Norm Jouppi, a distinguished hardware engineer at Google, detailed the company's public disclosure of the Tensor Processing Unit (TPU) last week after the CEO Sundar Pichai's earlier announcement at Google I/O.
ASIC optimizations, reportedly favoring machine learning with TensorFlow (TF) specifically, include reduced computational precision, thereby requiring fewer transistors per operation. Performance-testing parameters or metrics aren't available at this time, but Google claimed the optimizations help increase the number of operations per second the chip can process.
Google noted that the project was started several years ago and that it was fast-forwarding current technology by about seven years but hasn't provided data for the community to analyze. Jouppi noted the time from testing a prototype of the chip to data-center deployment was 22 days and that it was an example of Google putting research into practice.
Several questions came up around how the TensorFlow-optimized chipset could compete with publicly available hardware like Nvidia's Tesla P100 and even PaaS providers like Nervana that provide machine learning services. Google's public disclosure of the TPU may have been related to Nvidia's release of the Tesla P100 in April.
Recent inquiries directed at Google covered the topic of Google designing and producing its own chips, and the potential impact to industry leaders like Intel. Jouppi noted that Google wants to lead the industry in machine learning and make the innovation available to its customers, but didn't disclose specific plans or offerings to do so at this time.
Commenters in the original post brought up the Nvidia P100 and TX1, as well as IBM's TrueNorth as potentially fungible chips to compare against the TPU but no specifics benchmarks or comparisons were provided. Google hasn't disclosed availability plans for the TPU outside of their internal use cases, which include among other things RankBrain, Street View, and the recent highly publicized AlphaGo custom hardware used in the TF-based Go game stack that beat Lee Sedol last February.
