As we explained before, the recent surge of AI and its rapidly becoming a dominant discipline are partially due to the exponential degree of technological progress we faced over the last few years. What it is interesting to point out though is that AI is deeply influencing and shaping the course of technology as well.
First of all, the Graphics Processing Units (GPUs) have been adapted from traditional graphical user interface applications to alternative parallel computing operations. NVIDIA is leading this flow and is pioneering the market with the CUDA platform and the recent introduction of Telsa P100 platform (the first GPU designed for hyperscale data center applications). On top of P100, they also created the first full server appliance platform (named DGX-1), which will bring deep learning to an entirely new level. Very recently, they also released the Titan X, which is the biggest GPU ever built (3,584 CUDA cores).
In general, the most impressive developments we observed are related to chips, especially Neuromorphic Processing Units (NPUs) ideated to emulate the human brain. Specific AI-chips have been created by major incumbents: IBM has released in 2016 the TrueNorth chip, which it is claimed to work very similarly to a mammalian brain. The chip is made of 5.4 billion transistors, and it is able to simulate up 1 million neurons and 256 million neural connections. It is equipped with 4,000 cores that have 256 inputs lines (the axons) and as much output lines (neurons), which send signals only when electrical charges achieve a determined threshold.
This structure is quite similar to the Neurogrid developed by Stanford, although the academic counterpart is made of 16 different chips instead of the single one proposed by the software colossus.
Google, on the other hand, announced the introduction design of an application-specific integrated circuit (ASIC) thought and tuned specifically for neural networks — the so-called Tensor Processing Unit (TPU). The TPU optimizes the performance per watt specifically for machine learning problems, and it both powers RankBrain (i.e., Google Search) and DeepMind (i.e., AlphaGO).
Intel is working on similar chips as well, i.e., the Xeon Phi chip series, and the latest release has been named Knights Landing (KNL). KNL has up to 72 cores, and instead of being a GPU, it can be a primary CPU that reduces the need to offload machine learning to co-processors.
Even Qualcomm has invested enormous resources in the Snapdragon 820, and eventually into the deep learning SDK Snapdragon Neural Processing Engine and their Zeroth Machine Intelligence Platform.
The cost for all those chips is huge (on the order of billions for R&D, and hundred thousand dollars as selling cost), and they are not viable for retail consumers yet but only thought for enterprise applications. The main exception to this major trend is the mass-scale commercial AI chip called Eyeriss, released earlier in 2016 by a group of researchers at MIT. This chip — made of 168 processing engines — has been built on a smartphone’s power budget and thus is particularly energy-friendly, but it presents anyway computational limitations.
Even though this is a cost-intensive game, several startups and smaller companies are considerably contributing to the space: Numenta open-source NuPIC, a platform for intelligent computing, to analyze streaming data. Knowm, Inc. has brought memristor chips to the market, which is a device that can change its internal resistance based on electrical signals fed into it (and used as a non-volatile memory). KnuEdge (and its subsidiaries KnuPath) created LambaFabric, which runs on a completely innovative architecture different not only from traditional GPUs but also from TPUs. Nervana Systems released an ASIC with a new high-capacity and high-speed memory technology called High Bandwidth Memory. Horizon Robotics is another company actively working in the space, as well as krtkl, which has produced a new low-cost dual-core ARM processor (FPGA, Wi-Fi, Bluetooth) named Snickerdoodle.
A final note has to be made in favor of Movidius, which introduced a completely new concept, i.e., an all-in-one USB for deep learning. Codenamed Fathom Neural Compute Stick, it contains a chip called Myriad 2, which has been thought in partnership with Google specifically to tackle down any advanced image recognition issue (but it has been used also to power drones and robots of a diverse kind).