PXL Vision has acquired its first supercomputer - unleashing its AI computing potential!

26, October 2021

Identity Verification Liveness Detection Machine Learning Face Verification

PXL Vision is very excited about its purchase of an NVIDIA DGX A100 supercomputer, which will provide its data scientists and machine learning (ML) engineers with much needed compute cycles for their work. This increased compute capability will allow PXL Vision to continue innovating on its areas of technological expertise; namely machine learning and computer vision.

The Nvidia DGX A100 is a beast, it is a veritable data center in a box – a box so jam packed with circuitry that it takes two to carry (weighing in at 123 kilograms). At peak performance, it draws 7000 watts of power, which could easily heat a medium-sized apartment.

Nvidia specs

Source: NVIDIA DGX A100 promotional material (pdf)

The Nvidia line of supercomputers are by no means cheap and such an acquisition requires a company to carefully consider its needs, as the desire is to maximize the returns from the investment. This is not a problem for PXL Vision’s extended team of computer engineers, who are now able to complete compute processes, that used to take months - in a matter of days. This speeds up development cycles by a large degree enabling rapid model training and rigorous machine learning development.

Nvidia’s DGX A100 has 2 CPUs and 8 GPUs and can compute 5 Petaflops; or 5 thousand trillion operations per second (more on flops below), providing PXL’s engineers an abundance of compute cycles to work with and permits them to run numerous experiments at once.

Nvidia just installed     Nvidia just mounted

2 of PXL’s employees celebrating the just mounted NVIDIA DGX A100

What exactly is a supercomputer?

Britannica’s online dictionary defines a supercomputer as “any of a class of extremely powerful computers. The term is commonly applied to the fastest high-performance systems available at any given time.” So, the supercomputers of yesteryear would be referred to as normal computers today. According to Samsung Insights, the smartphones that most of us have in our pockets today are 1000 times more powerful than the Cray-2 supercomputer from the 1980s.

The Cray-2 followed on the footsteps of earlier Cray computers, which are considered to be the first supercomputers (given their huge performance edge over other computers at the time). The man behind these earliest supercomputers, Seymour Cray, also known as the father of supercomputing, built the first supercomputer in 1964, which was able to execute 3 million floating-point operations per second (FLOPS). Compare this to Nvidia’s DGX A100, which can compute an astonishing 5000 trillion FLOPS, which is around 1 600 000 times faster than the first supercomputer.

The compute speed of today’s supercomputers is constantly increasing. There is even an arms race - mostly between China and the United States - to build the fastest computer; however Japan is the current winner with its Fugaku supercomputer, which achieved a top speed of 442 petaflops.

The ability to connect numerous supercomputers in parallel into clusters is how these supercomputers are able to attain such high throughput. The Nvidia DGX A100 is also capable of clustering, meaning that PXL Vision will be able to continually scale its compute potential for the foreseeable future.

All this flippin’ talk about flops…

The standard performance measurement for supercomputers are FLOPS, which is an acronym for FLoating point OPerations per Second. The standard metric system prefixes of kilo, mega, giga, tera, peta… etc. are applied to flops. And one petaflop is equal to one thousand trillion operations per second. Let that sink in for a minute. It is actually really difficult to fathom how much a trillion is, so some comparisons are useful.

The New York Times has published a handy graph that compares a million to a trillion using a scrollable measuring bar. The article focuses on the United States national debt as it approaches 30 trillion dollars. An unmistakably huge amount of money, but I digress here a little; one petaflop is equal to 1000 trillion and that is a much bigger number.

Another way to look at it is with a stopwatch timer that counts in seconds. Consider that travelling back in time one trillion seconds would drop you off at around 30 000 B.C.; given that one trillion seconds of ordinary clock time = (1012 sec)/( 3.16 x 107 sec/year) = 31 546 years!

Why does PXL Vision need a supercomputer anyway?

Partial to machine learning and computer vision technologies, PXL Vision works in the data economy. Data is required to train machines to learn. Fortunately, with the emergence of high-performance computing, we now have systems that are capable of handling staggering amounts of data in processing times that seemed unimaginable only a few years ago.

To better understand this, one needs to look no further than the complexity of the human brain. Our brains have been evolving for millions of years in order to make sense of the otherwise messy world around us, and as a result our brains have become incredibly complex organs, containing an estimated 86 billion neurons.

Take the human sense of sight, for instance, which relates to the field of computer vision. The human retina and brain can detect edges really well, and from these edges our brains can build lines, then shapes, then forms and finally something that resembles a face. This is not a coincidence of nature but rather something that evolved over millions of years.

To mimic sight, machine learning computers must learn from data in order to “see” – much as our brains did through evolution. Machine learning algorithms have to be fed copious amounts of data and once that data is computed and returned, ML engineers make minor adjustments based on the desired results and then input the data again, until the parameters of the experiment have been fulfilled. To put it succinctly, machine learning engineers are putting the neurons into the neural networks of machines.

Needless to say that the amount of data computed in these processes is massive and this is precisely why PXL Vision requires a supercomputer. With this acquisition, PXL has joined the league of other companies working in artificial intelligence and has future-proofed itself towards training the machines of the 4th Industrial Revolution.

Interested in all things machine learning? Apply today at PXL Vision

Access to an onsite supercomputer has done wonders for our recruitment strategy. Join our team today! You can check out our Careers page here if you are interested!

Don’t miss the latest news, trends and insights in digital identity

Related insights

Machine Learning and How It Applies to Facial Recognition Technology

Read more

Related posts

Guide: How fake accounts hurt the sharing economy & how to prevent them

Read more

The Swiss Method: Innovative use of facial biometrics challenges global identity verification players

Read more

Machine Learning and How It Applies to Facial Recognition Technology

Read more

Form title