ASCI Red in 1997

The coming age of free supercomputing

(This is a followup post to (now) earlier posts on forecasting.

  • The first in May 2015 forecast both blimp-based and dedicated building-based drone deployments (later patented by Amazon);
  • The second in October 2015 largely predicted Elon Musk’s Tesla Masterplan Part Deux by 9 months.
  • The third in July 2016 among other things correctly hypothesised the use of Model X falcon wings for future Tesla bus designs.
  • The fourth was on SpaceX telecoms plans. We will see what happens there.
  • I try to get it mostly right but I mainly love the idle speculation).

Over twenty years ago a supercomputer called the Accelerated Strategic Computing Initiative (ASCI) Red became fully operational at Sandia National Laboratory in the United States.

It was a big beast — 104 cabinets covering 2,500 square feet (230 m2). It consisted of 6,000 200MHz Pentium Pros, 1,212 gigabytes of total distributed memory and 12.5 terabytes of disk storage. The supercomputer, built by Intel, was then ranked the number one supercomputer in the world. It cost 
$46 million at the time to build.

It was the summer of 1997 — not that long ago for many of us. We were whiling away our time on a buggy Windows 95 box and looking at the launch of the much anticipated Pentium II processor by Intel, while also awaiting the launch of Windows ‘98.

We connected to the internet through 28k dialup; Bertie Ahern became Taoiseach of Ireland in June; the first Harry Potter book was published; we watched Face/Off with Nicholas Cage and John Travolta; Radiohead released OK Computer and Puffy Daddy’s I’ll Be Missing You was riding high in the charts. We had a new movie called Men in Black to watch.

It was also back in the days before most people had mobile phones — it would be the following year before pay-as-you-go concepts started to appear fully in Ireland, with offerings from Eircell and Esat Digifone.

The tech boom was starting in earnest too. Amazon IPO’d in May 1997 — the dot com boom was well underway. If you decided to put $10,000 dollars into Amazon shares on the IPO day, you would be sitting on about $5,000,000 now. Inspired by Amazon, a small company was founded later in the Summer of 1997 called Netflix, with the idea of distributing DVD rentals by post (and later by broadband).

ASCI Red reached a big milestone — it was the first supercomputer to reach greater than one TeraFLOP — that is a unit of computing speed equal to one million million (1012) floating-point operations per second. With later upgrades it would stay as the fastest supercomputer in the world for four more years.

But let’s fast forward to 2017.

In March, graphics chip manufacturer Nvidia announced the launch of a new device called the Jetson TX2 and said that it can reach 1.5 teraFLOPS — or 50 per cent more calculations per second than the ASCI Red reached back in 1997.

And no, the TX2 is not a supercomputer that takes up 2,500 square feet.

The Jetson TX2 fits in your pocket, costs $599 — and it’s the size of a credit card. As they say on Twitter these days — let that sink in.

The TX2 runs in two power modes — low power and high power. Low power means it consumes 7.5 Watts — though means it performs less, while in high power mode (15 Watts) it consumes more power but ramps up its performance. It comes with 8GB of RAM and 32GB of onboard storage, USB 3.0 and WiFI built-in.

What on earth would you need 1.5 teraFLOPS of processing for on a device so small? Well this where things either get really interesting, or really scary, depending on your viewpoint.

Nvidia say that TX2 is designed for robots, drones and general connected devices. Its architecture is designed for tasks such as facial recognition, object and voice recognition. Connected to a camera, the device can perform hundreds of facial recognition tasks per second. It’s not built for any old drone or robot — it’s built for the high end ones coming down the line.

The reason for the two power modes is that if the device is connected to a battery powered device — for example a drone — it will outperform other graphics devices while using less power thus allowing a drone to stay in the air for longer. Obviously it’s light and small enough to be embedded in a drone since — again — it’s the size of a credit card.

A drone flying overhead that could potentially recognise hundreds of faces per second is no longer a Minority Report style dystopia — it’s already here.

And let’s look at it another way. Factoring in inflation, the ASCI Red cost around $70m. That amount of cash today would buy us about 117,000 TX2s. That would get us to 175,000 teraFLOPS (or 175 petaFLOPS, or one hundred and seventy five thousand million million floating-point operations per second) in potential processing.

This isn’t even Nvidia’s top of the line processor — Nvidia recently launched a new desktop graphics processor designed for the computer sitting in your office or living room. This graphics processor — the Titan Xp — can reach 12 teraFLOPS — and costs just $1,200.

There is an important proviso when looking at supercomputers and teraFLOPS. Computers are usually only as fast as their slowest bottleneck — so comparisons between computers based solely on operations per second — can be misleading.

‘Normal’ supercomputers require vast amounts of memory and storage to go with their processing abilities in order to ease bottlenecks on the processing of huge amounts of information. Clearly, newer designs such as the Jetson TX2 have very small amounts of memory and storage — while having extremely small and efficient processors.

But it is important to consider just how far processors have come in the past 20 years — and the comparison stands for just how far we have come in fitting more processing power on to ever smaller chips.

But let’s take this a step further — as a thought exercise.

Two things are likely to happen in the coming years. One is the increased use of solar energy combined with battery storage — which is being pursued by several companies. The second is the continued speed and efficiency of processors, combined with cost reductions in memory.

This is leading to the what is called ‘edge’ computing, or what could also be called the inverse of ‘cloud’ computing. Instead of centralised processing in big data centers, things may shift to millions of smaller devices (hopefully secured), which are powered by a growing distributed power and energy storage network.

Distributed power generation and storage — combined with distributed data processing — could radically change our understanding of cost. Or a supercomputer could be running in your home or in your hand for a very low (or maybe no) operating cost.

What happens to society when massive compute power is approaching free? That is the question for the next decade.