Next-Gen Computing Tech Digest - March 2017

Google TPU

Google has revealed study information on its first machine learning chip or Tensor Processing Unit (TPU). Google has been using the TPU for machine learning applications in its data centres since 2015. Google has published a list of findings:
On AI workloads that use neural network inference the TPU is 15x to 30x faster than GPUs and CPUs. The TPU is 30x to 80x more energy efficient than conventional chips. Neural networks powering these machine learning applications require a small amount of code: just 100 to 1500 lines. 
Google credited the TPUs as being instrumental in the victory of AlphaGo, improvements to Google Translate service, Google Image Search, Photos and Cloud Vision API. 


France, Germany, Italy, Luxembourg, the Netherlands, Portugal and Spain have signed the EuroHPC declaration to deploy exascale computers – capable of a billion billion calculations per second. The aim of the EuroHPC project is to have two exascale computers by 2020 with fully exascale performance by 2023. Once functional the computers will be available across the EU for use by public and private institutes and communities.  

Brain based computing

Researchers at Stanford University and Sandia National Laboratories, in the USA, have created an organic electrode that mimics a brain synapse’s learning process. The artificial synapse was programmed (taught) by repeatedly charging and discharging it. The artificial synapse showed lower power consumption that traditional computing storage and processing systems as it can process and store information in one procedure. Furthermore, digital transistors can only be in one of two states (1 or 0) whereas the scientists claim their organic electrode can be programmed for 500 states. The scientists see this as a step towards neural-model computing that works on visual and audio input (that might be used for example in voice controlled devices and driverless cars). The team plan to create an array of artificial synapses for further testing and research.

DARPA explores molecular potential

American agency DARPA (Defense Advanced Research Projects Agency) has announced its Molecular Informatics programs that will explore approaches to store and process information in molecules. The project brief says that the right molecule holds promise as an enabler of dense data representations and highly versatile computing concepts that move beyond digital computing approaches. 

Writing to magnetic atoms

Scientists at IBM have created a single atom magnet that can act as a storage device. The device stored one bit of data on the magnet. This compares favourably with other hard drives which might use as many as 100,000 atoms to store the same amount of information.  IBM says that using this tech would enable the 35 million song library of iTunes to be stored on a credit card sized device. In the study IBM also demonstrated the reading and writing of information to two magnetic atoms spaced one nanometre apart by applying electric current.

Add this: