Hello world! March 11, 2022 Heading 1 Heading 2 Heading 3 Heading 4 Heading 5 Heading 6 Sorry, this video requires cookie consent. Please accept marketing-cookies to watch this video.
GlobalFoundries Announces Next Generation in Silicon Photonics Solutions and Collaborates with Industry Leaders to Advance a New Era of More in the Data Center March 7, 2022First-of-its-kind silicon photonics platform available now addresses explosive growth of soaring data volumes while significantly reducing power consumption MALTA, N.Y., March 7, 2022 – GlobalFoundries Inc (Nasdaq: GFS) (GF) today announced it is collaborating with industry leaders including Broadcom, Cisco Systems, Inc, Marvell and NVIDIA, along with breakthrough photonic leaders including Ayar Labs, Lightmatter, PsiQuantum, Ranovus and Xanadu, to deliver innovative, unique, feature-rich solutions to solve some of the biggest challenges facing data centers today. The more than 42 billion connected IoT devices generating ~177 ZB of data annually1 combined with the rise of power consumption in data centers, is driving the need for innovative solutions to move and compute data faster and more energy efficiently. These key market trends and implications have catalyzed GF’s focus on groundbreaking semiconductor solutions that harvest the potential of photons, instead of electrons to move data and position GF to continue to be the manufacturing leader in the optical networking module market that is projected to grow at a CAGR of 26% between 2021 and 2026, reaching about USD 4 billion by 2026.2 Today GF is proud to announce GF Fotonix™, its next generation, widely disruptive silicon photonics platform. GF has active design wins with major customers, significant market share today and expects its growth in this segment to outpace the market. GF also announced it is partnering with industry leader Cisco Systems, Inc., on a custom silicon photonics solution for DCN and DCI applications, including an interdependent Process Design Kit (PDK) in close collaboration with our GF manufacturing services team. Customers and Partner support for GF Silicon Photonics Solutions “We’re working closely with GlobalFoundries to design high-bandwidth, low-power optical interconnects for some of our leading-edge data center products. NVIDIA interconnect solutions manufactured with the monolithic GF Fotonix platform will boost high performance computing and AI applications, enabling breakthrough advances.” Edward Lee, vice president of Mixed-Signal Design, NVIDIA “As one of our trusted semiconductor partners across a broad range of technologies and process nodes, we are happy to see Global Foundries extend their investments for enabling a photonics ecosystem across components and integrated solutions,” Liming Tsau, vice president of Foundry Engineering, Broadcom Inc. “The demands of today and tomorrow’s networking and communications infrastructure are requiring higher performance technologies for the design and manufacture of our optical transceiver modules,” said Bill Gartner, senior vice president and general manager of Optical Systems and Optics Group, Cisco Systems, Inc. “Our heavy investment and leadership in silicon photonics, combined with GF’s feature rich manufacturing technology, allows us to deliver best in class products.” “Marvell continues to lead the industry with the highest performance transimpedance amplifiers and modulator drivers for next-generation optical connectivity solutions for the cloud data centers and carrier markets,” said Dr. Loi Nguyen, executive vice president, Optical and Copper Connectivity Group, Marvell. “GF’s latest silicon germanium (SiGe) technology enables us to achieve the high bandwidth speeds and power efficiencies that our customers require to meet their ever-increasing data demands.” “Since our earliest days, Ayar Labs and GlobalFoundries have partnered on the development of GF Fotonix, from incorporating our PDK requirements and process optimizations to demonstrating the first working silicon on the platform,” says Charles Wuischpard, CEO, Ayar Labs. “The combination of our leading monolithic electronic/photonic solution and GF Fotonix unlocks a tremendous market opportunity for chip-to-chip optical I/O and sets the stage for us to deliver substantial volume shipments by year end.” “At Lightmatter, our technology delivers compute that is faster and more efficient than anything else in the market—results like these aren’t possible with conventional chips,” said Nicholas Harris, CEO of Lightmatter. “Our next-generation technology is made possible by GlobalFoundries’ best-in-class photonic foundry technology and together we’re changing the way the world thinks about photonics. This is just the start.” “Our telecoms, defense and data center customers need innovative new ways to transmit, connect and compute data at the speed of light,” said Martin Zirngibl, vice president and general manager at Macom. “GlobalFoundries offers the features in its silicon photonics platform that can be leveraged to scale communications to the next level.” “We are leveraging GF’s new Fotonix™ platform to develop custom silicon photonics chips that meet our advanced quantum computing requirements,” said Fariba Danesh, chief operating officer of PsiQuantum. “We are delighted to share our multi-disciplinary silicon-photonics IP cores and chiplets, and advanced packaging solutions with our customers who are driving the adoption of novel data center architectures based on integrating best-in-class chiplets and co-packaged optics,” said Hojjat Salemi, chief business development officer of RANOVUS. “Our close collaboration with GlobalFoundries underlines our joint commitment to deliver a fully featured set of qualified IP cores and chiplets with OSAT-ready high-volume manufacturing flows and supporting ecosystem to enable the huge potential of monolithic silicon photonics.” “Many chips, operating in parallel and networked together, are required to process the large number of qubits involved in fault-tolerant quantum computing algorithms,” said Zachary Vernon, head of hardware at Xanadu. “Leveraging an existing advanced 300mm platform like GF Fotonix gives us a huge advantage in the race to deliver a useful, error-corrected quantum computer.” “Silicon photonics is now recognized as a necessary technology for the datacenter revolution, and our leading semiconductor manufacturing technology platform accelerates adoption into the mainstream,” said Amir Faintuch, senior vice president and general manager of Compute and Wired Infrastructure at GF. “GF Fotonix is a feature-rich platform that addresses the most urgent, complex and difficult challenges in areas such as optical networking, super and quantum computing, fiber-to-the-home (FTTH), 5G networks, aeronautics and defense.” GF solutions to move and compute data at speed of light GF Fotonix is a monolithic platform, the first in the industry to combine its differentiated 300mm photonics features and 300GHz-class RF-CMOS on a silicon wafer, delivering best-in-class performance at scale. GF Fotonix consolidates complex processes that were previously distributed across multiple chips onto a single chip by combining a photonic system, radio frequency (RF) components and high-performance complementary metal–oxide–semiconductor (CMOS) logic on a single silicon chip. GF is the only pure-play foundry with a 300mm monolithic silicon photonics solution that has demonstrated the industry’s highest data rate per fiber (0.5Tbps/fiber). This enables 1.6-3.2Tbps optical chiplets, which deliver faster, more efficient transmission of data, more efficiently with better signal integrity. In addition, the up to 10,000x improvement in system error rate enables next generation artificial intelligence (AI). GF Fotonix enables the highest level of integration onto a photonics integrated circuit (PIC) so customers can integrate more product functions and simplify their bill of materials (BOMs). End customers can realize greater performance through increased capacity and capability. The new solution also enables innovative packaging solutions, such as the passive attachment for larger fiber arrays, support for 2.5D packaging, and on-die lasers. GF Fotonix solutions will be manufactured at the company’s advanced manufacturing facility in Malta, N.Y., with the PDK 1.0 available in April 2022. EDA partners Ansys, Cadence Design Systems, Inc., and Synopsys provide design tools and flows to support GF’s customers and their solutions. GF provides customers with reference design kits, MPWs, testing, pre- and post-fab, turnkey and semiconductor manufacturing services to help customers get to market faster. In addition, for customers needing discrete, high-performance RF solutions for optical systems, GF also announced it is adding new features onto the GF SiGe platform. High-performance silicon germanium (SiGe) solutions from GF are designed to deliver the speed and bandwidth required to transport information through next generation fiber optic high speed networks. Quotes from supporting EDA partners For more information visit gf.com/siliconphotonics About GF GlobalFoundries® (GF®) is one of the world’s leading semiconductor manufacturers. GF is redefining innovation and semiconductor manufacturing by developing and delivering feature-rich process technology solutions that provide leadership performance in pervasive high growth markets. GF offers a unique mix of design, development and fabrication services. With a talented and diverse workforce and an at-scale manufacturing footprint spanning the U.S., Europe and Asia, GF is a trusted technology source to its worldwide customers. For more information, visit www.gf.com. Contact: Shannon Love GlobalFoundries (480) 225-7325 [email protected]
More Compute Power, Better Connectivity: Combining Proven CMOS Chip Fabrication with Silicon Photonics to Move Data at the Speed of Light March 7, 2022In our previous blog, we looked at the rapidly shifting landscape for semiconductors. The explosion of data to be created, shared and connected demands more efficient chips that consumer less power. To satisfy these demands in the data center as well as in emerging sensing and computing applications including artificial intelligence (AI), the metaverse, streaming, gaming and social media, hardware designers and customers are looking for an almost-fantastical list of requirements: flexibility to mix and match chip capabilities in a variety of chip sizes and packages, more compute power with less power consumption, affordability and faster data transmission. As part of the acquisition of IBM Microelectronics in 2015, GF acquired photonics know-how and IP from IBM Research, and has since refined and developed ways to bring it to volume production as the market demand grew for more capability, better power efficiency and higher bandwidth. Now, GF has added two new high-performance solutions to its portfolio: the GF FotonixTM photonics platform and new, improved features to our already available silicon germanium (SiGe) 9HP platform. GF Fotonix, available from the GF Fab 8 in Malta, New York, monolithically integrates RF, digital and silicon photonics circuits on the same chip, while leveraging the scale, efficiencies and optimization of 300 millimeter silicon manufacturing. GF’s high-volume silicon manufacturing—combined with innovative 2.5D and 3D stacking—enables ultra-high integration of electrical, optical and communication capabilities into a single monolithic design. This solution lets customers integrate more product functions on-chip and simplifies the bill of materials. Key features include: Low loss passive components (such as SiN and Si WGs, tapers, MMIs) High performance active photonic components (such as MZM, MRM and GePDs) High performance active and passive RFCMOS components Digital standard cell library V-grooves supporting passive fiber attach For customers needing discrete high-performance component solutions for optical transceivers, GF is also announcing new features for the GF SiGe (9HP) portfolio. High-performance silicon germanium (SiGe) solutions from GF are designed to deliver the speed and bandwidth required to transport information through fiber optic high speed networks. This new feature set enables customers to integrate higher performance digital and RF functionality and exploit silicon economies of scale. This option will enable 800Gbps data rates to reduce power consumption of data center optical connections. This new feature set is also ideal for 5G and future 6G and similar telecom and cellular applications. GF’s SiGe platform offers maximized bandwidth and functionality, while minimizing power consumption and costs. GF is providing an Electro-Optical Process Design Kit (PDK) that includes p-cells for photonic and electrical components to help its customers get started designing for the speed of light. Reference designs, design services and post-fab services will be provided as well, along with a digital standard cell library. GF is collaborating with industry leaders including Broadcom, Cisco Systems, Inc, Marvell and NVIDIA, along with breakthrough photonic computing leaders including Ayar Labs, Lightmatter, PsiQuantum, Ranovus and Xanadu to deliver innovative, unique, feature-rich solutions to solve some of the biggest challenges facing data centers today. Learn more here. To help customers get their designs to market more efficiently, EDA leaders Ansys, Cadence Design Systems, Inc. and Synopsys, Inc. are offering design tools supporting integrated silicon photonics- based chips and chiplets. More info is available here.
Ranovus Delivers World’s First Monolithic 100G Optical I/O Cores for Next-Generation Data Centers based on the GF Fotonix™ Platform March 7, 2022
Ansys Collaborates with GlobalFoundries to Deliver Next-Gen Silicon Photonics Solutions to Advance New Era of Datacenters March 7, 2022
A Perfect Storm: Bigger Data, Power Consumption and Bandwidth Driving Alternatives to Traditional CMOS Chipmaking Process March 3, 2022 Almost five billion Internet users are using, creating and sharing ginormous amounts of data. Even as the amount and types of data increase, the number of opportunities to create and share data has exploded across devices ranging from home security systems, appliances, gaming systems, computers and phones to huge data centers that handle social media, streaming content, games and enterprise applications. A study by Ericcson notes there will be more than 42 billion connected IoT devices generating ~177 ZB of data (by 2026). One report notes that 2.5 quintillion bytes of data are created every day and 90 percent of the world’s data has been created in the last two years. Much of that data has been driven by increases in almost four billion users of social media sites worldwide. Not surprising, given remote learning and work-from-home, the ongoing pandemic has only further increased the world’s hunger for more data -and more bandwidth to share the data. According to one industry report, home data usage increased 38 percent from March 2019 to March 2020. The same report found that work-from-home increased during the pandemic from an average of 17 percent of workers to 44 percent, putting increased strain on networks and increasing data usage. Covid-19 also resulted in an increase of 138 percent in a group called power users consuming more than 1 terabyte of Internet data quarterly. Google alone counts more than 63,000 searches every second, or 5.6 billion daily searches. Sorry, this video requires cookie consent. Please accept marketing-cookies to watch this video. Amir Faintuch, SVP & GM, Computing & Wired Infrastructure SBU at GF explains why silicon photonics is a vital technology platform for the data revolution. These factors do not begin to comprehend potential impacts of the emerging metaverse, which will demand the creation, storage and connectivity of even more data – data that will need to be transmitted with very low latency at high speeds. The metaverse joins AI, machine learning and virtual reality as well as the continued expansion of connected devices in driving data creation and transmission. In the data center, power consumption has become a key consideration along with bandwidth. Historically, the chip industry has relied on electrical connectivity over metal (copper) connections for interconnects between systems. Electrical SerDes (serial deserializers), the most common form of electrical I/O, is reaching its limits and there is no attainable roadmap beyond 112 Gb/sec because the large signal losses in copper-based interconnects at a board level make it difficult to transmit data further than a few centimeters at such a high data rate. The next wave of high-performance computing architectures requires a new form of I/O that avoid the bottlenecks created by electrical I/O. By 2028, most data center short-distance physical interconnects will be optical instead of electrical. The pluggable module has been the key widget that converts electrical signals to optical signals and vice versa i.e. it is the electro-optic interface. There have been two key advantages of pluggable modules: Standardization and interoperability – data center operators can source modules from multiple vendors which has driven down the “per Gbps” cost through innovation and competition. Modularity – data center operators can use short range optics to traverse to the end of a row of racks in a data center, and long-range optics to go longer distances to a different data center. This modularity has also driven the form factor and standardization of the switch boxes and the box faceplates into which these pluggable modules are plugged into. In a typical switch box at the top of a data center rack, multiple pluggable modules are plugged into the faceplate. The conversion from optical to electrical signals occurs at the faceplate with high-speed electrical signals having to traverse the Cu traces on the board to the switch ASIC. Optical communications solutions are poised to enable new levels of performance in hyperscale data centers, cloud computing and 5G-driven network transformation. The silicon photonics technology, used for optical communications, will also become the foundation for rapidly emerging compute and sensing applications. Look for our next installment in this series to understand the role of new options to combine the proven benefits of CMOS technology with new capabilities for more powerful chips based on silicon photonics.
Microsoft Executive Bobby Yerramilli-Rao Joins GlobalFoundries Board of Directors March 3, 2022Executive with deep understanding of technology strategy brings experience leading corporate growth MALTA, N.Y., March 3, 2022 – GlobalFoundries Inc. (Nasdaq: GFS) (GF) today announced the appointment of Bobby Yerramilli-Rao as an independent director of the company’s board of directors. Dr. Yerramilli-Rao is currently chief strategy officer and corporate vice president of corporate strategy at Microsoft, where he is responsible for developing and driving growth-oriented strategies. Dr. Yerramilli-Rao’s appointment is effective immediately and expands the GF board to 11 directors, including five independent directors. He will be a member of the Strategy & Technology Committee. “We are delighted to welcome Bobby to GF’s board of directors,” said Ahmed Yahia Al Idrissi, board chair of GF. “As technology becomes more pervasive in our everyday life, and as GF embarks on the next phase of its growth journey to better serve our customers, Bobby’s distinctive perspectives and capabilities will be a welcome addition to our board. His experience in leadership roles in software, bioscience, communications and other fast-growing sectors will certainly help shape GF’s next chapter of innovation and growth.” “As GF continues to focus on redefining innovation in semiconductor technology and manufacturing, Bobby broadens our strategic and technology expertise to advance our long-term growth strategy,” said Tom Caulfield, CEO of GF. “Bobby brings a deep understanding of technology strategy and solutions execution that will help catalyze value creation and expand our reach to meet the future needs of our customers.” Prior to joining Microsoft in 2020, Dr. Yerramilli-Rao served as co-founder and managing partner of Fusion Global Capital, which focused on accelerating the growth of emerging software and cloud computing companies. Before that, he served as corporate strategy director of Vodafone Group responsible for strategy globally and acquisitions related to digital services. He spent more than a decade in the at McKinsey & Company rising to partner and co-leader of the TMT practice. Dr. Yerramilli-Rao is currently a member of the board of directors of Cambridge Epigenetix, a privately held biosciences company. He received a master’s degree in Electrical Engineering from the University of Cambridge and a Doctorate in Robotics from the University of Oxford. More information about GF’s board of directors can be found here. About GF GlobalFoundries® (GF®) is one of the world’s leading semiconductor manufacturers. GF is redefining innovation and semiconductor manufacturing by developing and delivering feature-rich process technology solutions that provide leadership performance in pervasive high growth markets. GF offers a unique mix of design, development and fabrication services. With a talented and diverse workforce and an at-scale manufacturing footprint spanning the U.S., Europe and Asia, GF is a trusted technology source to its worldwide customers. For more information, visit www.gf.com. Contact: Erica McGill GlobalFoundries (518) 795-5240 [email protected]
格芯®（GF®）瞄准高能效人工智能 March 1, 2022近年来，人工智能(AI)技术取得了长足的进步，从少数应用中的有限使用发展为各种系统的重要驱动技术，现已渗透到我们生活的方方面面。 “智能”恒温器、门铃和语音助手；半自动驾驶汽车；具有预测能力的医疗监测设备；以及许多领域的众多其他应用现在都依赖于AI技术。 然而，AI及其专用子集（机器学习、深度学习和神经形态计算）存在一个致命弱点，那就是电能需求巨大且不断增长，这阻碍了其进一步发展。随着AI的计算要求越来越高及其整体使用的日益普及，AI计算和数据传输所需的电能迅速增加，进而导致了电能资源的过度使用和全球碳足迹的大幅增加。 这种电能使用增长是不可持续的。以大量使用AI的数据中心为例。2017年，美国的数据中心用电量约占全国总用电量的3%，而到2020年，这一数字翻了一番，达到6%，而且仍看不到尽头。根据行业预测，如果依旧采用当今低效的计算架构，从理论上来说，数据中心到2041年将用掉全球所有的发电量。 2020年，数据中心用电量约占全美总用电量的6% AI能源挑战不仅限于数据中心。位于网络边缘的电池供电型物联网(IoT)设备，其整体电力需求也非常大。随着更多AI处理向边缘迁移，日益复杂的物联网设备必须变得更加高效，这样其锂离子电池才能为更多功能供电、续航时间更长且/或体积更小。这也有助于减少因废弃电池而带来更多的潜在危险锂离子废弃物。 为了应对AI能源挑战，格芯(GF)对其产品路线图做出了调整，将一系列技术创新融入其12LP/12LP+FinFET解决方案（用于数据中心和物联网边缘服务器）和22FDX® FD-SOI解决方案（用于物联网边缘）。此外，格芯还携手领先的AI研究人员，共同开发更高效的新型计算架构和算法，从而打开AI新世界的大门。 格芯正在致力于解决数据中心和电池供电物联网边缘设备（现已渗透到日常生活）不断增长的能源需求 AI的范式变革 AI系统会收集大量结构化或非结构化数据，然后根据为给定应用编写的算法对其进行处理。其目标是在数据中找到相应的关联性和模式，以此为依据做出推理和决策，并以满足应用需求的方式基于这些推理采取行动。鉴于数据集的大小和算法的复杂性，需要密集的计算机处理。 格芯无线基础架构(CWI)战略业务部首席技术官兼计算和副总裁Ted Letavic表示：“目前，大多数AI任务都在云中运行，但馈入云端算法的数据集来自外部世界，并采用边缘物联网设备等模拟接口传输。基于云的AI范式能效低下，因为它需要将大量数据从网络边缘（物联网边缘）传输到数据中心，在数据中心执行计算并推导出结果，然后再将结果传输回边缘设备。这种方法不仅能效低下，而且与数据传输相关的时间也会导致系统的整体延迟，因此无法用于许多安全关键型AI应用。” 起初，AI和机器学习使用传统的通用中央处理器(CPU)。Letavic表示：“这些CPU原本是为随机存储器访问而设计的，鉴于需要不断减少在处理器和存储器之间传输数据所需的时间和能耗，这种设计带来了很多问题。我们需要改变该范式，在存储数据的存储器网络内部处理数据，而无需进行数据传输。” 他指出，计算架构因此正在发生根本性的转变。一场向特定领域计算架构演进的“设计复兴”正在拉开帷幕，这些架构对于数据流和计算路径定义明确的AI推理（训练）任务来说非常节能。这些优化的加速器类似于存储器层次结构，通常称为“数字存内计算”或“模拟存内计算”。这些加速器执行并行操作，使其成为AI核心计算类型的理想选择，并且大幅降低了总功耗，从而能够在网络边缘更充分地利用AI。 格芯12LP+使存储器效率提高4倍 为了适应架构上的这些变化，格芯进行了技术改进并启用了新的设计流程。 Letavic表示：“在我们研究的几乎每一种AI工作负载中，存储器带宽和存储器访问功率都限制了整体性能，因为必须在固定的功率预算内完成一定数量的操作，而且存储器消耗了太多的功率。因此，我们将从7nm技术开发工作中得到的一些经验应用到我们的12LP/LP+技术中，推出了支持1 GHz的0.55V SRAM存储器宏，对于典型工作负载而言，它将与存储器访问相关的能耗降低到了原来的四分之一。该解决方案针对脉动阵列处理器，可直接用于处理AI和机器学习工作负载。” Letavic指出，格芯接下来研究了阵列架构。 “我们发现，每个客户都有不同的数据流架构，基本上没办法选择一个最佳设计。”他表示，“为了解决这个问题，我们创建了一个将逻辑和存储元件合成在一起的新颖设计流程，使它们可以非常接近地进行构建，并具有高度的灵活性。这种设计流程打破了逻辑和存储器宏合成的传统范式，这种逻辑和存储元件的混合可用于实现非常新颖的AI架构。” 格芯推出的差异化12LP+解决方案针对人工智能训练和推理应用进行了优化 Letavic指出，格芯的先进技术与新型的独特设计和合成流程相结合，构成了实现全新计算范式的强大工具，并进一步开启了AI时代。格芯正携手领先研究机构，推动该领域的重要科研工作。 Marian Verhelst博士和格芯的大学联系项目 格芯正在与一些全球领先的研究人员合作，研究这些创新架构，并为其确立客观效益和佐证点，从而让格芯的客户可以利用它们来设计更高效的AI系统。 这些研究工作大多通过与IMEC等研究联盟合作展开，以及通过格芯大学合作计划(UPP)与大学教授合作展开。在该计划下，格芯与全球学术研究人员密切合作，开展利用格芯技术的创新项目。 Marian Verhelst博士是格芯的主要学术合作人员之一，她是比利时鲁汶大学的教授，同时也是Imec的研究主任。Verhelst博士是高效处理架构的全球权威专家之一。她之前曾在美国英特尔实验室工作，从事数字增强模拟和射频电路研究，并于2012年加入鲁汶大学，并创立了一个研究实验室，该实验室目前拥有16名博士生和博士后研究人员。 她的实验室科研项目涵盖各方面，从欧盟资助的长期宏观项目，到涉及向广泛从业者进行技术转让的中短期研究。她曾获得比利时André Mischke YAE奖，该奖项旨在表彰国际领先的学术研究、管理和循证决策成就。 她作为比利时青年学院和佛兰德STEM平台的前成员，是科学与教育的大力倡导者，并曾登上比利时国家电视台多个科普类节目的专访。2014年，她创立了InnovationLab，旨在为高中教师及高中生开发交互式工程项目。她也是IEEE“Women in Circuits”倡导计划的成员之一，并积极参与许多其他宣传和教育活动。 DIANA芯片——AI向前迈进的重要一步 Verhelst博士致力于研发混合神经网络芯片，该芯片不仅是全球首款将模拟存内计算和数字脉动阵列结合到一起的芯片，而且还可以在这些异构资源之间无缝划分AI算法，以实现最佳能耗性能、准确性和延迟。 该芯片名为DIANA（DIgital和ANAlog，即数字和模拟），在格芯的22FDX平台上构建，相关的论文将在本月末举行的极具声望的2022年国际固态电路会议(ISSCC)上发表。 Verhelst表示：“机器学习正在蓬勃发展，每家企业都有一个针对机器学习优化的处理器，但大多数情况下，它们都是纯粹在数字领域中设计的，使用0和1进行计算，这并不总是能实现最高效率。因此，许多研究人员现在正在研究模拟领域中的计算，甚至在SRAM存储器内部，使用各个SRAM单元之间的电流累积而不是0和1。从电能角度来看，这将更有效，从芯片密度的角度来看也是如此，因为它允许在每平方毫米上进行更多的计算。” “到目前为止，我们已经取得了一些不错的成果，但仅适用于恰好与存储器形状完美匹配的特定机器学习网络。对于其他网络来说，算法不一定能有效运行。”她补充道，“DIANA芯片包含一个主机处理器以及一个数字和模拟存储器协处理器。对于神经网络的每一层，它都可以将指定层分派给推理加速器或协处理器，以确保尽可能高效地运行。所有操作都是并行运行，中间数据在各层之间有效共享。” 为了实现这一目标，Verhelst的团队开发了先进的调度程序和映射程序，用于分析芯片的硬件特性，以确定最优能效或最优延迟的“计算顺序”，即如何在芯片上运行给定算法。 “算法运行可以采用很多方法，具体取决于存储器大小、它的特性、处理阵列中有多少计算元件等。”她表示，“因此，我们开发了一些工具，您可以在其中输入硬件特性，并帮助您根据工作负载找到适合的最佳解决方案。” 正在进行的合作 DIANA芯片是Verhelst与格芯的最新合作成果，该次合作大约始于五年前，当时格芯为她的一名博士生提供了机会，使用22FDX技术流片视频处理芯片，该芯片可以高效并行执行数百个操作。 格芯的22FDX边缘AI加速器经优化可缩短延迟和可操作性响应时间，通过在边缘管理数据来增强安全性和数据隐私 此外，Verhelst还使用格芯的12 LP+技术，为高度密集的计算结构构建了深度学习芯片，该芯片包含超过2,000个乘法器和大量SRAM内容。另一个处于初始阶段的项目是使用格芯的22FDX平台构建一个高占空比的机器学习芯片，专注于超低功耗运行，面向物联网、机器监控或其他须以毫瓦级功率运行的传感器节点。 她指出，格芯提供的芯片和技术合作伙伴关系非常宝贵。她表示：“生产功能完备的芯片成本极为昂贵，尤其是对于体积很大的数字处理器。与格芯合作既为我们降低了芯片门槛，又为我们提供了获得最新相关IP的途径。” “此外，格芯还为我们提供建议和支持，解决有时候遇到的物理设计收敛工作难题，对于如此先进的技术，该工作不再是小问题。在后端需要考虑的事情有很多，当我们试图确保快速IO、出色的振荡器、最佳电源门控等性能时，格芯的制造经验确实对我们很有帮助。” 展望 当被问及格芯在更高效AI领域的下一步举措时，Letavic提到了公司在计算芯片本身的集成电压调节以及用于更高水平传输和计算效率的硅光子学方面的研发工作。 他表示：“改进供电是一种弥补较小节点功率扩展不足的方法，这已成为系统层面的真正限制。要节省应用总功耗，关键方法之一就是提高向处理器内核提供电流和电压的效率。我们正在探索各种可选方案，鉴于格芯在双极性CMOS和DMOS功率器件方面的悠久传统，这对我们来说会是一个巨大的商机。” Letavic还提到，光子加速，即使用光（光子）替代电（电子），不仅可以通过光纤传输信号，还可用于计算本身，将会在AI中发挥重要作用。“我想说这种技术发展速度比我预期的要快得多。这是我们已有一些大学明确参与合作的另一个领域。” 阅读其他通过格芯大学合作计划开展的研究项目： 格芯推动新一代汽车雷达发展 与学界合作助力格芯加快奠定6G领先地位 格芯携手领先研究机构共同推动6G技术研发
