Online retail giant Amazon has so many Traveling Salesman Optimization NP difficulties it’s crazy. In fact, Amazon has nested or chained NP-hard problems. Figuring out how to get an item from bin 23 to bin 42 across the country by any number of possible routes in the least amount of time (or energy, if you want to optimize for that) to your doorstep is enough to make your head spin break.
The problem is so bad that Amazon’s cloud computing arm, Amazon Web Services, invested in the AWS Center for Quantum Computing at the California Institute of Technology back in October 2021, and it’s still bad enough that AWS is… got to the bottom of various solvers from many finite element analysis programs and electromagnetic simulators and decided to build its own called Palace, which it has just released open source on GitHub to try and build a community around it.
Palace is short for Parallel Large Scale Computational Electromagnetics – this is how acronyms are tortured to make a real word. (PLSCEM rolls off the tongue, so we get it.) The macroscopic form of “Maxwell’s Wonderful Equations” that many of us encountered in college and how my own physics professor referred to the laws of electricity and magnetism from James Clerk Maxwell in 1861 and 1862 are difficult enough; The microscopic equations that apply at the atomic level can be very unwieldy, but are useful when designing a superconducting quantum chip, as AWS is doing right now because it wants to solve its ultrascale – what’s beyond hyperscale? – Optimization problems for traveling salesmen and become more profitable.
(Interesting and possibly a physics joke aside: it doesn’t take quantum electrodynamics simulation to build a quantum computer. Strange, isn’t it?)
AWS already offers a quantum computing service called Bracket that allows companies to play around with existing machines from IonQ, OQC, Rigetti, Xanadu, QuEra (but oddly not D-Wave Systems). Parent company Amazon has no doubt toyed with these devices and no doubt encouraged AWS to offer them as a service so we can all help Amazon foot the bill for its quantum computing experiments. But just as Amazon correctly thinks that it needs to innovate at the CPU SoC level to push server design forward, and therefore has developed Nitro DPUs and Graviton CPUs based on the Arm architecture, even if it has virtual parts of X86- Selling calculators on EC2, Amazon knows that to drive innovation in quantum computing by developing its own superconducting quantum computing chip.
More than the feature picture at the beginning of this story showing an AWS quantum processor, the phrase “superconducting quantum computer chip” that one of the researchers, Sebastian Grimberg, said to us, and the center’s setup at Caltech, nobody knows very much about it , what AWS has achieved in its quantum chip design or what approaches it is pursuing. What we do know, however, is that as far as AWS was concerned, the existing tools were too expensive and didn’t do the job properly at scale.
“It’s a really tough problem, especially when it comes to error correction,” Ian Colle, general manager of HPC at AWS and longtime user and developer of HPC systems, tells The Next Platform. “Our quantum computing team realized that this gap exists. There are really expensive tools here. There’s a bit of open source, and we thought that if we put a lot of our skill into it, maybe we could develop a new tool that would essentially be a very powerful, scale-based solver in this effort to build this quantum computer. ”
The resulting Palace solver is now said to run on Arm CPUs and X86 CPUs, but Grimberg, a senior research scientist at the Center for Quantum Computing, says Palace isn’t limited to electromagnetic simulation, but is also applicable to computational fluid dynamics . which also have very complex partial differential equations. (Grimberg has a PhD in Aerospace Engineering from Stanford University, and his colleague Hugh Carson, who also worked on Palace, has a PhD in Computational Science with a focus on CFD from MIT. Carson previously worked on the Amazon Prime Air drone program Amazon is also interested in tools that improve CFD.)
In a blog showing what Palace does, the physics of simulating a transmon qubit and a readout resonator in fine-mesh and coarse-mesh models is pretty hairy indeed, which is why you build a simulator. (Even Einstein had a mathematician.) What interests us is the interplay between the wall clock time in the simulation and the number of cores used to scale the simulation of the transmon qubit and its resonator, and how Graviton3 (C7g) and Graviton2 (C6g) instances have outperformed the X86-based instances in simulation, and how upscaling degrees of freedom in the simulated finite elements requires more cores.
Here are the plots showing the wall clock time and acceleration factor for adding more cores for the coarse-grain model with 15.5 million degrees of freedom in the simulation:
At least for this application with the Palace solver, the Graviton chips hold their own against the latest Xeon SPs, but it’s not clear why the “Skylake” Xeon SPs perform so well in the C5n instances. (Imagine that.)
Here is the higher-resolution simulation with 246.2 million degrees of freedom, which requires thousands of cores to run versus hundreds of cores and takes 12 minutes (more or less) to run, compared to 1.4 minutes for the coarse-grained model. This 15.8x increase in degrees of freedom is not free:
To push Palace even further, the AWS quantum team ran a simulation of a metamaterial superconducting waveguide based on a chain of lumped-element microwave resonators — and if you know what that is, Jeff Bezos wants to hire you. The idea is to predict the transmission properties of the waveguide in the 4 GHz to 8 GHz range in 1 MHz increments, and the test scales from 242.2 million degrees of freedom with a single metamaterial unit cell to 1.4 billion degrees of freedom with 21 metamaterial unit cells . In this case, all simulations are run on the C6gn Graviton2 instances and range up to 200 instances with a maximum of 12,800 cores. Have a look:
There are adaptive and uniform sampling modes shown in the diagram above, but again the interesting thing is that the Palace solver can scale over 12,800 Graviton2 cores. Presumably, AWS doesn’t have enough Graviton3 cores lying around to run the test on C7g instances, but if it did, we’d assume the wall clock time would go down quite a bit, and the scales would come out that far and about the same slope.
Incidentally, the instances used in these tests are set up by the ParallelCluster tool, created by AWS to easily set up virtual MPI clusters to connect nodes running the Palace solver.
When AWS will have a self-developed quantum chip is a mystery for now and probably for a while.
“We are committed to being able to develop a quantum computer,” says Colle. “I can tell you that, and we will bring it to market as soon as we feel it is feasible.”
