广发证券发布研报称，在大模型应用中，相比依赖外置HBM，SRAM可显著降低权重与激活数据的访 延迟与抖动，从而改善Time-to-First-Token与尾时延表现。目前，Groq与Cerebras都相继推出基于 SRAMAI芯片。SRAM架构进入主流视野，根据Groq官网以及市场媒体报道，英伟达此前斥资200亿美 元获得Groq的知识产权的非独家授权；OpenAI与Cerebras签署100亿美元合同，部署多达750兆瓦的定制 AI芯片。 广发证券主要观点如下： SRAM是片上高带宽存储层 存储分级为SRAM、HBM、DRAM和SSD，其中SRAM(静态随机存取存储器)集成在CPU、GPU计算核 心附近的片上存储，具备纳秒级访问时延与高度确定性的带宽特性，带宽高但容量小、成本高。 SRAM可提升AI推理速度 根据Cerebras官网，其晶圆级引擎3(WSE-3)芯片集成44GB SRAM，片上存储带宽达21PB/s，在OpenAI GPTOSS120B推理任务中实现>3000tokens/s的输出速度，较主流GPU云推理快约15×。此外，2026年2 月，OpenAI推出首个运行在Cerebras Syst ...

Investment Rating - The report provides a "Buy" rating for the industry, indicating an expectation of stock performance exceeding the market by more than 10% over the next 12 months [45]. Core Insights - SRAM (Static Random Access Memory) is identified as a high-bandwidth on-chip storage layer that can significantly enhance AI inference speed by reducing latency and jitter compared to external HBM (High Bandwidth Memory) [3][11]. - The architecture of SRAM is gaining mainstream attention, with significant investments and partnerships, such as Nvidia's $20 billion acquisition of Groq's intellectual property and OpenAI's $10 billion contract with Cerebras [3][32]. - The report emphasizes the growing importance of AI memory-related upstream infrastructure, suggesting that investors should focus on key beneficiaries within the industry chain [3][39]. Summary by Sections SRAM as a High-Bandwidth Storage Layer - SRAM is positioned as an essential component in the multi-tier storage architecture, providing high bandwidth but with limited capacity and higher costs [3][11]. SRAM Enhancing AI Inference Speed - SRAM can improve AI inference speed, with examples such as Groq's LPU chip achieving a bandwidth of 80 TB/s and maintaining stable inference speeds of 275-276 tokens/s, outperforming other platforms [3][15][21]. - Cerebras' WSE-3 chip integrates 44GB of SRAM, achieving over 3000 tokens/s in inference tasks, significantly faster than mainstream GPU cloud inference [3][23][39]. SRAM Architecture Gaining Mainstream Attention - The report notes that major companies are investing in SRAM technology, highlighting Groq's partnership with Nvidia and Cerebras' funding round that values the company at $23 billion [3][32][39]. Investment Recommendations - The report suggests that the ongoing expansion of AI memory capabilities will enhance model performance and accelerate the deployment of AI applications, recommending a focus on core beneficiaries in the industry chain [3][39].

Group 1 - Nvidia has reached a non-exclusive licensing agreement with AI chip startup Groq, reportedly worth $20 billion, acquiring Groq's founder Jonathan Ross and engineering team [1] - Groq focuses on LPU chips for inference, achieving an output speed of 500 tokens per second per card, which is ten times faster than Nvidia's GPUs, utilizing a temporal instruction set architecture to mitigate HBM shortages and reduce costs [1] - This transaction represents a "technology licensing + talent acquisition" model, allowing Groq to continue its cloud business independently while Nvidia aims to enhance its inference computing capabilities targeting the Google TPU market [1] Group 2 - Tsinghua TSAIL Laboratory and Shengshu Technology have jointly open-sourced the TurboDiffusion video generation acceleration framework, reducing the processing time of a 1.3B-480P model on a single RTX 5090 from 184 seconds to 1.9 seconds, achieving a 97-fold acceleration [2] - The framework integrates four core technologies: SageAttention2++ quantization, SLA sparse linear attention, rCM step distillation, and W8A8 quantization, decreasing end-to-end latency from 900 seconds to 8 seconds [2] - SageAttention has been successfully integrated into NVIDIA TensorRT and deployed on platforms such as Huawei Ascend and Moole Technology, with major companies like Tencent, ByteDance, and Alibaba already applying it [2] Group 3 - Shanghai Municipal Planning and Resources Bureau and SenseTime have launched the first 600 billion parameter foundational model in the national planning and resources field, named "Yunyu Xingkong," which can answer questions, adjust maps, perform statistics, recognize images, and generate reports [3] - The model is trained on the Kunyu Jinglue corpus and is integrated with the government intranet's professional version and core business systems, achieving a 98% accuracy rate for specialized terms and a 95% approval rate for human Q&A [3] - It employs a "1+6" (base + vertical) model system and an intelligent scheduling engine, supporting natural language calls for 2D and 3D spatial data, exploring a new paradigm for data productization and service-oriented government models [3] Group 4 - Tencent Cloud and Anhui Yilu Weixing have launched the first AI assistant in the ETC field, named "Assistant Agent," based on Tencent's Mix Yuan model, which has served over one million users since its internal testing began in April [4] - The assistant integrates multimodal interaction technology, supporting both text and voice input, achieving a 95% accuracy rate in Q&A and a 90% problem-solving rate, capable of handling complex requests such as device inquiries, traffic record checks, and invoicing [4] - It deploys 105 state monitoring algorithms to collect real-time device operation data, enabling voice interaction and key status reporting for a "service find person" capability, allowing users to control devices via voice commands [4] Group 5 - Dexmal has proposed the GeoVLA framework, utilizing a dual-stream architecture to retain VLM semantic understanding while endowing robots with 3D geometric perception capabilities through point cloud embedding networks and spatial awareness action experts [6] - In the LIBERO-90 long-range multi-task test, it achieved a 97.7% success rate, surpassing OpenVLA-OFT, and reached an average success rate of 77% in ManiSkill2, with an overall average of 86.3% in real-world tasks [6] - It demonstrated outstanding performance in out-of-distribution scene robustness tests, maintaining a 60% success rate with varying basket heights and a 70% success rate with a 45° viewpoint shift, proving its understanding of true 3D spatial structures [6] Group 6 - The SciMaster team, composed of Shanghai Jiao Tong University's TSAIL Laboratory, Shanghai Algorithm Innovation Research Institute, and DeepSense Technology, has launched ML-Master 2.0, achieving a 56.44% medal rate in the MLE-bench, topping the leaderboard [7] - This system is designed for real machine learning engineering, introducing a hierarchical cognitive caching mechanism that models context as Experience, Knowledge, and Wisdom [7] - It employs a "generate-validate" protocol to achieve ultra-long-range autonomous capabilities, with applications already in theoretical computational physics and embodied intelligence, currently open for Waiting List applications via the SciMaster platform [7] Group 7 - Jim Fan, head of embodied intelligence at Nvidia, stated that Tesla's FSD v14 is the first AI to pass the physical Turing test, with Elon Musk noting that "perception is maturing," and the software has been launched in seven countries including the US [9] - Tesla has established 14 technical barriers, including a sensor freezing scheme for 4-6 years to accumulate data, an instant value judgment engine for intelligent data filtering, and a Neural Codec for processing raw Bayer data [9] - The end-to-end transformer facilitates the transition from photon input to motor torque output, with hardware-in-loop quantization training conducted on the Cortex supercomputer's vehicle chip, updating 12 versions within 77 days, although issues remain with lane switching and lane change decisions [9]