Quantum Computing with Spectre's Ultra-Low Temperature Models

Equal1 has just announced a breakthrough in quantum computing with a fully integrated quantum processor operating at 3.7K using commercially available FD-SOI technology from GLOBALFOUNDRIES (GF) in Dresden, Germany. The chip was designed with Cadence software, in particular the Spectre Simulation Platform's new support for ultra-low temperature models.

Britain is famous for serving warm beer. Actually, this is a misunderstanding—it should be cellar temperature and so cool rather than ice cold. In a similar way, "hot qubits" means 3.7K as opposed to ~50mK, so cellar temperature as opposed to truly cold. The reason temperature is so important is that thermal activity destroys quantum coherence, the fundamental basis of quantum computing.

I recently talked with Jason Lynch, the COO of Equal1. They are based in Dublin, Ireland, a city more famed as the home of 42.8°F Guinness than for 3.7K quantum computing. In my post What Is Quantum Supremacy? I wrote about Google and IBM's quantum programs. They are working in the milli-Kelvin type of quantum computing. Equal1 is working with hot qubits at 3.7K. By the way, it is not correct to say 3.7°K nor to talk about degrees Kelvin. Although a Kelvin is the same as a degree Celsius, they are not called degrees.

Jason gave me a potted history of the company. CEO Dirk has a PhD in Quantum Physics and had the dream of building a quantum computer. In 2005, they were looking at single-electron transistors at Texas Instruments (TI), but they decided it was too early, and so the founders went their separate ways. After seeing a presentation of GF's 22nm FD-SOI process that GF calls 22FDX, the team realized the time was right for a silicon quantum computing company. This process is manufactured in GF's Fab 1 in Dresden (the old AMD fab that GF inherited when it was spun out). For details on FD-SOI and 22FDX in particular, see my post Cadence Tool Suite Qualified for 22FDX Reference Flow. This post gives a history of FD-SOI and compares it to FinFET.

After hearing about FD-SOI, as Jason put it:

The founders decided to get the band back together. Now there was a way to do quantum on commercially available CMOS.

If you've seen any pictures of Google or IBM's quantum computers, they are more like a huge art installation than a computer, with all those beautiful brass and gold fittings. It is obvious that they are expensive. You are not even seeing the literally tons of equipment required to get the temperature down to the milli-Kelvin level. The big advantage of working at 3.7K is that those temperatures are used in medical MRI systems, and so are commercially available.

The chip that Equal1 just announced is its second-generation quantum processor. It is a flip-chip that goes onto the PCB board and then into the cold chamber. It has two demonstrator systems that have been running for 18 months, one in Dublin and one in San Carlos, California. The chip contains a quantum dot array of 4424 dots. The first iteration is charge-based. The second has proved quantum behavior. At the end of May, Equal1 taped out the third chip.

With everything integrated onto a single chip, the mechanism to get the data in and out of the computer is done directly with electronics, as opposed to the milli-Kelvin approaches where it requires microwaves and waveguides. The entire system including the cooling is still a lot bigger than just a chip, but it is rack-sized rather than room-sized.

At 3.7K, the big challenge is poorer decoherence time. Jason explained:

Superconducting qubits have decoherence times in the milliseconds but we’re more like 150ns with a roadmap to get up to a microsecond. But decoherence time is just one lever. If you can flip the gates faster then that's another lever. Because the superconducting qubits are at a much lower temperature, the gate flip times are much longer. We are clocking in picoseconds but the other technologies are an order of magnitude above that.

The question we get asked is 'how can you guys be on a commercial process when these other guys are spending hundreds of millions of dollars?' They have been doing it for decades and are up to somewhere under 100 qubits. We are getting a commercial CMOS process to be good enough and there are lots of applications that there are lots of things that it is applicable too. So not cryptography, but neural networks can live with a shorter coherence time as long as there is local entanglement. So we don't think we are competing with the Google's of the world.

I think targeting artificial intelligence and neural networks is a good match. Quantum computing is inherently probabilistic and so are neural networks. In fact, in the non-quantum implementation of neural networks, one attractive approach is to use resistive RAM (ReRAM) but use values between 0 and 1, and substitute having lots of bits instead of accuracy of storing the weights. 

Equal1 have a number of collaborations, many of which it cannot talk about. There is a publicly announced project with IBM and Mastercard. It also has a close relationship with the Stanford Linear Accelerator Center. By the way, if you live (or visit) Silicon Valley, if you don't know it is there, you might never have noticed the accelerator. But Interstate 280 goes over the top of it near the Sand Hill Road exit. Since the accelerator building is a mile long, it is hard to miss if you are looking for it. Other engagements are on genomics/genetics trying to detect cancer, and one looking at detection of ships from satellites. These are all difficult AI problems. Equal1 is also looking for huge improvements in standard models like ResNet.

As I mentioned in the opening paragraph, the chip was designed with Cadence tools. The Equal1 team designed all the complex mixed-signal circuits such as high-speed pulse-generators, ADCs, DACs, and cryogenic memory. The chip has over 10M transistors, and includes not just the quantum dot array, but all control and read-out electronics. One key, essential to cryogenic design validation, was the Spectre Simulator's new support for ultra-low temperature models. You can't just take a normal model and then set the temperature to 4K instead of -20°C.

I asked Elena Blokhina, Equal1's CTO, who joined the call, what made FD-SOI so appropriate for this:

Buried oxide (BOX)  is what makes it attractive. Also the high purity silicon channel with isolation from the wafer. We can also bias the channel.

And the future? Jason told me:

We plan to scale to 1M qubits to and produce an AI quantum accelerator/ demonstrator called Aquarius that is significantly more cost effective with a form factor in the server rack size.

My Photo Tour

I said earlier that Equal1 are based in Dublin, and one of their systems is there. But there is a system in San Carlos, up near San Francisco Airport. So I took a trip up there to get some photographs. At first glance, the system looks surprisingly like a normal vending machine...until you look in through the glass front panel. The innocent-looking panel shows just how cold it is (3.413K, which is warm for quantum computing but really cold for almost anything else).

Some of the internals laid bare:

And the future:

Sign up for Sunday Brunch, the weekly Breakfast Bytes email.