Equity Research 25 June 2025 U.S. Semiconductors & Semiconductor Capital Equipment Laying Out the Scale-Up Architecture Landscape Interconnect is more critical than ever for the success of XPU efforts, and this is no more prevalent than in scale-up designs where dominance is being fought today amongst three factions (UALink, Scale Up Ethernet, NVLink) with major implications for the future of AI. We appreciate your 5-star vote in the 2025 Extel All-America Research Survey in the Semiconductors & Semiconduct ...

Core Insights - Astera Labs (ALAB) is strategically positioned in next-generation AI and cloud infrastructure, showcasing a strong start to 2025 with a focus on technological leadership in interconnect standards [1][2] Group 1: Product and Technology Leadership - Astera Labs has established a first-mover advantage with a comprehensive suite of PCIe Gen 6 solutions, including retimers, smart gearboxes, optical modules, and fabric switches, tailored to meet the high performance and signal integrity requirements of modern AI racks [2] - The Leo CXL product family addresses the increasing demand for memory expansion and pooling, supporting both AI and general-purpose compute workloads as data center CPUs adopt CXL 2.0 and 3.0 [2][3] - The introduction of UALink 1.0 in early 2025 is expected to lead to commercial solutions by 2026, creating a scalable, open interconnect for AI accelerators and unlocking a multibillion-dollar opportunity [3] Group 2: Competitive Positioning - Compared to Marvell Technology, which has strengths in custom ASICs and broader market focus, Astera Labs benefits from tighter integration of hardware and software through its COSMOS suite, enhancing AI-specific system observability and fleet management [4] - Broadcom, while a scale leader in PCIe switches and networking silicon, offers less tailored solutions for AI rack-scale deployments, allowing Astera Labs to maintain agility in addressing emerging infrastructure needs like UALink and NVLink-based clustering [5] Group 3: Market Performance and Valuation - Astera Labs' shares have increased by 26.2% over the past three months, outperforming the industry growth of 10.2% and the sector's growth of 7.5%, while the S&P 500 index rose by 3.6% during the same period [6][8] - The company is currently trading at a forward 12-month price-to-sales ratio of 19.19X, slightly below its one-year median of 19.86X, but remains overvalued compared to the industry [9]

Core Insights - Marvell Technology (MRVL) is enhancing its position in AI infrastructure by expanding its custom chip capabilities, integrating new components to improve performance and scalability across large-scale systems [1][6] Financial Performance - In Q1 FY26, Marvell reported record Data Center revenues of $1.44 billion, representing a 76% increase year over year, driven by the rapid scaling of custom AI silicon [2][10] - Marvell's forward price-to-sales ratio is 7.36X, which is lower than the industry average of 8.15X [13] Strategic Developments - Marvell partnered with NVIDIA in May 2025 to offer NVLink Fusion technology, enhancing the flexibility of its custom cloud platform silicon for next-generation AI infrastructure [3] - The introduction of a new multi-die packaging solution based on proprietary interposer technology aims to improve die-to-die interconnect efficiency, reduce power consumption, and lower product costs [4] - Marvell launched the Ultra Accelerator Link (UALink) scale-up solution, providing an open-standards-based interconnect platform that enhances compute utilization and reduces latency [5] Competitive Landscape - Advanced Micro Devices (AMD) is advancing its AI solutions through the acquisition of ZT Systems, which will reduce deployment time for hyperscalers [7] - Broadcom (AVGO) reported a 170% year-over-year increase in AI networking revenues, now comprising 40% of its total AI semiconductor revenues, and introduced the Tomahawk 6 switch to enhance AI cluster performance [8] Market Outlook - The Zacks Consensus Estimate for Marvell's fiscal 2026 and fiscal 2027 earnings indicates year-over-year growth of 77.71% and 27.73%, respectively, with recent upward revisions in earnings estimates [16]

市值再创新高的博通正迎来新对手——英伟达。 近日英伟达推出NVLink Fusion，直指博通的高成长型市场：AI定制芯片。这也意味着ASIC芯片市场有 新的加入者。 在全球云服务厂商竞相选择自研芯片用于AI推理、英伟达GPU芯片用于AI训练的组合拳下，ASIC芯片 之争正加速转向生态之争。一方面，UALink和UEC两大连接协议联盟在加速聚合，以对抗英伟达的 NVLink协议；另一方面，英伟达在ASIC方面以半开放态度，已经拉拢Marvell、联发科等合作伙伴。 这背后是ASIC芯片市场的可持续性成长机遇。博通近日发布的财报显示，2025财年第二财季，公司实 现营收150.04亿美元，创历史新高。 博通公司总裁兼首席执行官陈福阳表示："受人工智能网络的强劲需求推动，第二季度公司人工智能业 务营收同比增长46%，超过44亿美元。随着我们的超大规模客户继续加大投资，预计第三季度人工智能 半导体营收将加速增长至51亿美元，实现连续十个季度增长。" 英伟达的入局无疑给聚焦计算芯片市场的玩家阵营带来变数。本质上来说，随着全球对AI计算的需求 从训练转向推理为重心，AI芯片市场也面临一场需求重构。英伟达的动态背后，反映 ...

Optimized custom offering enables end-to-end UALink architecture for rack-scale AI Delivering high compute performance with low power and latency Open standards-based technology supports flexible interconnect solutions that allow customers to innovate at scaleSANTA CLARA, Calif., June 11, 2025 /PRNewswire/ -- Marvell Technology, Inc. (NASDAQ: MRVL), a leader in data infrastructure semiconductor solutions, today announced its custom Ultra Accelerator Link (UALink) scale-up offering. As part of the Marvell® ...

如果您希望可以时常见面，欢迎标星收藏哦~ 两大联盟架构正面交锋 然而，多放1颗别人制作的AI芯片进场，就代表少卖1颗自家的处理器。在记者会上，本刊记者对 英伟达执行长黄仁勋提问："可否谈谈ASIC伺服器未来五年的成长状况？"黄仁勋的回答相当强 势："现在，90%的ASIC专案都将失败⋯，你打造的ASIC芯片必须比我们打造的ASIC强，我们可 是很有竞争力的。"他强调："我们的ASIC会比其他竞争者成长速度更快！" 几天之后，联发科股东会上，当股东问到联发科未来重要的成长动能时，董事长蔡明介看好ASIC 芯片的未来商机。他引用市调公司的数据指出，到2028年时，用于云端的ASIC芯片市场规模将达 到400亿美元以上。 巧合的是，就在英伟达宣布推出NVLink Fusion前一个月，NVLink Fusion的竞争对手UALink联 盟首次推出UALink AI加速器互连1.0版规范，摆明要和英伟达争夺AI数据中心的高速传输标准。 其中值得注意的是，UALink联盟成员中，不少公司都是云端AI芯片的大买家，主要发起人包括阿 里 巴 巴 、 AWS 、 Ａ MD 、 苹 果 、 思 科 、 谷 歌 、 HPE 、 ...

Group 1 - BofA Securities highlighted two key AI scale-up interconnect protocols: Nvidia's NVLink Fusion and AMD's UALink, which are essential for connecting AI accelerators through fast, low-latency networking [1] - UALink is an open-standard initiative aimed at providing an alternative to Nvidia's proprietary NVLink, formed by a consortium of industry leaders including AMD, Broadcom, Intel, and others [2][3] - UALink is seen as a significant step for the non-Nvidia ecosystem, potentially challenging Nvidia's dominance in AI scale-up from 2027 onward [4] Group 2 - Nvidia's NVLink Fusion has opened up approximately $12 billion in networking total addressable market (TAM) that was previously inaccessible, with AMD being a founding member of the UALink Consortium [5] - UALink 1.0 supports up to 800GB/s bandwidth per accelerator, which is less than half of NVLink 5.0's 1.8TB/s [6] - Nvidia is expected to maintain a 1-2 generation advantage in AI scale-up networking as UALink-compatible systems ramp from 2027 [7] Group 3 - NVLink Fusion is projected to expand Nvidia's TAM in networking associated with custom accelerators, with custom silicon representing about 15% of the estimated $500 billion AI compute TAM by 2029-30 [8] - The emergence of UALink as an open standard backed by major industry players is crucial for fostering innovation and competition in the AI interconnect space [10] - The coming years, particularly 2026 and beyond, will be critical for observing UALink's adoption and Nvidia's evolution of its proprietary technology [10]

如果您希望可以时常见面，欢迎标星收藏哦~ 来源：内容来自半导体行业观察综合 。 在本文中，我们来谈一下GPU集群的横向和综合拓展。 让我们从"AI Pod"的概念开始。这个术语对不同的人可能意味着不同的东西，但它通常指的是一种预先配置的模块化基础设施解决方案，旨在简化 和加速AI工作负载的部署。 这些"pod"将计算、存储、网络和软件组件集成为一个紧密相连的单元，从而促进高效的 AI 运行。这就是我们遇到"纵向扩展"和"横向扩展"等术语 的地方。以下是可视化示例： 每个 XPU 刀片通常包含 2 到 8 个 XPU 设备。每个设备可以形成为单片芯片（即由单个半导体切片制成），也可以形成由一组称为"芯片集"的芯 片组成的多芯片系统。 我们这里讨论的计算处理能力令人难以置信，XPU 设备本身也同样如此。例如，NVIDIA 的 B200 GPU 拥有超过 2000 亿个晶体管（当然，我可 没亲自数过）。 CPU（中央处理器） GPU（图形处理单元 NPU（神经处理单元） TPU（张量处理单元） DPU（数据处理单元） FPGA（现场可编程门阵列） ASIC（专用集成电路） END 半导体精品公众号推荐 扩展是有限制 ...

With broad industry support from 100+ companies in the UALink Consortium and the ratification of the UALink 200G 1.0 specification, UALink is emerging as the essential open standard for scale-up AI infrastructureSANTA CLARA, Calif., May 06, 2025 (GLOBE NEWSWIRE) -- Astera Labs, Inc. (Nasdaq: ALAB), a global leader in semiconductor-based connectivity solutions for AI and cloud infrastructure, today announced its upcoming webinar hosted by J.P. Morgan on Ultra Accelerator Link™ (UALink™) technology. Astera La ...