IBM held its Quantum Developers Conference at its research headquarters in Yorktown Heights, New York this week and provided a progress report on their development activities and updates to its roadmap. At a high level, there have two themes in their development efforts:
- Quantum must be performant
- Quantum must be easy-to-use
The company advanced towards these goals through a combination of hardware improvements, software improvement, new tools, and collaboration with partners. We will cover each of these areas in this article.
Hardware Improvements
There were no dramatic changes from the roadmap IBM showed last year, but they have continued executing towards it and achieved the bulk of the goals they discussed a year ago. Here are some of the improvements they have made in the past year.
Currently, their flagship quantum chip family is named Heron. A year ago they showed Heron R1 which contained 133 qubits and was their first chip that contained tunable couplers to reduce crosstalk. They now have installed Heron R2 which increased the number of qubits to 156, and, equally important, made a number of improvements to improve gate fidelity. These improvements included implementation of two-level system (TLS) mitigation techniques, better tunable coupler calibrations, and implementation of fractional gates which will reduce the number of levels needed to implement certain gate operations. Next year they are planning on improving the performance of dynamic circuits by reducing the duration through parallel execution of conditional blocks.
The Heron technology will be the building block for IBM’s quantum systems for the next several years. This technology will only be used for NISQ algorithms and will not enable error correction. However, IBM and others will continue to implement incremental improvements in the chip as well as increased use of error mitigation to provide better results with a goal of achieving quantum utility for some applications
The big efforts using the Heron technology will be to scale up the number of qubits by implementing multi-module systems using multiple Herons connected together. To do this, they have been developing two types of couplers. The first is called an L-coupler that will connect two Herons as far as a meter apart using cables. The second is called an M-coupler that will connect two chips adjacent to each other. IBM has built in their lab a prototype of a processor called Flamingo that will connect two Heron’s using this L-coupler as shown in the picture below. They plan on continue to optimize the design and plan to make it available in the second half of 2025 to provide a system with over 300 qubits.
IBM is also developing a module called Crossbill that consists of three Herons connected with the M-coupler as shown in the picture below. In 2023, IBM produced an 1121 qubit test chip called Condor to help test out system scaling and fridge capacity. The picture below shows a comparison of a prototype Crossbill to the Condor including the associated packaging which demonstrates how the new architecture packs is more space efficient.
Although these developments will provide increase processing power for the next several years, we all know that supporting fault tolerance is essential for the future and IBM is working on a new architecture to support this. To do this, they need to move away from the heavy-hex architecture used in Heron and many of its predecessor which had a limited qubit-qubit connectivity of three neighboring qubits. They have designed an error correction architecture based upon Quantum Low Density Parity Check (Q-LDPC) that is more efficient that the surface code, but it requires a qubit to be able to connect to 6 others as shown in the diagram below. IBM described this approach in a recent paper published in Nature magazine and will be using it in future chips that are under development.
To support this, IBM is developing a third type of coupler called a C-coupler as shown in the diagram below:
One additional innovation that IBM is working on for the longer term machines is to use cryo-CMOS control chips for providing a further improvement in qubit quality as well as a method for implementing the control signal connections necessary for systems that contain thousands of qubits. IBM has also built a prototype chip for this and has performed some preliminary testing of it as shown in the picture below.
So there is a lot going on, and putting it all together provides a longer term roadmap that is shown in the chart see below. If you look carefully, you will see green checkmarks in the upper right corner of the Qiskit Code Assistant, Qiskit Functions Service, Qiskit Transpiler Service, Heron (5K), Flamingo, and Crossbill. These denotes technical goals that were set a year ago that have been achieved during 2024.
100 x 100 Challenge
One way of measuring your progress and motivating engineers is to set a challenge for some goal that needs to be obtained. IBM did that in 2023 by establishing the 100 x 100 challenge which specified that they wanted to run a circuit with 100 qubits and 100 gate levels, 5000 two qubits gates, and get an accurate answer within one day’s run time. At the conference they announced they have achieved this milestone with the Heron R2 processor.
Massive Improvements in Runtime Performance
It’s not often that we see a company make a 50X improvement in a performance metric over a couple of years, but IBM has made this type of improvement in its CLOPS (Circuit Layer Operations per Second) metric. It turns out that a lot of the time needed to run a quantum program is used in all the classical computing activities needed to support the calculation. This includes delays in queueing, compilation, post processing, etc. IBM used a lot of lessons it has learned from its classical HPC activities to make major improvements. These techniques include parallel compilation, implementing parameterized circuits, job scheduling, code pipelining, optimizing data movement and other things. The result is that the CLOPS measure has improved from about 3K in 2023 to about 30K at the beginning of 2024 to 150K now. This latest level has just been made available at a beta level for premium customers to try out. Improving runtime performance will be important for end users because it will improve the time to solution. It also benefits IBM because they can process more customer jobs every day. In addition, IBM is introducing GPU support as part of their Quantum Serverless capabilities. We are not aware of anyone else who has optimized the integration between their quantum and classical computing resources to this extent.
In HPC environments, it is a complex task to manage the CPU and GPU resources in orer to achieve the most effective utilization. Some of these installations use a tool called SLURM to help them do this. IBM has been working with the Rensselaer Polytechnic Institute to help integrate their IBM Quantum System 1 with RPI’s AiMOS supercomputer and their SLURM resource manager and has developed Resource Management software for this. This allows them achieve a heterogenous workflow in a fully realized quantum-centric supercomputing environment.
Qiskit Transpilation Advances
One area that has seen significant improvement over the past year is the transpilation function in Qiskit. This is the function that takes the quantum program inputted by a user and converts it to the most efficient implementation using the native gates of the processor. This can be a very complex and resource intensive process.
In order to make these transpilations quicker, the company has converted the bulk of the transpilation pipeline from its original implementation to the much more efficient Rust programming language. In doing so, they showed improvements in transpilation times by as much as 60X. Not only are the transpilations performed faster, but the resulting output can create results with significantly fewer gates. In some examples shown at the conference, the latest version of Qiskit 1.3 was able to achieve a 18% and a 21% reduction in 2-qubit gate count over the nearest competitor.
An additional improvement is the offering of a Qiskit Transpiler Service that is available in preview mode for Premium Users. This service adds additional capability that uses AI-powered transpiler passes to reduce gate depth by as much as 30%.
Qiskit Functions, Qiskit Add-Ons, Qiskit Code Assistant, and Debugging Tools
In order to make it easier for end users to develop programs, IBM has created Qiskit Functions, which is a catalog of services to provide certain useful functions. This catalog already includes three application functions and four circuit functions provide by IBM and a few software partners. And we expect this number will grow quite a bit in 2025.
The application functions currently available include:
Circuit Functions
- IBM is providing a circuit function that provides accurate hardware results using AI powered circuit optimization and advanced error mitigation.
- Q-CTRL is providing a circuit function that includes its Fire Opal performance management software that applies AI-driven error suppression without extra overhead to allow users to run larger circuits at peak performance.
- Algorithmiq is providing a Tensor Network Mitigation method (TEM) that performs noise mitigation entirely at the classical post processing stage. This method results in unbiased estimators with a fewer number of shots and less runtime than other methods such as probabilistic error cancellation (PEC) or zero-noise extrapolation-probabilistic error amplification (ZNE-PEA).
- QEDMA is providing a function that implements efficient and accurate characterization of the quantum processors and uses that data to implement error suppression and error mitigation.
Application Functions
- Qunasys is providing their Quri Chemistry software that computes the ground state energy and electron configuration distribution of a molecule.
- Q-CTRL is providing a function for an optimization solver that lets a user input the definition of a problem at a high level and then proceeds an optimal solution for that problem.
- Multiverse is providing a Singularity machine learning classification tool for supervised learning tasks
Another capability that IBM has added to make it easier for a customer to develop a program for an application is called Qiskit Add-Ons. These are modular tools that a customer can plug into their workflow to help design new algorithms. The currently available add-ons include:
- Multi-product formulas (MPF) – reduce the Trotter error of Hamiltonian dynamics through a weighted combination of several circuit executions.
- Approximate quantum compilation (AQC-Tensor) – Enables the construction of high-fidelity circuits with reduced depth.
- Operator back propagation (OBP) – Reduces circuit depth by trimming operations from the end at the cost of more operator measurements.
- Sample-based quantum diagonalization (SQD) – Classically post-processes noisy quantum samples to yield more accurate eigenvalue estimations of quantum system Hamiltonians, for example in chemistry applications.
Another service available to IBM’s premium users is the Qiskit Code Assistant. This service uses Generative AI to help the customers write code based upon Qiskit models and Watsonx. In the future, this will be able to help users migrate their code to new versions of Qiskit, detect wrong code, and suggest fixes. Another new tool in development is a debugger. When something goes wrong, this tool will help a user understand why.
Education and Community Building
From the very beginning, IBM realized that to make sure they needed to make investments to ensure that users were trained and ready to make use of the quantum technology they developed. A current effort is to provide a Learning Paths module including courses, tutorials, and educational resources taught by leading quantum experts. They have formed working groups with leading scientists in the areas of Optimization, Materials+HPC, High-Energy Physics, Healthcare & Life Sciences, and Sustainability to study how quantum computing can be leveraged to solve problems in these areas. They have an Open Plan offering that provides free time of up to 10 minutes/month for users to start learning quantum. The company has also sponsored many hackathons. And, of course, they held this three day Quantum Developers Conference at their research headquarters site with a couple hundred quantum developers to provide updates on IBM’s latest quantum offerings and to train the developers on the latest tools and learn how these developers can best utilize the new tools.
Conclusion
The best way to understand IBM’s quantum strategy at the high level is in terms of GQI’s Quantum Tech Stack framework as shown below:
The important thing to understand about IBM’s strategy is that they are making significant efforts at ALL levels of this tech stack, from the Quantum Chips at the bottom of the stack to the User Community at the top and all levels in-between. Although some may argue that a competitor’s technology might be stronger for an individual piece, we know of no one who is working at all the levels to the same extent that IBM has been doing. Being able to provide a complete solution and implement a tight integration between the different levels provides many benefits that makes their overall quantum offering very competitive.
November 16, 2024
