ASCENT: Collaborative Research: Scaling Distributed AI Systems based on Universal Optical I/O

我们的社会正迅速变得依赖于基于神经网络的人工智能计算。每天都有新的算法被发明出来，增加了推理和训练的内存和计算需求。这种爆炸性的增长为分布式机器学习（ML）训练和推理创造了巨大的需求。OpenAI的估计表明，自2012年以来，计算需求每两年稳定增长100倍，这比我们在过去半个世纪中通过半导体行业的摩尔定律实现的计算改进速度快50倍。这种新的计算需求已经部分地通过硬件加速器和软件栈的快速发展来满足，以支持这些专门的计算。硬件加速器提供了大量的加速，但今天的训练任务仍然需要几天甚至几周的时间。原因是：随着工作者（例如，计算节点）的数量增加，每个工作者的计算时间减少，但是节点之间的通信需求增加，从而在计算节点之间的互连中产生瓶颈。未来的分布式机器学习系统将要求每个节点的互连带宽提高1-2个数量级，这迫切需要全新的方法来构建分布式机器学习系统的互连。该提案旨在通过开发基于推进集成电子和光子技术的可扩展互连解决方案，为扩展分布式ML计算创建一个新的范例，该解决方案能够实现直接的节点到节点光纤连接。拟议的跨堆栈协作多学科工作将使教育和培训一批独特的工程师和科学家，这些工程师和科学家跨越了机器学习、网络和电子光子系统和设备的界限，这些都是迫切需要的。主要研究人员与高中生直接接触，在伯克利无线研究中心和麻省理工学院的妇女技术项目提供暑期实习，以及在波士顿大学的示范性本科生研究活动的良好记录。PI实施的教育和推广活动将确保从K-12到本科和研究生学习以及继续劳动力教育的新一代领导者的早期接触和持续培训，特别关注代表性不足的学生。互连已成为实现分布式ML充分潜力的关键瓶颈。未来的机器学习工作负载可能需要每个设备数十Tbps的带宽。只有通过新的通用I/O才能实现逻辑连接、物理分布式计算在机架、机架和行规模上的无处不在部署，该通用I/O能够以封装内互连的能量、延迟和带宽密度实现插座到插座的通信。目前还没有这样的技术。基于硅光子学的光学I/O有可能解决这一关键挑战，但仍需要从芯片制造到路由算法的根本性进步，以确保这些互连系统的可扩展性。为了实现高带宽密度和能量效率，必须使用密集波分复用。需要高效的基于环形谐振器的调制器和梳状激光源来实现每个光纤连接上的Tbps速率和从10秒到100秒Tbps的插座带宽缩放。需要新的链路架构，如所提出的激光转发相干链路，以实现高效率的外部集中式梳状激光源，每个光纤端口每个波长的功率适中（低于mW）。拟议的工作还将开发新的调度算法、网络架构和工作负载并行策略，以利用通用光学I/O的带宽密度和低延迟，将具有大量数据集的大型AI工作负载映射到可扩展的分布式计算系统。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查进行评估，被认为值得支持的搜索.

Our society is rapidly becoming reliant on neural networks based artificial intelligence computation. New algorithms are invented daily, increasing the memory and computational requirements for both inference and training. This explosive growth has created an enormous demand for distributed machine learning (ML) training and inference. Estimates by OpenAI illustrate the steady growth of computational requirements of 100x every two years since 2012, which is a 50x faster than the rate of computation improvements enabled previously through Moore’s Law of semiconductor industry that we have enjoyed in the last half-century. This new computation demand has been partly met by rapid development of hardware accelerators and software stacks to support these specialized computations. Hardware accelerators have provided a significant amount of speed-up but today’s training tasks can still take days and even weeks. The reason for this: as the number of workers (e.g. compute nodes) increases, the computation time per worker decreases, but the communication requirements between the nodes increase, creating a bottleneck in the interconnect between the compute nodes. Future distributed ML systems will require 1-2 orders of magnitude higher interconnect bandwidth per node, creating a pressing need for entirely new ways to build interconnects for distributed ML systems. This proposal aims to create a new paradigm for scaling distributed ML computation, by developing a scalable interconnect solution based on advancing the integrated electronics and photonics technology that enables direct node-to-node optical fiber connectivity. The proposed cross-stack collaborative multi-disciplinary work will enable the education and training of a unique crop of engineers and scientists that cross the boundaries of machine learning, networking, and electronic-photonic systems and devices, which are in severe demand. The principal investigators have an established track record of direct engagement with high-school students providing summer internships at Berkeley Wireless Research Center and MIT’s Women’s Technology Program, as well as exemplary undergraduate research activities at Boston University. The educational and outreach activities the PIs have put in place will ensure early exposure and continued training of new generation of leaders in this field, from K-12, through undergraduate and graduate studies, and continuing workforce education, with special focus on underrepresented students.The interconnect has emerged as the key bottleneck in enabling the full potential of distributed ML. Future ML workloads are likely to require tens of Tbps of bandwidth per device. Ubiquitous deployment of logically-connected, physically distributed computation across shelf, rack and row scale can only be enabled by a new universal I/O that enables socket to socket communication at the energy, latency and bandwidth density of in-package interconnects. No such technology currently exists. Silicon-photonics based optical I/O has the potential to address this critical challenge, but fundamental advances–from chip manufacturing to routing algorithms–are still needed to ensure the scalability of these interconnect systems. To enable high-bandwidth density and energy-efficiency, dense wavelength division multiplexing must be used. High-efficiency ring resonator-based modulators and comb laser sources are needed to enable Tbps rates over each fiber connection and socket bandwidth scaling from 10s to 100s of Tbps. New link architectures like the proposed laser-forwarded coherent link are needed to enable high-efficiency external centralized comb laser sources with modest (sub-mW) power per wavelength per fiber port. The proposed work will also develop new scheduling algorithms, network architectures, and workload parallelism strategy to leverage the bandwidth density and low-latency of the universal optical I/O, to map large AI workloads with massive datasets to a scalable distributed compute system.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.