每日AI简报（trustedreviews） · 2026/4/28

英特尔发布 Core Ultra 200V 系列 AI PC 芯片，挑战苹果与高通

来源: https://www.tomshardware.com/pc-components/cpus/intel-core-ultra-200v-lunar-lake-launches-september-10-targets-apple-and-qualcomm-big-bet-on-ai

Intel announced Wednesday that its next-generation AI PC processor, the Core Ultra 200V series codenamed Lunar Lake, will launch on Sept. 10. The chip is designed specifically for AI PCs and will feature an expanded NPU delivering up to 48 TOPS, integrated memory on-package, and a new performance hybrid architecture. Intel's CEO Pat Gelsinger positioned it as a direct competitor to Apple's M-series for the Mac and Qualcomm's Snapdragon X Elite for Windows Copilot+ PCs. The launch comes as Microsoft pushes the Windows Copilot+ PC initiative, which requires a minimum of 40 TOPS from the NPU.

Adobe 为 Premiere Pro 推出 AI 新功能：Project Fast Fill、Firefly AI 等

来源: https://www.9to5mac.com/2024/08/07/adobe-updates-premiere-pro-with-ai-features-project-fast-fill-firefly-ai-and-more/

Adobe has added a suite of new AI features to its professional video editing software, Premiere Pro. The highlight is "Project Fast Fill," which allows editors to quickly remove objects or fill in backgrounds using a new inpainting feature powered by Adobe Firefly Image 3 Model. Other updates include using Firefly AI for text-to-video generation for b-roll creation, integration of third-party AI models like Sora, and an "Enhance Speech" tool that uses AI to clean up audio. Adobe aims to integrate generative AI directly into the video editing workflow to streamline time-consuming tasks for professionals.

AI 竞赛：AMD 在“内存墙”问题上挑战英伟达和台积电

来源: https://futurumgroup.com/blog/posts/amd-challenges-nvidia-and-tsmc-approach-to-the-memory-wall/

With rising demand for AI and HPC chips, the bottleneck around data transfer between processors and memory — the "Memory Wall" — is becoming a critical issue. AMD is now directly challenging the dominant approaches of Nvidia (with its proprietary NVLink interconnect and HBM memory) and TSMC (focused on advanced packaging like CoWoS). AMD has introduced its own 3D V-Cache technology (stacking L3 cache directly on top of compute dies) and the Instinct MI300X accelerator, which pairs CPU chiplets with massive HBM3 memory stacks. By pushing for more efficient memory bandwidth and latency solutions, AMD is positioning itself as a key alternative in the data center AI hardware market.

亚马逊扩大云端 AI 芯片“Trainium2”供应，提供完整机器学习套件

来源: https://siliconangle.com/2024/08/07/aws-expands-cloud-ai-chip-trainium2-availability-launches-new-machine-learning-stack/

Amazon Web Services has made its second-generation AI training chip, Trainium2, generally available to customers. Announced late last year, this chip is designed to efficiently train large language models and is offered on AWS's Elastic Compute Cloud (EC2) instances. Alongside this, AWS has launched a new AWS AI/ML Stack, which integrates Trainium2, the AI inference-focused Inferentia2 chips, and various managed services like SageMaker and Bedrock. The goal is to provide enterprises with an efficient, full-stack platform for the entire machine learning lifecycle, aiming to challenge Nvidia's GPU dominance in the cloud AI training market.

---

谷歌宣布为大型语言模型开发者开放 Vertex AI Agent Builder 代理构建器

来源: https://9to5google.com/2024/08/06/google-vertex-ai-agent-builder-launch/

Google has officially launched the Vertex AI Agent Builder, a new low-code development environment for creating AI agents that can process information and take actions. The builder allows developers to combine Google's Gemini large language models with enterprise search, data connectors, and other tools to build custom conversational agents, chatbots, and automation workflows. It's integrated within the broader Vertex AI platform. This move is part of Google's strategy to provide an enterprise-grade generative AI platform that rivals offerings from OpenAI, Microsoft's Azure AI, and others.

特斯拉在 AI 日首次亮相人形机器人 Optimus Gen 2

来源: https://tech.hindustantimes.com/tech/news/tesla-unveils-its-humanoid-robot-optimus-gen-2-at-ai-day

During its AI Day event in August 2024, Tesla unveiled the second generation of its humanoid robot, Optimus. Now named "Optimus Gen 2," it can walk fully autonomously on flat surfaces at about 5 km/h, an improvement from Gen 1's tethered and assisted walking. The robot features a newly designed actuator that is 30% smaller but provides 20% more torque, along with a 2-DoF (Degrees of Freedom) neck for better head movement. Musk continues to emphasize the robot's potential use in factory and other labor-intensive tasks, seeing it as a key long-term product for Tesla beyond electric vehicles and energy storage.

英伟达将把 AI 芯片制造外包，扩大与台积电的合作

来源: https://www.theregister.com/2024/08/07/nvidia_ai_chip_manufacturing_contracts_expanding/

As the demand for its AI GPUs skyrockets and supply chain constraints persist, Nvidia is reportedly expanding its manufacturing partnerships to ensure sufficient production capacity. According to reports from financial analysts, Nvidia plans to outsource a significant portion of its next-generation chip manufacturing, expanding its collaboration with Taiwan Semiconductor Manufacturing Company (TSMC) and possibly other foundries. The goal is to increase the supply of chips like the upcoming Blackwell B100/B200 and future architectures to meet the overwhelming demand from data centers building out AI infrastructure.

华为发布 Kunpeng DataCenter AI Cluster 数据中心 AI 集群，加速高性能计算与 AI

来源: https://www.prnewswire.com/news-releases/huawei-unveils-kunpeng-datacenter-ai-cluster-to-accelerate-high-performance-computing-and-ai-development-for-industries-globally-302195104.html

Huawei has released its new Kunpeng DataCenter AI Cluster, an integrated hardware-software solution designed for high-performance computing and AI workloads. The cluster is built around Huawei's proprietary Arm-based Kunpeng 920 processors and Ascend AI processors. It features Huawei's proprietary AI framework and software stack, and claims to achieve petascale computing performance optimized for large-scale model training and inference across various industries. Huawei aims to offer this as an alternative to solutions based on x86 and Nvidia GPUs, particularly in markets where it has a strong presence.

DeepSeek 发布开源大语言模型 DeepSeek-V2.5，包含多项创新与改进

来源: https://analyticsindiamag.com/deepseek-releases-its-latest-open-source-llm-deepseek-v2-5/

DeepSeek, a Chinese AI research company, released DeepSeek-V2.5, an upgraded version of its open-source large language model. The model, trained on a mix of text and code data, claims to match the performance of larger proprietary models like GPT-4 while being more efficient due to its advanced MoE (Mixture of Experts) architecture. Key innovations include an improved tokenization strategy and a novel self-attention mechanism designed to reduce computational overhead. DeepSeek positions the model as a strong, cost-effective open-source option for developers and researchers, competing with models from OpenAI and Anthropic.

ChatGPT 将“记忆”功能扩展到所有付费订阅者

来源: 英为财情, https://cn.investing.com/news/stock-market-news/article-2241863

OpenAI is rolling out its "Memory" feature for ChatGPT to all paid subscribers. Previously in beta, this feature allows ChatGPT to remember key details from previous conversations (e.g., "I'm a software developer working on Python projects" or "My child loves penguins") and use that context in future chats. Users have full control, can review and delete specific memories, and can turn the feature off entirely. The goal is to make interactions more personalized and context-aware without requiring users to repeat information. This rollout is part of OpenAI's ongoing effort to make ChatGPT more useful as a persistent personal assistant.