Computer architects are building special-purpose chips for runtime use of trained neural networks to compensate for power and performance limits.

Judge as of 19 March 2019. True: by that date, special-purpose chips designed to run already-trained neural networks (inference) had been built, announced or deployed, with power or performance efficiency given as a reason. Examples would be Google's TPU or neural engines in phones. Mostly true: such chips existed, but they were mainly aimed at training, or the stated motive was not power or performance limits. Mostly false: such chips existed only as research proposals, with no hardware being built. False: no special-purpose inference chips were being designed or built by that date.

Set before the research began.

1. E1 Quantifying the performance of the TPU, our first machine learning chip | Google Cloud Blog Google Cloud Blog (Google) · 25 June 2018 Own record

   “custom chips, which have been running machine learning applications in our data centers since 2015.”

2. E2 Quantifying the performance of the TPU, our first machine learning chip | Google Cloud Blog Google Cloud Blog (Google) · 25 June 2018 Own record

   “This first generation of TPUs targeted inference (the use of an already trained model, as opposed to the training phase of a model, which has somewhat different characteristics)”

3. E3 Quantifying the performance of the TPU, our first machine learning chip | Google Cloud Blog Google Cloud Blog (Google) · 25 June 2018 Own record

   “The TPU also achieves much better energy efficiency than conventional chips, achieving 30x to 80x improvement in TOPS/Watt measure”

4. E4 Huawei Mate 10 Review TechSpot · 28 December 2017

   “allowing the Kirin 970 to process neural network operations at a faster rate in a highly efficient fashion.”

Google's 2018 blog says its first-generation TPU was built for inference, meaning running already-trained models (E2), and had been running in its data centers since 2015 (E1). Google also says the chip achieved 30x to 80x better energy efficiency in TOPS/Watt than conventional chips (E3). TechSpot's December 2017 review of the Huawei Mate 10 says the phone's Kirin 970 can process neural network operations at a faster rate in a highly efficient fashion (E4). Together these show that special-purpose inference chips had been built and deployed before 19 March 2019, with power and performance efficiency given as reasons, so the claim resolves true.