IBM quantum summit 2023 | NotionNext BLOG

主题：

utility

Road maps

IBM Condor, a 1,121 superconducting qubit quantum processor based on our cross-resonance gate technology. Condor pushes the limits of scale and yield in chip design with a 50% increase in qubit density, advances in qubit fabrication and laminate size, and includes over a mile of high-density cryogenic flex IO wiring within a single dilution refigerator. With performance comparable to our previous 433-qubit Osprey, it serves as an innovation milestone, solving scale and informing future hardware design.

the first IBM Quantum Heron processor on the ibm_torino quantum system. Featuring 133 fixed-frequency qubits with tunable couplers, Heron yields a 3-5x improvement in device performance over our previous flagship 127-qubit Eagle processors, and virtually eliminates cross-talk. With Heron, we have developed a qubit and the gate technology that we’re confident will form the foundation of our hardware roadmap going forward.

IBM 团队并不期望 Condor 处理器在执行量子计算方面像 Heron 一样有用，并将其更多地视为旨在升级公司硬件和软件堆栈的研究任务。另一方面，Heron 标志着门保真度和性能的显着改进，预计在复杂计算上会产生更好的结果。

https://www.allaboutcircuits.com/news/closer-look-at-ibms-heron-and-condor-quantum-processors/

IBM 量子系统二号是可扩展量子计算的基石，现已在我们位于纽约州约克敦高地的实验室运行。它宽 22 英尺，高 12 英尺，目前配备了三个 IBM Quantum Heron 处理器。它将低温基础设施与第三代控制电子设备和经典运行时服务器相结合。

IBM Quantum System Two 是模块化架构量子计算平台，我们将使用它来实现以量子为中心的超级计算的并行电路执行。

https://youtu.be/Qndz54SGCAs

IBM 通过在 IBM Quantum Platform 上运行的全球首个使用强化学习的电路编译服务，将 AI 的力量引入量子计算。此初始预览表明，与标准启发式方法相比，二量子位门数减少了 20-50%。

为了进一步优化执行多个独立作业时的吞吐量，我们引入了批处理模式——一种新的执行模式，相对于单个作业提交，执行时间提高了 5 倍。此外，对于公用事业规模的迭代工作负载，我们发布了扩展会话，它允许将多个会话组合在一起，以无缝地启用高级量子经典工作负载。

IBM 推出了Qiskit Patterns，这是一个概述量子程序结构的编程模板，以及用于大规模构建量子算法和应用程序的逻辑框架。利用 Qiskit 模式提供的可组合性、容器化和抽象性，用户可以从一组基础构建块无缝创建量子算法和应用程序，并使用异构计算基础设施（例如 Quantum Serverless）执行这些模式。这允许对现有企业规模工作流程进行有针对性的量子加速，并提供对量子电路和操作员的抽象。IBM 宣布通过 Qiskit Patterns 部署 Quantum Serverless 作为 Beta 版，以实现大规模托管、无人值守的模式执行。

watsonx 上的量子生成人工智能

为了更好地简化量子开发流程，IBM 率先通过IBM 企业 AI 平台watsonx使用生成式 AI 进行量子代码编程。我们演示了通过 watsonx 提供的生成式 AI 如何帮助自动开发 Qiskit 的量子代码。我们通过对IBM Granite 200 亿参数代码基础模型的微调来实现这一目标。

查看到 2033 年的扩展 IBM 量子开发路线图，以及到 2029 年的 IBM 量子创新路线图。到 2033 年的路线图十年来的量子创新。该路线图强调了我们的处理器和系统能够执行的门数量的改进。该路线图以 Heron 在 2024 年达到 5,000 个门的目标为起点，规划了多代处理器，每代处理器都利用质量的改进来实现越来越多的门数。

然后，在 2029 年，我们遇到了一个拐点：使用我们的 Starling 处理器在 200 个量子位上执行 1 亿个门，并采用基于新颖的纠错技术毛是一打的单位。阅读有关近期量子计算机的纠错码的更多信息。总代码。接下来是 Blue Jay，这是一个到 2033 年能够在 2,000 个量子位上执行 10 亿个门的系统。自 2016 年我们将第一个设备放在云端以来，这意味着执行的门数增加了 9 个数量级。我们的新创新路线图将演示通过 l-、m- 和 c-耦合器实现 Gross 代码所需的技术，分别由 Flamingo、Crossbill 和 Kookaburra 处理器演示。

IBM 运行了3万亿个线路。

A noisy quantum computer is capable of producing accurate expectation values on 127 qubits, surpassing the capabilities of brute force classical computation.

[1]Y. Kim et al., “Evidence for the utility of quantum computing before fault tolerance,” Nature, vol. 618, no. 7965, Art. no. 7965, Jun. 2023, doi: 10.1038/s41586-023-06096-3.

heron 可以用了

Tunable couplers - a fundamental shift powering our future roadmap

Qiskit 1.0 Now with increased performance, stability, and reliability.

使用了AI 减少双Qubits门数

Reducing 2-qubit gates using a reinforcement learning (RL) AI transpiler involves the application of RL techniques to optimize quantum circuits. This process is often referred to as "qubit routing," which modifies quantum circuits to satisfy the connectivity constraints of a target quantum computer. This involves inserting SWAP gates into the circuit so that the logical gates only ever occur between adjacent physical qubits, with the goal of minimizing the circuit depth added by the SWAP gates[1][8].

A study by Pozzi et al. proposed a qubit routing procedure that uses a modified version of the deep Q-learning paradigm. The system was able to outperform the qubit routing procedures from two of the most advanced quantum compilers currently available, on both random and realistic circuits, across near-term architecture sizes[1].

IBM has also developed a new Transpiler Service based on Qiskit and new AI models trained using reinforcement learning. The new AI Transpiler passes produce highly optimized circuits, with a 20-50% improvement in circuit depth and CNOT count compared to equivalent Qiskit transpiler passes[5].

To implement such a system, you would need a solid understanding of reinforcement learning and its key concepts, such as action, reward, environment, and agent[3][7]. You would also need to be familiar with deep learning libraries like TensorFlow or Keras, and quantum computing libraries like Qiskit[5][7].

In terms of reinforcement learning algorithms, the AlphaZero agent is a notable example. It guides a Monte Carlo tree search (MCTS) planning procedure using a deep neural network. The input to the neural network is the state, and the output is a policy and value prediction. The policy prediction is a distribution over actions, and the value function is a prediction of the cumulative returns that the agent should expect to receive from the current state[2].

In conclusion, reducing 2-qubit gates using a reinforcement learning AI transpiler involves the application of RL techniques to optimize quantum circuits, with the goal of minimizing the circuit depth added by the SWAP gates. This is a complex task that requires a deep understanding of both quantum computing and reinforcement learning.

Citations: [1] https://arxiv.org/pdf/2007.15957.pdf [2] https://www.nature.com/articles/s41586-023-06004-9 [3] https://youtube.com/watch?v=cO5g5qLrLSo&t=45 [4] https://dl.acm.org/doi/pdf/10.5555/3455716.3455818 [5] https://www.ibm.com/quantum/blog/qiskit-patterns [6] https://link.springer.com/article/10.1007/s10462-022-10205-5 [7] https://blog.paperspace.com/getting-started-with-reinforcement-learning/ [8] https://dl.acm.org/doi/full/10.1145/3520434 [9] https://www.bu.edu/caadlab/Shahzad22.pdf [10] https://youtube.com/watch?v=bD6V3rcr_54 [11] https://www.nature.com/articles/s42256-023-00687-5 [12] https://paperswithcode.com [13] https://neptune.ai/blog/best-reinforcement-learning-tutorials-examples-projects-and-courses [14] https://www.nature.com/articles/s42005-021-00684-3 [15] https://github.com/zwang4/awesome-machine-learning-in-compilers [16] https://www.sliceofexperiments.com/p/an-actually-runnable-march-2023-tutorial [17] https://esynctechnologies.com/updates/f/chapter-4-using-the-transpiler [18] https://arxiv.org/pdf/1805.03441.pdf [19] https://www.baeldung.com/cs/reinforcement-learning-neural-network [20] https://iopscience.iop.org/article/10.1088/1367-2630/abe0ae [21] https://blog.salesforceairesearch.com/coderl/ [22] https://www.toptal.com/machine-learning/deep-dive-into-reinforcement-learning [23] https://link.aps.org/pdf/10.1103/PhysRevResearch.3.043088 [24] https://tomassetti.me/language2language-transformers-machine-learning-to-build-transpilers/ [25] https://www.javatpoint.com/reinforcement-learning

Using reinforcement learning (RL) for qubit routing in quantum compilers offers several benefits:

Performance Improvement: RL-based qubit routing procedures have been shown to outperform the routing procedures from some of the most advanced quantum compilers currently available, on both random and realistic circuits, across a range of near-term architecture sizes[1][2][3].

Dynamic Adaptation: Unlike traditional methods that rely on static heuristics, RL methods can dynamically adapt to the problem at hand. This makes them more flexible and capable of handling a wider range of scenarios[4].

Efficiency: RL methods can potentially reduce the overall execution time of quantum compiling, improving the tradeoff between the length of the sequence and execution time, potentially allowing real-time operations[5].

Noise-Adaptive: RL strategies can be adapted to handle noise, a critical aspect in real-world quantum hardware. This makes them more robust and practical for real-world applications[7].

Future-Proof: As quantum computing technology evolves, the ability of RL to learn and adapt makes it a future-proof strategy for qubit routing[1][2][3].

In conclusion, the use of reinforcement learning for qubit routing in quantum compilers can lead to performance improvements, increased efficiency, and greater adaptability to noise and future technological advancements.

Citations: [1] https://dl.acm.org/doi/abs/10.1145/3520434 [2] https://arxiv.org/abs/2007.15957 [3] https://dl.acm.org/doi/fullHtml/10.1145/3520434 [4] https://arxiv.org/pdf/2007.15957.pdf [5] https://www.nature.com/articles/s42005-021-00684-3 [6] https://www.researchgate.net/publication/343390184_Using_Reinforcement_Learning_to_Perform_Qubit_Routing_in_Quantum_Compilers [7] https://repositorio-aberto.up.pt/bitstream/10216/153420/3/646720.pdf [8] https://openreview.net/pdf/db1b0c030120604e61cd4ab2ebac4947d6fc6653.pdf