Getting to the next generation of supercomputer isn’t just a function of making everything go faster — it’s about making things go faster without requiring multiple nuclear power plants dedicated to running the new machines. That’s been a problem dogging the researchers trying to build the next generation exascale supercomputers for the last few years, but scientists at MIT have discovered a new type of on-chip information transport that replaces wires with light that might help the problem.
Yup, a researcher named Michael Watts (let’s just take a moment to appreciate how he is fulfilling the destiny offered by his name) has built an on-chip photonics transport technology that only requires one femtojoule to transfer a bit. A femtojoule is a measurement of how much energy it takes to transfer a bit of data, but isn’t directly comparable to watts.
You see, the wires that connect the transistors on a chip shuttling information between them have long been problematic as we cram ever more of them on a chip. They consume a lot of power, they leak energy that interferes with the bits — it’s a hot mess. Literally.
Or from the MIT Technology Review article:
And therein lies a serious problem. An interconnecting wire on a chip just 1 millimetre long consumes about 100 femtoJoules for every bit it carries. That may sound small–a femtoJoule is 10^-15 Joules. But with data rates now hitting petabits per second (1 petabit = 10^15 bits), a large chip will eat about 100 Watts. And that’s just for the interconnecting wires. The power that transistors use up is on top of this.
The bottom line is that conventional interconnecting wires are at least ten times more power hungry than the next generation of chips can handle.
So researchers have turned to light to do this, but generating the pulses of light has a few problems. First, it takes power to generate those pulses of light. When we discuss photonics in the data center or on chip, transmitting data via light as opposed to over wires requires a power investment that’s not always practical. Second, you have to build separate structures on the chip to transmit photons.
Watts’ breakthrough lets people use the traditional CMOS manufacturing process to build photonics onto a chip — an element that makes it both cheaper and more feasible. The MIT article notes that the structure Watts has built also doesn’t react to temperature changes on the chip — another issue that previous generations of on-chip photonics parts have struggled with.
Now, getting the power budget for really fast supercomputers won’t solve all of the challenges the high performance and academic computing industry faces — there are still challenges of cost and even resilience — but this bit of research could eventually have applications in the data center and even future broadband networks.