1. I am initiating coverage on Nvidia with a Strong Buy rating due to robust demand for AI accelerators and the upcoming Blackwell Superchips; 2. Nvidia's Q2 FY2025 results shattered records with $30.04 billion in revenue, a 122% YoY growth; 3. Foxconn's massive manufacturing scale-up and Nvidia's $50 billion share repurchase program signal strong confidence in sustained long-term demand for Blackwell chips.
Recent #AI accelerators news in the semiconductor industry
➀ IBM's researchers are inspired by the human brain to develop neuromorphic computing, aiming to mimic the brain's efficiency in processing vast amounts of data required for tasks like AI.
➁ The concept of neuromorphic computing involves designing a computing system that reflects the efficiency of the human brain, which consumes very little power and solves tasks effectively even with ambiguous or undefined data.
➂ IBM's NorthPole chip, an example of neuromorphic architecture, integrates memory and computation on the same chip to overcome the von Neumann bottleneck and reduce energy consumption.
➃ IBM's research teams are working on materials and algorithms to improve the performance and efficiency of neuromorphic chips, with the goal of creating more powerful and energy-efficient AI accelerators.
➀ SiFive推出针对加速高性能AI工作负载的XM系列；➁ XM系列是SiFive的第一个包含可扩展AI矩阵引擎的IP，旨在加速边缘物联网、消费设备、下一代电动/自动驾驶汽车、数据中心等领域的芯片解决方案上市时间；➂ XM系列采用SiFive Kernel Library (SKL)的开源参考实现，提供高效的内存带宽。
➀ Yole Group predicts a CAGR of 9% for the advanced IC substrate market between 2024 and 2029, reaching $25.53 billion in 2029; ➁ The market is seeing a shift from organic core substrates to glass core substrates, driven by demand for FC BGA substrates and 2.5D/3D advanced packages; ➂ Investment in expansion and new plants for high-end substrates is increasing, particularly in the U.S. and China, supported by local government incentives.
➀ Discusses the limitations of using RDMA in ScaleUP networks, highlighting issues with latency, CPU overload, and chip area constraints. ➁ Explores alternative protocols suitable for ScaleUP, emphasizing the benefits of Ethernet and direct memory access. ➂ Analyzes the challenges and solutions in implementing efficient interconnects for AI accelerators, focusing on network convergence and memory semantics.
➀ Tesla introduces TTPoE for Lossy Fabric, combining iWARP TCP congestion control and RoCEv1 layer-2 forwarding. ➁ Azure Maia 100 targets OpenAI's inference business with high-density deployment and standard Ethernet integration. ➂ Meta MTIA focuses on search, advertising, and promotion with a dense model approach and cost-effective LPDDR5 usage.
1. Marvell Technology introduces the new Alaska P PCIe retimer product line, designed for scaling data center compute fabrics. 2. The first products in this line are 8- and 16-lane PCIe Gen 6 retimers, enhancing connections between AI accelerators, GPUs, CPUs, and other components. 3. The 16-lane PCIe 6 retimer operates at an industry-low power consumption of 10 watts, addressing signal degradation and ensuring reliable communication.