广告
加载中

量枢QSim:跑出世界领先性能,让更多设备承接更大规模的模拟任务

龚作仁 2026-11-19 10:30
龚作仁 2026/11/19 10:30

邦小白快读

EN
全文速览

总1:性能与规模是核心干货。

1.大规模随机线路测试中,六个规模里五个取得参测方案最短耗时;35量子比特耗时114秒,业内领先模拟器200秒,减少约43%。

2.单GPU完成35量子比特模拟,多卡多机拓展到38量子比特;8G显存RTX 3070借助内存完成需32G数据的模拟,数据量达显存4倍。

总2:实操门槛和可用方式值得关注。

1.把内存当大仓库、显存当高速工作台,分批搬运并连续计算,34量子比特任务从462秒降到50.3秒,加速约9.17倍。

2.统一架构支持英伟达与AMD显卡、Arm与x86处理器;计划2027年第一季度通过量枢云提供完整模拟器服务,用户提交程序后由平台安排计算并返回结果,还可自动选择模拟方法。

总1:技术领先与跨硬件适配构成产品卖点。

1.大规模随机线路测试六个规模中五个取得最短耗时;35量子比特114秒,业内领先模拟器200秒,减少约43%,形成可传播的性能优势。

2.统一架构支持英伟达与AMD显卡、Arm与x86处理器,覆盖家用设备到超算平台;内存与显存协同扩大规模,单GPU完成35量子比特,多卡多机到38量子比特。

总2:产品化与云服务渠道值得观察。

1.基于科研版模拟器做性能调优与工程化完善,形成面向科研和行业应用的商业版本,把执行效率、可处理规模和跨硬件支持整合进同一架构。

2.计划2027年第一季度通过量枢云提供完整模拟器服务,用户提交程序后平台安排计算并返回结果,减少寻找硬件、搭建环境和反复调优的准备时间。

3.量枢云已有面向石油、电力、金融、材料等领域的开箱即用应用;研究者可按任务和条件选择家用显卡、工作站或超算,个人开发者和资源有限团队可用现有设备扩大实验。

总1:增长市场与正面机会较为明确。

1.量子模拟正成为高性能计算的重要应用方向;2026年10月英伟达披露在德国于利希超算中心JUPITER超算上开展大规模量子计算模拟,于利希也公开介绍利用处理器与GPU协同内存扩大模拟规模。

2.更快模拟速度和更大可处理规模,让团队在相同时间与资源条件下探索更多方案;已有内存可转化为更大模拟容量,减少部分任务单纯为存储数据而增加显卡的需求,个人开发者和资源有限团队受益。

总2:商业模式、可学习点与风险提示值得留意。

1.量枢QSim计划2027年第一季度通过量枢云提供完整模拟器服务,用户提交程序后由平台安排计算并返回结果,平台还可根据程序特点和结果要求自动选择模拟方法。

2.量枢云已提供面向石油、电力、金融、材料等领域的开箱即用应用,并进一步连接行业问题、量子算法和底层算力。

3.可学习点是数据少搬家、硬件多干活;风险在于充分释放性能需围绕具体硬件和任务细致调优,单机单卡、单机多卡、多机多卡的计算安排、内存使用和设备协作各有侧重,往往需大量测试与反复调试。

总1:产品设计与生产需求集中在架构协同。

1.模拟器需同时规划计算方法、数据位置和设备分工;先根据程序特点和结果要求选择适合的模拟方法,对需要完整保存量子状态的任务再优化每批数据的计算与流转。

2.把适合的量子运算转化为矩阵计算,结合成熟高性能工具;为电脑处理器、英伟达GPU和AMD GPU分别开发、优化底层计算程序,上层共享任务组织方式,底层发挥各类硬件长处。

总2:商业机会与数字化启示可落到现有设备利用。

1.内存与显存协同扩大模拟规模,单GPU完成35量子比特,多卡多机到38量子比特;8G显存RTX 3070借助电脑内存完成需要保存32G数据的量子程序模拟,数据量达显存容量4倍。

2.计划2027年第一季度通过量枢云提供完整模拟器服务,用户提交程序后平台安排计算并返回结果,自动选择模拟方法;量枢云已有石油、电力、金融、材料等领域开箱即用应用。

3.数字化启示是提前规划、连续执行,记住位置、延后整理,就近处理、减少远程调取,把省下的等待转化为更多有效计算。

总1:行业趋势与新技术方向清晰。

1.量子模拟正在成为高性能计算的重要应用方向;2026年10月英伟达披露在德国于利希超算中心JUPITER超算上开展大规模量子计算模拟,于利希也公开介绍利用处理器与GPU协同内存扩大模拟规模。

2.中科量枢基于科研版模拟器,经性能调优与工程化完善形成商业版本,并将执行效率、可处理规模和跨硬件支持整合进同一套架构;统一架构支持英伟达与AMD显卡,以及Arm与x86两类处理器。

总2:客户痛点与解决方案对应明确。

1.痛点是线路规模扩大后完整量子态数据呈指数增长,迅速触及单张显卡显存上限,仅提高GPU计算速度也难以扩大规模。

2.方案是把电脑内存当大仓库、GPU显存当高速工作台,分批搬运并连续完成多步计算,摊薄往返搬运成本;34量子比特任务从462秒降至50.3秒,加速约9.17倍。

3.性能上35量子比特114秒,较业内领先模拟器200秒减少约43%;单GPU完成35量子比特,多卡多机到38量子比特;计划2027年第一季度通过量枢云提供完整模拟器服务,减少找硬件、搭环境和反复调优时间,并自动选择模拟方法。

总1:平台需求与问题集中在硬件环境和调优成本。

1.用户充分释放模拟器性能,需要围绕具体硬件和任务进行细致调优;单机单卡、单机多卡、多机多卡在计算安排、内存使用和设备协作上各有侧重,往往需要大量测试与反复调试。

2.不同程序特点和结果要求需要匹配不同模拟方法;寻找硬件、搭建环境和反复调优会占用实验与验证前的准备时间。

总2:平台最新做法与运营管理思路可归纳为云化承接。

1.量枢QSim计划于2027年第一季度通过量枢云提供完整的模拟器服务,用户提交程序后由平台安排计算并返回结果。

2.平台可根据程序特点和结果要求,自动选择适合的模拟方法,提升云端虚拟量子处理器的运行效率,让用户把精力集中在算法和问题本身。

3.量枢云已提供面向石油、电力、金融、材料等领域的开箱即用应用,随着高性能模拟能力接入,进一步连接行业问题、量子算法和底层算力;量枢QSim向下组织处理器、显卡、内存和集群资源,向上支撑算法研究、云端虚拟量子处理器与行业应用。

4.风向规避点是跨硬件、跨任务调优复杂,需围绕具体硬件和任务持续测试调试。

总1:产业新动向与新问题值得跟踪。

1.量子模拟正成为高性能计算的重要应用方向;2026年10月英伟达披露在德国于利希超算中心JUPITER超算上开展大规模量子计算模拟,研究未来量子计算机怎样更可靠地工作;于利希也公开介绍利用处理器与GPU协同内存扩大模拟规模的工作。

2.新问题是线路规模扩大后完整量子态所需数据量呈指数增长,迅速触及单张显卡显存上限,需同时提高执行效率、扩大模拟规模并适应不同硬件。

总2:技术路径与商业模式提供启示。

1.量枢QSim把适合的量子运算转化为矩阵计算,结合产业界数十年积累的算法、软件库与芯片优化;执行策略包括提前规划连续执行、记住位置延后整理、就近处理减少远程调取。

2.团队为电脑处理器、英伟达GPU和AMD GPU分别开发优化底层计算程序,上层共享任务组织方式,底层发挥各类硬件长处;单GPU完成35量子比特,多卡多机到38量子比特。

3.商业模式上,科研版经调优与工程化形成商业版本,计划2027年第一季度通过量枢云提供完整模拟器服务,用户提交程序后平台安排计算并返回结果,自动选择模拟方法;量枢云已覆盖石油、电力、金融、材料等领域应用,启示是将现有计算资源更充分转化为量子研发能力。

返回默认

声明:快读内容全程由AI生成,请注意甄别信息。如您发现问题,请发送邮件至 run@ebrun.com 。

我是 品牌商 卖家 工厂 服务商 平台商 研究者 帮我再读一遍。

Quick Summary

Summary 1: Performance and scale are the core takeaways.

1. In large-scale random circuit tests, the tested solution achieved the shortest runtime in five of six problem sizes; 35 qubits took 114 seconds, about 43% less than the 200 seconds achieved by the industry-leading simulator.

2. A single GPU completed a 35-qubit simulation, while multi-GPU, multi-node setups scaled to 38 qubits; an 8GB RTX 3070 used system memory to run a simulation requiring 32GB of data—four times its VRAM.

Summary 2: Practical barriers and usage modes are worth noting.

1. Treating system memory as a large warehouse and VRAM as a high-speed workbench, it moves data in batches and performs continuous computation, cutting a 34-qubit task from 462 seconds to 50.3 seconds—about 9.17x faster.

2. A unified architecture supports Nvidia and AMD GPUs as well as Arm and x86 processors. It plans to offer a full simulator service via Liangshu Cloud in Q1 2027: users submit programs, the platform schedules computation and returns results, and it can automatically select simulation methods.

Summary 1: Technology leadership and cross-hardware adaptability form the product value proposition.

1. In large-scale random circuit tests, it achieved the shortest runtime in five of six problem sizes; 35 qubits took 114 seconds, about 43% less than the 200 seconds achieved by the industry-leading simulator, creating a communicable performance advantage.

2. The unified architecture supports Nvidia and AMD GPUs and Arm and x86 processors, covering devices from home PCs to supercomputing platforms; coordinated system memory and VRAM expand scale, with a single GPU completing 35 qubits and multi-GPU, multi-node systems reaching 38 qubits.

Summary 2: Productization and cloud service channels are worth watching.

1. Building on the research edition, performance tuning and engineering refinement have produced a commercial version for research and industry applications, integrating execution efficiency, manageable scale, and cross-hardware support into one architecture.

2. It plans to offer a full simulator service via Liangshu Cloud in Q1 2027; after users submit programs, the platform schedules computation and returns results, reducing time spent finding hardware, setting up environments, and repeatedly tuning.

3. Liangshu Cloud already offers out-of-the-box applications for oil and gas, power, finance, materials, and other fields; researchers can choose consumer GPUs, workstations, or supercomputers based on tasks and conditions, while individual developers and resource-constrained teams can expand experiments using existing equipment.

Summary 1: The expanding market and positive opportunities are relatively clear.

1. Quantum simulation is becoming an important application area in high-performance computing; in October 2026, Nvidia disclosed large-scale quantum computing simulations on the JUPITER supercomputer at the Jülich Supercomputing Centre in Germany, and Jülich also publicly described using coordinated processor and GPU memory to expand simulation scale.

2. Faster simulation speeds and larger manageable scale allow teams to explore more options under the same time and resource constraints; existing memory can be converted into greater simulation capacity, reducing the need for some tasks to add GPUs solely for data storage, benefiting individual developers and resource-constrained teams.

Summary 2: The business model, lessons, and risks deserve attention.

1. Liangshu QSim plans to offer a full simulator service via Liangshu Cloud in Q1 2027; after users submit programs, the platform schedules computation and returns results, and can automatically select simulation methods based on program characteristics and output requirements.

2. Liangshu Cloud already provides out-of-the-box applications for oil and gas, power, finance, materials, and other fields, further connecting industry problems, quantum algorithms, and underlying compute.

3. The lesson is to move less data and keep hardware busier; the risk is that fully releasing performance requires careful tuning around specific hardware and tasks. Single-GPU, single-node multi-GPU, and multi-node multi-GPU configurations differ in computation scheduling, memory use, and device coordination, often requiring extensive testing and repeated debugging.

Summary 1: Product design and production requirements center on architectural coordination.

1. The simulator must simultaneously plan simulation methods, data placement, and device division of labor; it first selects a suitable simulation method based on program characteristics and output requirements, then optimizes the computation and flow of each data batch for tasks requiring full quantum state retention.

2. It converts suitable quantum operations into matrix computations and combines them with mature high-performance tools; it separately develops and optimizes low-level compute programs for CPUs, Nvidia GPUs, and AMD GPUs, sharing task organization at the upper layer while leveraging each hardware type's strengths at the lower layer.

Summary 2: Commercial opportunities and digitalization lessons can be applied to using existing equipment.

1. Coordinated system memory and VRAM expand simulation scale: a single GPU completes 35 qubits, and multi-GPU, multi-node systems reach 38 qubits; an 8GB RTX 3070 used system memory to run a quantum program simulation requiring 32GB of data—four times its VRAM capacity.

2. It plans to offer a full simulator service via Liangshu Cloud in Q1 2027; after users submit programs, the platform schedules computation and returns results and automatically selects simulation methods; Liangshu Cloud already has out-of-the-box applications in oil and gas, power, finance, materials, and other fields.

3. The digitalization lesson is to plan ahead and execute continuously, remember locations and defer consolidation, process locally to reduce remote fetches, and convert saved waiting time into more useful computation.

Summary 1: Industry trends and new technology directions are clear.

1. Quantum simulation is becoming an important application area in high-performance computing; in October 2026, Nvidia disclosed large-scale quantum computing simulations on the JUPITER supercomputer at the Jülich Supercomputing Centre in Germany, and Jülich also publicly described work using coordinated processor and GPU memory to expand simulation scale.

2. Zhongke Liangshu built a commercial version based on its research edition through performance tuning and engineering refinement, integrating execution efficiency, manageable scale, and cross-hardware support into one architecture; the unified architecture supports Nvidia and AMD GPUs and both Arm and x86 processors.

Summary 2: Customer pain points and solutions map clearly to each other.

1. The pain point is that as circuit scale grows, complete quantum state data grows exponentially and quickly hits the VRAM limit of a single GPU; improving GPU compute speed alone is also insufficient to expand scale.

2. The solution treats system memory as a large warehouse and GPU VRAM as a high-speed workbench, moving data in batches and continuously completing multiple computation steps to amortize round-trip data movement; a 34-qubit task dropped from 462 seconds to 50.3 seconds, about 9.17x faster.

3. On performance, 35 qubits take 114 seconds, about 43% less than the 200 seconds achieved by the industry-leading simulator; a single GPU completes 35 qubits, and multi-GPU, multi-node systems reach 38 qubits. It plans to offer a full simulator service via Liangshu Cloud in Q1 2027, reducing time spent finding hardware, setting up environments, and repeatedly tuning, while automatically selecting simulation methods.

Summary 1: Platform requirements and problems center on hardware environments and tuning costs.

1. For users to fully release simulator performance, careful tuning is needed around specific hardware and tasks; single-GPU, single-node multi-GPU, and multi-node multi-GPU setups differ in computation scheduling, memory use, and device coordination, often requiring extensive testing and repeated debugging.

2. Different program characteristics and output requirements need matching simulation methods; finding hardware, setting up environments, and repeated tuning consume preparation time before experiments and validation.

Summary 2: The platform's latest practices and operational management approach can be summarized as cloud-based delivery.

1. Liangshu QSim plans to provide a complete simulator service via Liangshu Cloud in Q1 2027; after users submit programs, the platform schedules computation and returns results.

2. The platform can automatically select suitable simulation methods based on program characteristics and output requirements, improving the operating efficiency of cloud-based virtual quantum processors and letting users focus on algorithms and the problems themselves.

3. Liangshu Cloud already provides out-of-the-box applications for oil and gas, power, finance, materials, and other fields. As high-performance simulation capabilities are connected, it further links industry problems, quantum algorithms, and underlying compute; Liangshu QSim orchestrates processors, GPUs, memory, and cluster resources downward and supports algorithm research, cloud-based virtual quantum processors, and industry applications upward.

4. The risk-mitigation point is that cross-hardware and cross-task tuning is complex and requires continuous testing and debugging around specific hardware and tasks.

Summary 1: New industry developments and new questions are worth tracking.

1. Quantum simulation is becoming an important application area in high-performance computing; in October 2026, Nvidia disclosed large-scale quantum computing simulations on the JUPITER supercomputer at the Jülich Supercomputing Centre in Germany, studying how future quantum computers can operate more reliably; Jülich also publicly described work using coordinated processor and GPU memory to expand simulation scale.

2. The new question is that as circuit scale grows, the data required for a complete quantum state grows exponentially and quickly hits the VRAM limit of a single GPU, requiring simultaneous improvements in execution efficiency, simulation scale, and adaptability to different hardware.

Summary 2: The technical path and business model offer lessons.

1. Liangshu QSim converts suitable quantum operations into matrix computations, combining decades of industry-accumulated algorithms, software libraries, and chip optimization; execution strategies include planning ahead for continuous execution, remembering locations and deferring consolidation, and processing locally to reduce remote fetches.

2. The team separately develops and optimizes low-level compute programs for CPUs, Nvidia GPUs, and AMD GPUs, sharing task organization at the upper layer while leveraging each hardware type's strengths at the lower layer; a single GPU completes 35 qubits, and multi-GPU, multi-node systems reach 38 qubits.

3. On the business model, the research edition has been tuned and engineered into a commercial version. It plans to offer a full simulator service via Liangshu Cloud in Q1 2027; after users submit programs, the platform schedules computation and returns results and automatically selects simulation methods. Liangshu Cloud already covers applications in oil and gas, power, finance, materials, and other fields, suggesting that existing compute resources can be more fully converted into quantum R&D capacity.

Disclaimer: The "Quick Summary" content is entirely generated by AI. Please exercise discretion when interpreting the information. For issues or corrections, please email run@ebrun.com .

I am a Brand Seller Factory Service Provider Marketplace Seller Researcher Read it again.

量子算法走向真实硬件之前,往往需要先在模拟器中反复验证:用现有计算机推演量子程序,检查算法设计、比较参数方案、研究设备误差的影响。模拟速度、可处理规模和硬件选择,直接关系到研发迭代的效率。GPU(显卡)加速是提高模拟效率的重要手段,但随着线路规模扩大,完整量子态所需的数据量呈指数增长,迅速触及单张显卡的显存上限。如何同时提高执行效率、扩大模拟规模,并适应不同硬件,成为量子模拟走向实用的重要课题。

针对这些挑战,中科量枢研发了量枢QSim量子线路模拟器。它基于科研版模拟器,经进一步性能调优与工程化完善,形成面向科研和行业应用的商业版本,并将执行效率、可处理规模和跨硬件支持整合进同一套架构。高性能计算(HPC)平台的一组大规模随机线路测试显示,量枢QSim在六个规模中的五个取得参测方案的最短记录耗时;在AMD GPU的四个共同测试规模中,也全部取得最短记录耗时。

除了提速,量枢QSim还通过内存与显存协同扩大模拟规模:单GPU测试已完成35量子比特模拟,多卡、多机协作进一步拓展到38量子比特;家用显卡也能借助电脑内存承接超出显存容量的任务。其统一架构支持英伟达与AMD显卡,以及Arm与x86两类处理器,覆盖从家用设备到超算平台的多种计算资源。

01 大规模模拟,跑出领先性能

模拟器的性能,最终要通过实际任务检验。随着量子线路规模扩大,软件能否充分发挥硬件的计算与存储能力,决定了任务的执行效率。

在相同硬件配置下的一组大规模线路测试中,量枢QSim在六个规模中的五个取得参测方案的最短记录耗时。以35量子比特任务为例,量枢QSim记录耗时114秒,业内领先模拟器为200秒,较后者减少约43%。此时,完整状态数据已超过单张GPU的显存容量。

另一篇架构论文和家用显卡补充测试,进一步展示了这套架构在不同硬件上的表现。

这些结果让跨硬件支持具备了实际的性能价值。研究者可以根据任务需求和现有条件,在家用显卡、工作站与超算之间选择计算资源,更高效地推进实验。

02 显存装不下的任务,也能用上GPU加速

量子模拟的规模,首先受制于存储容量。量子比特是量子计算处理信息的基本单位;在完整模拟中,每增加一个量子比特,需要保存的数据量就会翻倍。一旦完整数据超出显存容量,仅靠提高GPU的计算速度,也难以扩大模拟规模。

量枢QSim把电脑内存当作“大仓库”,把GPU显存当作“高速工作台”:完整数据保存在仓库中,每次只将当前需要的一批搬上工作台,连续完成多步计算后写回,再处理下一批。

关键是让每批数据进入显存后,尽可能多地完成计算,摊薄往返搬运的成本。量枢QSim通过统一安排数据搬运与运算顺序,让扩大容量后的任务继续发挥GPU的加速优势。

在单GPU测试中,量枢QSim已完成35量子比特的完整模拟,状态数据总量超过该GPU的显存容量。家用主机也验证了这一思路:8G显存的RTX 3070借助电脑内存,完成了需要保存32G数据的量子程序模拟,数据量达到显存容量的4倍。

数据如何存放和调度,同样影响大任务的运行速度。论文对量枢QSim自身的两种执行方式进行了比较:在同一个34量子比特任务中,通过主动安排数据搬运和计算,记录耗时从462秒降至50.3秒,约加速9.17倍。

通过这种方式,已有内存可以转化为更大的模拟容量,减少部分任务单纯为存储数据而增加显卡的需求。对个人开发者和资源有限的团队而言,现有设备可以承接更多实验,研究范围也随之扩大。

03把计算机数十年的积累,用到量子模拟

这些加速成绩背后,是一套同时规划计算方法、数据位置和设备分工的架构。量枢QSim先根据程序特点和结果要求选择适合的模拟方法;对于需要完整保存量子状态的任务,再优化每批数据的计算与流转。

矩阵乘法是科学计算与人工智能加速的重要基础,可以把它理解为对大批数字进行成组运算。围绕这类计算,产业界已积累了数十年的算法、软件库与芯片优化。量枢QSim将适合的量子运算转化为矩阵计算,再结合自身的执行策略,让量子模拟用上这些成熟的高性能工具。

第一,提前规划、连续执行。把后续计算安排成一组衔接的工序,让一批数据进入显存后连续用于多步运算,减少反复搬运。

第二,记住位置、延后整理。数据临时换位后,先记录新位置,让后续计算沿用当前的数据排列,等确有需要时再整理归位。

第三,就近处理、减少远程调取。多设备协作时,提前安排接下来要用的数据,让更多计算在本地完成。

数据少搬家,硬件多干活;把省下的等待,转化为更多有效计算。

团队还为电脑处理器(CPU)、英伟达GPU和AMD GPU分别开发、优化底层计算程序:上层共享任务组织方式,底层发挥各类硬件的长处,让跨平台适配与性能优化相互配合。

04从科研级能力,走向可以直接使用的云服务

量子模拟正在成为高性能计算的重要应用方向。2026年10月,英伟达披露,在德国于利希超算中心的JUPITER超算上开展大规模量子计算模拟,研究未来量子计算机怎样更可靠地工作;于利希也公开介绍了利用处理器与GPU协同内存扩大模拟规模的工作。

这一方向的产业价值,在于将现有计算资源更充分地转化为量子研发能力。算法设计、程序调试、设备误差研究和软硬件协同验证,都依赖反复运行、比较和迭代。更快的模拟速度和更大的可处理规模,让团队能够在相同的时间与资源条件下探索更多方案。

源自中国科学院计算技术研究所科研积累的中科量枢,正围绕程序开发、运行、算法与应用构建配套软件。量枢QSim为这套体系提供底层执行能力:向下组织处理器、显卡、内存和集群资源,向上支撑算法研究、云端“虚拟量子处理器”与行业应用。

充分释放模拟器的性能,需要围绕具体硬件和任务进行细致调优。无论是单机单卡、单机多卡,还是多机多卡,计算安排、内存使用和设备协作都各有侧重,往往需要大量测试与反复调试。为了让用户更便捷地获得这些经过调优的模拟能力,量枢QSim计划于2027年第一季度通过量枢云提供完整的模拟器服务。用户提交程序后,由平台安排计算并返回结果,减少寻找硬件、搭建环境和反复调优的准备时间,更快开始实验与验证。

平台还可根据程序特点和结果要求,自动选择适合的模拟方法,提升云端“虚拟量子处理器”的运行效率。用户可以更便捷地调用不同计算能力,将精力集中在算法和问题本身。

量枢云已提供面向石油、电力、金融、材料等领域的开箱即用应用。随着高性能模拟能力接入,平台将进一步连接行业问题、量子算法和底层算力,帮助行业团队开展方法筛选与原型验证,为研究成果走向应用提供更便捷的入口。

通过量枢QSim与量枢云的衔接,中科量枢希望把模拟速度、任务规模和硬件适配上的进展,转化为更多团队可以直接使用的研发能力,让量子算法和行业应用探索更容易开始、更高效地迭代。

注:文/龚作仁,文章来源:Laborer,本文为作者独立观点,不代表亿邦动力立场。

文章来源:Laborer

广告
微信
朋友圈

FAQ回顾

量枢QSim是什么?它主要用于什么场景?

量枢QSim是中科量枢研发的量子线路模拟器,基于科研版模拟器调优形成面向科研和行业应用的商业版本。它用于量子算法走向真实硬件前的模拟验证,支持算法设计、参数比较和设备误差研究,并通过GPU加速提高研发迭代效率。

量枢QSim在量子模拟性能上有哪些具体表现?

在高性能计算平台的大规模随机线路测试中,量枢QSim在六个规模中的五个取得参测方案最短记录耗时;在AMD GPU的四个共同测试规模中全部取得最短记录耗时。35量子比特任务耗时114秒,较业内领先模拟器的200秒减少约43%。

显卡显存不足时如何完成更大规模量子模拟?

量枢QSim采用内存与显存协同:完整量子态数据存放在电脑内存,按需分批搬入GPU显存计算并写回,通过统一安排数据搬运与运算顺序摊薄搬运成本。8G显存RTX 3070借助电脑内存完成了需保存32G数据的量子程序模拟,数据量达显存容量4倍。

量枢QSim支持哪些硬件和计算平台?

量枢QSim的统一架构支持英伟达与AMD显卡,以及Arm与x86两类处理器,覆盖从家用设备到超算平台的多种计算资源。团队还为CPU、英伟达GPU和AMD GPU分别开发、优化底层计算程序,使跨平台适配与性能优化相互配合。

量枢云将如何提供量子模拟服务?

量枢QSim计划2027年第一季度通过量枢云提供完整模拟器服务。用户提交程序后平台安排计算并返回结果,减少寻找硬件、搭建环境和反复调优时间;平台可根据程序特点和结果要求自动选择模拟方法。量枢云已提供石油、电力、金融、材料等领域开箱即用应用。

这么好看,分享一下?

朋友圈 分享

APP内打开

赞 +1
+1
微信好友 朋友圈 新浪微博 QQ空间
关闭
收藏成功
发送
/140 0