Summary of Lightmatter Passage Conference Call Industry and Company Overview - **Industry**: AI and Photonic Computing - **Company**: Lightmatter, known for its Passage M1000 "superchip" platform utilizing photonic technology to enhance AI training capabilities [1][3][13] Core Points and Arguments 1. **Exponential Growth of AI Models**: The scale of AI models has increased dramatically, with models now reaching hundreds of billions or even trillions of parameters, necessitating thousands of GPUs for training [3][4] 2. **Challenges in AI Training**: The industry faces significant challenges in scaling AI training, particularly due to the slowdown of Moore's Law and the limitations of traditional electrical interconnects, which create bottlenecks in data communication and synchronization [7][10][11] 3. **Lightmatter's Solution**: The Passage M1000 platform addresses the interconnect bottleneck by employing a 3D photonic stacking architecture, integrating up to 34 chiplets on a single photonic interposer, achieving a total die area of 4,000 mm² [13][14] 4. **Unprecedented Bandwidth**: The Passage platform delivers a total bidirectional bandwidth of 114 Tbps and 1,024 high-speed SerDes lanes, allowing each chiplet to access multi-terabit-per-second I/O bandwidth, effectively overcoming traditional I/O limitations [17][21] 5. **Comparison with Competitors**: Lightmatter's approach contrasts with other industry players like NVIDIA and Cerebras, who focus on maximizing single-chip performance or building ultra-large chips. Lightmatter emphasizes optical interconnects to achieve high bandwidth communication across chiplets [30][42][44][52] Additional Important Insights 1. **Nature Paper Validation**: A study published in *Nature* demonstrated the feasibility of photonic processors for executing advanced AI models, achieving near-electronic precision, which complements Lightmatter's focus on interconnect solutions [22][23][82] 2. **Future of AI Acceleration**: The combination of Lightmatter's optical interconnects and the advancements in photonic computing suggests a paradigm shift towards hybrid electronic-photonic architectures, breaking through performance ceilings in AI acceleration [82][83] 3. **Scalability and Efficiency**: Lightmatter's Passage aims to simplify AI deployments and improve efficiency by collapsing datacenter-level communication into a single "superchip," potentially offering better cost efficiency and flexibility compared to traditional methods [42][52][78] Conclusion - The emergence of Lightmatter's Passage platform represents a significant advancement in addressing the challenges of modern AI training, providing a breakthrough pathway through innovative photonic interconnect technology [84]

Core Viewpoint - The article discusses advancements in optical chip technology presented at the Hot Chips 2025 conference, highlighting innovations from companies like NVIDIA, Celestial AI, and Ayar Labs, which aim to enhance data center performance and efficiency through silicon photonics and optical interconnects [2][75][236]. Group 1: NVIDIA's Innovations - NVIDIA emphasizes the need for co-packaged optics to significantly enhance AI factory scalability, noting that AI factories consume approximately 17 times more optical power than traditional cloud data centers [3][5]. - The Spectrum-X Ethernet technology aims to reduce the cost of network photonics, which constitutes about 10% of the total computing power cost in AI factories [3][16]. - Spectrum-X Ethernet photonics is claimed to be the first to implement 200 G/channel SerDes technology, improving signal integrity and reducing power consumption by decreasing the number of lasers required for high-speed links [16][34]. Group 2: Celestial AI's Photonic Fabric - Celestial AI showcases its Photonic Fabric module, which utilizes optical connections to link next-generation GPUs and accelerators, moving away from traditional electrical connections [75][78]. - The company is developing its own optical MAC (OMAC) to achieve high energy efficiency by matching SerDes with channels [80][92]. - Celestial AI's technology aims to optimize the integration of optical I/O within chip designs, enhancing performance and reducing power consumption [92][94]. Group 3: Ayar Labs' Optical I/O Solutions - Ayar Labs introduces a UCIe optical I/O retimer, designed to facilitate the integration of optical I/O into chip packaging, achieving data rates of up to 8 Tbps [108][122]. - The company addresses the challenge of scaling interconnects for large clusters of chips, which is critical for AI applications [111][117]. - Ayar Labs' technology aims to decouple optical and electrical signal transmission, allowing for larger systems compared to traditional electrical I/O [162][169]. Group 4: Lightmatter's Passage M1000 - Lightmatter presents the Passage M1000, which integrates co-packaged optics and silicon photonics, promising bandwidths of up to 114 Tbps [170][184]. - The design focuses on compact optical I/O, utilizing silicon micro-rings for efficient modulation [191][196]. - Lightmatter aims to enable optical circuit switching for redundancy and scalability in high-performance computing environments [211][224]. Group 5: OpenLight's Photonic Solutions - OpenLight Photonics aims to accelerate the transition to silicon photonics with its proprietary photonic application-specific integrated circuits (PASIC), designed for high-density, low-power optical interconnects [234][236]. - The company emphasizes customization in its PASIC offerings, allowing clients to tailor components to specific applications, enhancing flexibility in rapidly evolving markets [237][239]. - OpenLight's technology is positioned to address the growing demand for faster data transmission in AI data centers, with plans to expand its component library and production capabilities [238][239].

来源：内容 编译自 IEEE ，谢谢。 如果将过多的铜线捆扎在一起，最终会耗尽空间——前提是它们不会先熔合在一起。人工智能数据 中心在GPU和内存之间传输数据的电子互连方面也面临着类似的限制。为了满足人工智能的海量 数据需求，业界正在转向更大尺寸、更多处理器的芯片，这意味着在机架内实现更密集、更长距离 的连接。初创公司正在展示 GPU 如何摆脱铜互连，用光纤链路取而代之。 光纤链路对数据中心来说并不陌生。它们使用可插拔收发器在机架之间传输数据，将电信号转换为 光信号。为了提高能源效率，"将光学元件集成到芯片封装中一直是一个梦想，"加州大学圣巴巴拉 分校电气工程教授克林特·肖( Clint Schow)表示。这就是共封装光学器件（CPO），科技巨头们正 在全力支持它。英伟达 (Nvidia) 最近宣布量产一款网络交换机，该交换机使用嵌入在与交换机同 一基板上的光子调制器。"这震惊了整个行业，"加州桑尼维尔初创公司Avicena的首席执行官巴迪 亚·佩泽什基 (Bardia Pezeshki) 表示。 哥伦比亚大学电气工程教授、Xscape Photonics联合创始人Keren Bergman解释说， Nvid ...

Core Viewpoint - Lightmatter has launched two silicon photonic interconnect products to meet the increasing bandwidth demands of AI deployments, specifically targeting high-bandwidth multi-chip switches for XPU applications [1][2]. Group 1: Product Overview - The first product, Passage M1000, is an optical intermediary layer designed to facilitate high-bandwidth communication between stacked ASIC or GPU chips, with a total bandwidth capacity of up to 14.25 TB/s [2]. - The M1000 utilizes 56 Gb/s NRZ modulation and wavelength division multiplexing, allowing each fiber to support eight wavelengths, resulting in a bandwidth of 56 GB/s per fiber [2]. - Lightmatter plans to release a pair of smaller co-packaged optical designs in 2026 following the M1000 launch [2]. Group 2: Additional Products - The Passage L200 and L200X are designed as more traditional co-packaged optical devices, promising bidirectional bandwidths of 32 Tb/s and 64 Tb/s, respectively [3]. - The L200 uses 56 Gb/s NRZ, while the L200X employs 112 Gb/s PAM4 SerDes, with both series utilizing a 3D packaging method to support external communication speeds exceeding 200 Tb/s [3]. - These chips incorporate various technologies from Alphawave Semi, including low-power, low-latency UCIe and optical-ready SerDes [3].