter what kind of hardware someone produces, that same architecture should be the fastest and most efficient and, eventually, should even be adapted for quantum computing as well. Many assume a quantum computer can do all the number crunching anyone might need, but they are only useful for specific problems of the Traveling Salesman variety and will have to be combined with classical computers as well on the same chip for the highest speeds and efficiency possible but, first, we require the classical scalar architecture that is compatible with a quantum mechanical one.

AMD's new high bandwidth controller or high bandwidth memory is perhaps the best example to date being able to cut the amount of memory required for doing anything in half merely by moving the data around less and processing it on the spot whenever reasonable. This is more like the way the human brain works with each synapse functioning as both memory and processing because processing anything also serves the same function as forming a memory.