07-29-2017, 01:40 PM
In the last five years, three technology breakthroughs NVDA refers to as the "Big Bang of AI" have transformed certain AI technologies - in particular vision - from impractical to commercially viable. Investors are broadly aware of two of these breakthroughs - massively parallel processing (provided by NVDA and others), and massive data storage (provided by webscale companies). We highlight an under-appreciated third breakthrough of the Big Bang - improved algorithms - that emerged in 2012 and 2015. Stein goes through a whole history of algorithms: "beginning in 2006 researchers developed three major breakthroughs in regards to the algorithms used in deep learning: Restricted Boltzmann Machines (RMBs), Deep Belief Networks (DBNs), and Convolutional Neural Networks (CNNs).” He goes into each in detail. He comes back at the end of the report to point out Nvidia is the “key semiconductor and development tool player” in A.I.
Nvidia: ‘We’re Not Early,’ but, ‘It’s Not Too Late,’ Says SunTrust - Barron's
Volta is the new GPU architecture Nvidia revealed earlier this year. The new chips were promised to be such an improvement over current models that shares of the company jumped 17.8% in a single day after their announcement. AI research requires training a computer program to be as efficient as possible before it works well. This training requires multiplying matrices of data, which normally would have to be done single numbers at a time. The new Volta GPU architecture is able to multiply entire rows and columns of matrices data at once, rapidly speeding up the AI training process. Nvidia claims the new Volta architecture is 12 times faster at processing matrix multiplication than its previous "Pascal" architecture. It reduces the duration of an AI training task that used to take 18 hours to 7.4 hours, according to company data. Nvidia gave away 15 of its Volta-based Tesla V100 chips to top researchers attending the conference. The chips were some of the first ones available outside of the company, and were signed by CEO Jensen Huang.