Dynamiqs: GPUs are Just the Tip of the Iceberg
“Iceberg in the Arctic with its underside exposed” by AWeith is licensed under CC BY-SA 4.0. To view a copy of this license, visit https://creativecommons.org/
The English language idiom “tip of the iceberg” means that we’re only observing a small fraction of something. Inspired by an actual iceberg, we’re only seeing the tip of whatever it is. A much larger fraction of the iceberg, object, problem, or situation remains hidden from us. The rare photo above is an exception to the rule.
Alice & Bob recently published “Alice & Bob Supercharge Quantum Simulations with Dynamiqs by Integrating with Accelerated Computing,” in which it largely attributes a 60X performance boost in quantum simulations to NVIDIA GPUs. Several paragraphs down, however, there is a brief mention of differentiability, hinting that GPUs might be only the tip of the Dynamiqs iceberg. And indeed, they are.
GPU Compatibility
The standard library in use today, and for at least a decade, has been QuTiP. Interestingly, QuTiP is not yet compatible with GPUs. Dynamiqs, on the other hand, as the article states, can already take advantage of GPU acceleration. You can actually switch from CPUs to GPUs with a single line of code, resulting in an as-expected performance boost.
Differentiability
Dynamiqs is the first library to provide differentiability. This allows you to simulate differential dynamics and interactions with the environment, and to solve differential equations. It also enables you to compute gradients for quantum state tomography and…
…Optimal Control
Dynamiqs can help find optimal values for the controls of a parametrized quantum system. It provides the gradients of a given function to optimize (e.g., gate fidelity) with respect to input parameters (e.g., drive amplitude). It solves an equation that essentially provides an educated guess, starting closer to the solution, reducing the number of iterations required, and helping to push fidelity closer to 100% for optimization problems. The method is also applicable to system calibration, parameter estimation, and the training of neural networks describing quantum systems. It can also be used to optimize any state preparation, quantum gate, or state readout using any quantum computing modality (e.g., superconducting circuits, trapped ions, neutral atoms, etc.). Alice & Bob hopes that there will be many more applications for this within the next 5-10 years.
Tricks of the Trade
Much of the performance boost of Dynamiqs comes from little-known, highly technical techniques used in Academia. These “tricks” are too minor to be published, which keeps them little-known, but they’re what really help to squeeze out maximum performance.
Sponsored Open Source
Whereas QuTiP is open source and has grown organically, it is primarily a volunteer effort. In contrast, Dynamiqs is an academic collaboration and is sponsored by a company, Alice & Bob. In time, thanks to this sponsorship, users can be reassured that Dynamiqs will have durable support and maintenance – a safety net for long-term sustainability – whereas QuTiP may or may not.
Conclusion
Dynamiqs is an exploratory tool. The idea is to give this tool to Academia and see what it can do with it. While the general public recognizes the concept of “GPU acceleration,” and that’s the angle the article took, it’s important for researchers to know that – like an iceberg – there’s more beneath the proverbial surface. Dynamiqs has novel features hitherto unseen in QuTiP or elsewhere, and it will be interesting to observe what might come out of this.
