IBM today issued an extensive and detailed expansion of its Quantum Roadmap that calls for developing a new 1386-qubit processor – Kookaburra – built from modularly scaled chips, and delivering a 4,158-qubit POC system built using three connected Kookaburra processors by 2025. Kookaburra (Australian Kingfisher) is a new architecture that IBM says will “leverage three pillars: robust and scalable quantum hardware; cutting-edge quantum software to orchestrate and enable accessible and powerful quantum programs; and a broad global ecosystem of quantum-ready organizations and communities.”
The revised roadmap still includes delivery of previously-announced single-chip 433-qubit Osprey processor this year and 1121-qubit Condor processor next year, but introduces newly-architected processors intended for use in multi-chip modules, starting with a 133-qubit (Heron) processor in 2023, a 408-qubit processor (Crossbill) and 462-qubit processor (Flamingo) in 2024, and leading to Kookaburra in 2025. Besides hardware improvements, IBM discussed key software advances and introduced its broad concept of a quantum-centric supercomputer leveraging both classical and quantum technology.
Jay Gambetta, IBM Fellow and vice president, quantum computing, delivered the expanded IBM vision in a blog posted this morning.
“We aren’t just thinking about quantum computers, though. We’re trying to induce a paradigm shift in computing overall. For many years, CPU-centric supercomputers were society’s processing workhorse, with IBM serving as a key developer of these systems. In the last few years, we’ve seen the emergence of AI-centric supercomputers, where CPUs and GPUs work together in giant systems to tackle AI-heavy workloads. Now, IBM is ushering in the age of the quantum-centric supercomputer, where quantum resources — QPUs — will be woven together with CPUs and GPUs into a compute fabric. We think that the quantum-centric supercomputer will serve as an essential technology for those solving the toughest problems,” wrote Gambetta.
No doubt the devil will be in the details. IBM has so far hit most quantum milestones it set for itself as shown in the chart above (click to enlarge). Gambetta’s blog is wide-ranging and covers most aspects of IBM’s evolving quantum strategy. With apologies for the length of this excerpt, here’s Gambetta’s summary of the hardware news. There’s a good deal more around software:
“With the 433-qubit “Osprey” processor and the 1,121-qubit “Condor” processors — slated for release in 2022 and 2023, respectively — we will test the limits of single-chip processors and controlling large-scale quantum systems integrated into the IBM Quantum System Two. But we don’t plan to realize large-scale quantum computers on a giant chip. Instead, we’re developing ways to link processors together into a modular system capable of scaling without physics limitations.
“To tackle scale, we are going to introduce three distinct approaches. First, in 2023, we are introducing Heron: a 133-qubit processor withcontrol hardware that allows for real-time classical communication between separate processors, enabling the knitting techniques described above. The second approach is to extend the size of quantum processors by enabling multi-chip processors. Crossbill, a 408-qubit processor, will be made from 3 chips connected by chip-to-chip couplers that allow for a continuous realization of the the heavy-hex lattices across multiple chips. The goal of this architecture is to make users feel as if they’re just using just one, larger processor.
“Along with scaling through modular connection of multi-chip processors, in 2024, we also plan to introduce our third approach: quantum communication between processors to support quantum parallelization. We will introduce the 462-qubit Flamingo processor with a built-in quantum communication link, and then release a demonstration of this architecture by linking together at least three Flamingo processors into a 1,386-qubit system. We expect that this link will result in slower and lower-fidelity gates across processors. Our software needs to be aware of this architecture consideration in order for our users to best take advantage of this system.
“Our learning about scale will bring all of these advances together in order to realize their full potential. So, in 2025, we’ll introduce the Kookaburra processor. Kookaburra will be a 1,386-qubit multi-chip processor with a quantum communication link. As a demonstration, we will connect three Kookaburra chips into a 4,158-qubit system connected by quantum communication for our users.”
In keeping with its steady hardware advances, IBM says it will raise its Quantum Volume score (performance benchmark) from 256 to 1024 this year and also increase the highest CLOPS (circuit layer operations per second) from 2.9k to 10k, noting that faster processing should permit incorporation of more error mitigation measures. The QV benchmark, developed by IBM, has been used by others. For example, Quantinuum recently reported achieving a QV of 4096.
Bob Sorensen, chief analyst for quantum computing, Hyperion Research, said, “The most interesting part is the reference to quantum parallelization in the blog and quantum communication beyond 2026 in the roadmap. Single quantum processor qubits counts are a good indicator of technological advance at the processor level, but to truly get to high-qubit-count QC systems, QC architects will need to take a page out of the classical HPC designers handbook and begin to assemble QC systems that contain multiple quantum processors connected by a fast, efficient network.
“[Moreover], the interconnect/network between those quantum processors cannot be exclusively classical-based: it will need to support quantum processor to quantum processor communication that stays in the quantum realm. the ability to field such quantum networks becomes a major determinant of how fast QC systems can scale to reach the oft mentioned goal of 1 million qubits,” said Sorensen.
Hardware advances, of course, are insufficient to drive IBM’s vision forward. Gambetta’s blog spends equal time on software plans. Broadly, the idea is to develop tools that help shield developers and users from the underlying complexity of quantum computing hardware. It’s an idea that’s shared by virtually everyone in the quantum community, and IBM has framed this concept as Quantum Serverless computing.
“Quantum Serverless centers around enabling flexible quantum-classical resource combinations without requiring developers to be hardware and infrastructure experts, allocating just those compute resources a developer needs when they need it. In 2023, we plan to integrate Quantum Serverless into our core software stack in order to enable core functionality such as circuit knitting,” said Gambetta, who incidentally is an HPCwire 2022 Person to Watch.
IBM has long been an active participant in building out the needed quantum software ecosystem. Its Qiskit programming framework suite, now open source, has steadily gained wider use. Gambetta noted, “different users have different needs and experiences and we need to build tools for each persona: kernel developers, algorithm developers, and model developers.” He singled out three specific targets for software advancement: dynamic circuits; Qiskit primitives; and circuit knitting.
Dynamic circuits allow for feedback and feedforward of quantum measurements to change or steer the course of future operations. “They extend what the hardware can do by reducing circuit depth, by allowing for alternative models of constructing algorithms, and by enabling the fundamental operations at the heart of quantum error correction,” wrote Gambetta. “Debuting dynamic circuits required several advances: we needed new control hardware that could move data between components with lower latency and tight synchronization, which will only improve with the hardware advances, which we detail below. They also require a language to describe them, and as the community continues to mature the OpenQASM 3 language, we’ll be developing an OpenQASM 3-native compiler.”
Two Qiskit primitives were introduced earlier this year as part of Qiskit Runtime update. One of the primitives, Sampler, lets users reconstruct probability distributions of measurements from repeated runs of a circuit. It’s a useful tool for applications such as Grover’s algorithm. The other primitive is Estimator, which lets users “calculate the expectation values of operators, which can be useful for representing the electronic structure of a molecule, calculating the kernel of a machine learning problem, and much more,” says IBM.
Gambetta noted, “In 2023, we plan to update Qiskit Runtime with threaded primitives, allowing our systems to execute primitives on parallelized quantum processors, including automatically distributing work that is trivially parallelizable. We also plan to enhance primitive performance with low-level compilation and post-processing methods, like introducing error suppression and mitigation tools. These advanced primitives will allow algorithm developers to use Qiskit Runtime services as an API for incorporating quantum circuits and classical routines to build quantum workflows.”
Beyond primitives, IBM plans to continue beefing up Qiskit Runtime. “We want Qiskit Runtime to be an operating system focused on providing the tools that developers need in order to best utilize quantum resources. In 2024 and 2025, we’ll be introducing error mitigation and suppression techniques into Qiskit Runtime so that users can begin to calculate observables with less noise—and eventually, noise-free observables. These techniques will help lay the groundwork for quantum error correction in the future,” he wrote.
IBM and others have been discussing the need for circuit knitting for some time. The idea is to break larger circuits into smaller pieces to run on a quantum computer. “Earlier this year, we demonstrated a circuit knitting method called entanglement forging to double the size of the quantum systems we could address with the same number of qubits,” wrote Gambetta. “However, circuit knitting requires that we can run lots of circuits split across quantum resources and orchestrated with classical resources. We think that parallelized quantum systems with classical communication will be able to bring about quantum advantage even sooner.”
IBM has ambitious software goals. Gambetta says it will begin prototyping quantum software applications for users hoping to use Qiskit Runtime and Quantum Serverless to address specific use cases. “We’ll begin to define these services with our first test case — machine learning — working with partners to accelerate the path toward useful quantum software applications,” he wrote. “By 2025, we think model developers will be able to explore quantum applications in machine learning, optimization, finance, natural sciences, and beyond.”
There’s a lot to unpack in today’s announcement, which was timed to coincide with the start of IBM’s Think 2022 conference. This year, Think has been expanded into series of events – Think on Tour – commencing in Boston this week, followed by others across the globe including London, Berlin, Toronto, Singapore, Paris, Madrid, Sydney, Dallas, Mumbai, Beijing, and Tokyo. Some of today’s quantum roadmap is scheduled for discussion in Dario Gil’s (senior vice president, director of research, IBM) session on Future Technologies tomorrow at Think Boston.
Link to Gambetta blog, https://www.research.ibm.com/blog/ibm-quantum-roadmap-2025