Here's What It Took To Build IBM Quantum's Hummingbird Processor

 




This past summer, IBM Quantum researchers Neereja Sundaresan and Srikanth Srinivasan were tasked with masking up and heading into the lab at the Thomas J. Watson Research Center in Yorktown Heights, New York in order to help wire up and cool down the largest quantum device the team had ever built, the 65 qubit IBM Quantum Hummingbird, revision 2. Dozens of researchers on the team had been working around the clock in order to make the device available on the cloud before Labor Day, and these initial tests would likely determine whether or not the whole team would have to go back to the drawing board. However, when they cooled the fridge and first took benchmarking measurements, everything was wrong. 

The numbers didn’t make any sense—the IBM Quantum team had put months into designing and simulating this all-new processor from the bottom up, incorporating advances in wiring, fabrication, and qubit layout taken from the previous IBM Quantum Falcon revision 4. Design parameters were vetted to work in-step with the latest room temperature control electronics, amplifiers, packaging constraints, and backend calibration routines. “We thought we turned over every stone in preparing for this build; we were so confident it would work. Seeing that first horrible result was pretty shocking," said Sundaresan. After staring at the data coming out of the fridge side-by-side with designs for a couple hours, they started suspecting where they went wrong. Figuring out the 65-qubit device's new wiring scheme required rotating some designs and flipping others. The team had a hunch by the end of the day. They turned off the electronics, warmed up the fridge, and confirmed their suspicions were correct: they essentially mounted the whole thing upside-down. Indeed, they had to spend a few hours rewiring the dilution fridge that held the chip, plugged it back in, and voila—the chip passed its tests with flying colors, performing exactly as they’d hoped.

The second revision of the IBM Quantum Hummingbird processor is the most advanced device that the IBM Quantum team has produced yet, and one of its most important. This device was totally re-imagined to combine lessons learned across the team in order to develop a scalable architecture with error correction in mind. It carries with it three important advances that will be baked into future IBM Quantum computers: "breaking the plane," or the ability to isolate qubits by stacking multiple wafers together with bump bonds; 8-1 readout multiplexing, allowing one output cable to readout the signal from eight qubits; and a "heavy-hexagonal" layout, arranging the chip into a tessellated hexagon layout where qubits sit on the edge and vertices of each hexagon in order to produce more "good" devices during the fabrication while anticipating future error-correction strategies. 

What is a Quantum Computer... In fact, What is a Qubit?

Feel free to skip to the next section if you understand how superconducting qubits work, but for those that don't: you know how computers are devices that manipulate data encoded into "bits," physical systems that represent the zeroes and ones of binary code, controlled by logic gates and arranged into logic circuits? Quantum computers do the same thing, but two quantum states (like two different position of an electron around an atom) represent the zero and one, and these quantum bits or "qubits" are controlled by quantum gates and arranged into quantum circuits. Unlike regular bits, single qubits enter superpositions, taking on values that are a little bit zero and a little bit one during the computation, entangle, meaning that two qubit values are more tightly correlated than the classical rules of probability allow, and interfere, meaning certain combinations of zero and one become more likely than others at the end of the computation.  

We think that one day, a computer running these quantum circuits could have revolutionary applications in modeling new molecules (like for pharmaceuticals or materials science), solving optimization problems for businesses across industries, and who knows what else. Many people who work on these devices are just excited about them because they're super advanced pieces of technology that might be able to probe the laws of the universe and look under the hood of the quantum world like no devices ever built before.
 
Among the biggest challenges that quantum hardware engineers face are finding a physical system to represent qubits that has controllable quantum states, ensuring that the qubits don't forget their quantum information (also known as decohering), and building a device with enough of these qubits in order to actually run quantum calculations. The physical system that the IBM Quantum team builds to represent qubits are called transmons. Zlatko Minev, IBM Quantum Research Staff Member, does a pretty amazing job of explaining how transmon qubits work in his lecture here, but basically, each qubit is a circuit of superconducting wire consisting of a capacitor, which stores electrical energy, and a nonlinear inductor called a Josephson junction, which stores magnetic energy. Electrical current oscillates between each of these components. The lowest mode of oscillation (think of a wire fixed on two ends just bouncing up and down) is the qubit's zero state, while the first mode above that (a fixed wire vibrating with a full wave and a node in the middle) is the qubit's one state. Injecting this oscillator with a specially-tuned pulse of microwave radiation excites it from zero to one or to some state in between.

This system uses a Josephson junction, rather than a regular inductor, in order to ensure that it doesn't take the same amount of energy to excite the circuit from the lowest state to the first excited state as it would from the first to the second state or higher. Imagine a toy designed to mindlessly climb a ladder based solely on the distance between rungs, where the lowest two rungs represent the lowest two qubit states. The machine would keep climbing if the rungs were evenly spaced, but if you squished all the rungs above the lowest two, then the machine wouldn't advanced further than these first two states. A superconductive circuit with a capacitor and a regular inductor would be the regular ladder, while the transmon qubit with the Josephson junction is the one with squished higher rungs.

In order to access the quantum behavior of superconducting wires, and to ensure that energy from the outside world doesn't destroy the fragile state of the qubits, these devices must be held super cold in a vacuum. Dilution fridges, the chandelier-in-a-garbage-bin looking devices that have grown to symbolize quantum technology, are responsible for the chip chilling.

Logos for Falcon and Hummingbird. The hexagons represent our heavy hex layout.


Building a Canary, a Falcon, and a Hummingbird

Perhaps harder than designing a transmon qubit is figuring out how to build a lot of good ones. The IBM Quantum team etches qubits onto a chip with similar lithography principles used to build silicon microchips. First, the team fabricates the capacitors and resonators responsible for communicating with the qubit using Reactive Ion Etch (RIE) to remove metal from specific areas of a wafer with a deposited superconducting metal stack. Then, to build Josephson junctions, the wafer is covered in a thin layer of material called resist film. Then we pattern the resist with an electron beam, evaporate superconducting metal into the pattern left behind by beam, the wafer goes through a resist liftoff process, and eventually the only part that's left is the junction. Of course, all of this is done in a clean room, explained Cindy Wang, Research Staff Member on the IBM Quantum team, because anything—a stray particle of material, a dust speck from the air, or any residues left behind at all—can destroy the qubits' ability to stay coherent. The team continually tweaks the design of this chip and its control electronics—the arrangement of the qubits and the electronics that send signals to the qubits and read out their values—until they have a controllable, compact system with long coherence times, that they can scale up to larger and larger numbers of good qubits.

Ok, That last bit about tweaking and scaling was a major oversimplification. Refining our qubits, their layout, and their control electronics in order to minimize noise is a gargantuan challenge that has required constant experiments and redesigns to our chips in order to get us from 5-qubit chips that we first deployed on the cloud in 2016 to the IBM Quantum Falcon Processor, a 27-qubit chip with a Quantum Volume of 64, and the Hummingbird processor: a 65 qubit chip with a Quantum Volume of 32. Of course, quantum computers are far from a place where they can run applications useful to the average person or business—but it's still worth looking back at the progress we've made. Our devices represent the most advanced superconducting qubits with some of the best metrics in the world. They're the result of a number of innovations spanning 21 years, said Markus Brink, Manager of Device Design and Integration on the IBM Quantum team, maturing from devices with coherence times of just a few nanoseconds to more robust designs with coherence times over 100 microseconds today, built with the help of our top-notch research fabrication facilities. 

Designing the newest IBM Quantum Hummingbird required engineering feats across our smaller devices, like our IBM Quantum Canary 5-qubit processors and our IBM Quantum Falcon 27-qubit processors. As a note, this is a really complicated, non-linear story (because nothing interesting is ever linear!), explained Oliver Dial, Distinguished Research Staff Member and Quantum Science Portfolio lead of the IBM Quantum team. We don't think about advances based on number of qubits, and are instead always developing and releasing new smaller-scale devices in order to solve problems and identify things we'd like to carry forth to bigger processors. 

This is what our wire bonds looked like before we broke the plane. What a mess!


Breaking The Plane 

Our first quantum chips had the qubits and fan-out electronics that connected them to the outside world on the same plane. But qubits require lots of wiring and cables, which can lead to on-chip clutter (like the wires behind your desk) making it difficult to scale the chip to larger qubit counts, and provide an extra source of noise that might cause the qubits to decohere faster. This became obvious as we moved from our 5-qubit processors to our 27-qubit ones—we wouldn't be able to include all of the wiring on the same chip as the qubits without causing wires to cross. Early attempts to move wires out of the qubit plane with traditional CMOS (Complementary metal–oxide–semiconductor) techniques were affecting qubit performance, said John Cotte, Research Staff Member on the IBM Quantum team. So, the team had to figure out another way to move the fan-out electronics onto a different chip that sits above the chip with the qubits.

We call this "breaking the plane," since the processors' components now sit on multiple two-dimensional planes: one plane which we call the qubit plane, and the other the interposer plane. Our team developed our own methods, tailored for quantum, to connect these two planes via a flip-chip method, putting solder on one of the chips and then flipping one on top of the other, creating "bump bonds" where the solder balls connect the two. The circuitry on the top interposer plane connects through these bonds to the circuitry on the qubit plane below. The challenge for Hummingbird was finding ways to make these connections and wiring fan-out efficiently and with scalable design techniques while avoiding impact to performance through unwanted crosstalk. 

Our bump bond strategy has plenty of other benefits aside from keeping readout electronics separate from the qubits. It enables a smaller qubit chip footprint, which makes for a more reliable device by mitigating stresses due to temperature changes and allows us to increase the qubit chip yield, fabricating more good chips per wafer. Additionally, this two-stack process can further be augmented, as we start to apply more traditional CMOS techniques for ever more complex wiring schemes with the interposer chip. This will be especially useful as we scale up to larger qubit devices.
 

Eight different frequencies for multiplexing


Multiplexing readout

As our qubits grew in size, so too did the amount of components that we needed to control and readout. In earlier generations of our systems, including the 53-qubit version of Hummingbird we released last year, the readout of the qubits took an act of brute force. For 53 qubits, we used 53 readout circuits with 53 quantum amplifiers. If we hope to one day build a thousand or even a million-qubit device, it would simply be unreasonable to have the complex equipment that comes along with this many readout chains, including heavy permanent magnet isolators and circulators plus multiple flavors of amplifiers, passing through the fridge. So, we started to experiment with five-to-one readout multiplexing in our Falcon devices—that means that instead of having readout circuitry for each qubit, a single set of readout circuitry enables the measurement of five qubits. When we applied the same principles to the newest Hummingbird chip, we went with eight-to-one multiplexing instead, eight qubits per set of readout circuitry.
 
Bulky isolators that necessitate multiplexing

How does this work? Well, each qubit is coupled to a resonator, which resonates based on the state of the qubit after we hit it with a microwave pulse. Each resonator has a corresponding frequency that differs from its neighboring resonator. Since we can send different frequency signals over the same wire simultaneously, we can send microwave pulses to eight resonators at the same time without causing any interference.

But there's another issue here. The initial release of Hummingbird was using 53 amplifiers to amplify the extremely quiet signal from each qubit such that the room temperature electronics could actually hear them. Each of these amplifiers used a resonant circuit, meaning amplifier could only amplify a small range of frequencies, but they conflicted with the multiplexing strategy of passing signals of different frequencies of the same wire. So, the team developed Traveling Wave Parametric Amplifiers (TWPAs, pronounced two-puhs) customized to our quantum processors. These TWPAs are composed of a wire containing several thousands of Josephson junctions, driven by another microwave signal referred to as the pump tone; as the readout signals pass through, the Josephson junctions impart them with more and more energy from the pump tone in a process akin to pumping your legs on a swing set, said David Lokken-Toyli, IBM Quantum research physicist. These TWPAs can amplify a much broader range of frequencies, and using them allowed the team to decrease Hummingbird's total number of amplifiers to nine. 
 

Two different representations of the heavy hex


The Heavy Hex 

There are plenty more considerations than just building a compact and low-noise system. We need to ensure that our yield is high; while we can fabricate a dozen or more chips at a time, the highest fraction of them possible should be "good" chips. We also need to think about these systems in a modular way so we don't have to completely redesign and perfect the layout of our qubits each time we hope to make a bigger processor. Plus, our processors must be ready to incorporate error correction—sharing quantum information over multiple redundant qubits to protect them from giving us the wrong values during computations. That's why we debuted the heavy hexagon qubit arrangement, where qubits are arranged on a tessellation of hexagons, one qubit per edge and vertex of each hexagon. This pattern does't quite look like hexagons in real life; but it's still a tessalation of six-sided shapes where each qubit is connected to two or three other qubits.

Why the heavy-hex arrangement? Other error correction codes, such as the ubiquitous surface code, worked based on a system where each qubit connected to four other qubits. Moving to a system where the qubits are less connected decreases the amount of crosstalk noise and makes these systems easier to make, increasing the overall qubit chip yield. Additionally, each qubit needs to operate at a frequency distinct from its neighbors like stations on a radio dial so they don't interfere with one another, but setting the transmon frequencies is especially challenging; decreasing the number qubits in contact with one another helps reduce the chance for any interference. We understand that this is a sacrifice, since connectivity is of crucial importance to quantum computers of the future, and we'll have to overcome the decreased connectivity with better quantum gates. However, we think the heavy hexagon arrangement gives us the best balance of high yield plus lower crosstalk noise and frequency interference, explained Jared Hertzberg, Research Staff Member on the IBM Quantum team. It's a sacrifice we're willing to make and a reasonable way to think about scaling up our devices. Last year we demonstrated that we can still implement error correction strategies over this new heavy hex layout; now, our team is thinking about how we'd implement gates on these logical qubits.

Of course, none of this means we have an error-corrected logical qubit yet—that will require further improvements to our qubits—but we can at least build increasingly large chips with error correction in mind, and begin to experiment with error correction on our Hummingbird chip. The Heavy Hex is part of a continued focus on hardware-aware advances to quantum computing, explained Andrew Cross, Research Staff Member and Manager on the IBM Quantum team. We build these devices, we notice new constraints, tailor our thinking in order to adapt, try out our new approaches, and iterate. This has become a story of constant communication between our theory and experimental teams, and between theorists and experimentalists in the field as a whole. 

Looking Forward

The advances gleaned from our 5 and 27 qubit devices have brought us to a place where we can fabricate and maintain a 65-qubit device with a Quantum Volume of 32—and it wasn't easy. Each advance required overcoming roadblocks with creative problem solving from across IBM Quantum, representing an enormous team effort where every individual's expertise was required in order for the chip to be successful. Now, we'll be able to take these advances, and other ideas we're cooking up, and take them through to our 127-qubit Eagle chip, 433-qubit Osprey chip, and 1121-qubit Condor chip—but that's for another post. For now, the 65-qubit Hummingbird revision 2 is available for IBM Q Network clients, and we're excited to see what they can do with it.





No comments:

Post a Comment