Core Viewpoint - Increasing evidence suggests that AI training does not necessarily rely on high-end GPUs, which may slow down Nvidia's future growth [2][5][14] Group 1: Nvidia's Financial Performance - Nvidia's data center business has experienced strong growth, with revenue increasing by 216% in FY2024 and 142% in FY2025 [2] - Revenue growth rates for Nvidia are projected at 63% for FY2026, driven by a 70% increase in the data center segment, alongside a recovery in gaming and automotive markets [8][9] - The company's total revenue is expected to reach $430 billion in Q1 FY2026, with a slight fluctuation of 2% [6] Group 2: Competitive Landscape - Ant Group's research indicates that their 300B MoE LLM can be trained on lower-performance GPUs, reducing costs by 20%, which poses a significant risk to Nvidia's market position [2][5] - Major hyperscalers like Meta are developing their own AI training chips, reducing reliance on Nvidia's GPUs, with Meta's internal chip testing marking a critical milestone [5][14] - Custom silicon solutions from companies like Google and Amazon are emerging as attractive alternatives for AI training and inference [5] Group 3: Long-term Growth Challenges - Nvidia's high-end GPU growth may face increasing resistance as AI enters the inference phase and lower-cost models become more prevalent [14] - Analysts have revised growth expectations for Nvidia's data center business, projecting a slowdown to 30% growth in FY2027 and further declines to 20% from FY2028 to FY2030 [8][9] - The company's operating expenses are expected to grow by 19% from FY2028 to FY2030, impacting profit margins [9] Group 4: Capital Expenditure Trends - Major tech companies are significantly increasing capital expenditures, with a projected 46% year-over-year growth in 2025, which may boost demand for Nvidia's GPUs in the short term [12][13] - Nvidia has established its own custom ASIC division, potentially mitigating risks from competitors like Broadcom and Marvell [14]

白交 发自 凹非寺 量子位 | 公众号 QbitAI 注意看，机器人像人一样从容地走出大门了！ 甚至，还有一整支机器人队伍迎面走来。 人形机器人独角兽Figure，再次带来他们的新成果—— 利用强化学习实现自然人形行走 。 跟之前版本的机器人相比，确实更像人了许多，而且步态更加轻盈，速度也更快。 根据官方介绍，主要分成三个部分： 强化学习 ：强化学习利用模拟试验和错误，教Figure 02 人形机器人如何像人一样行走。 模拟训练 ：通过高保真物理模拟器学习如何像人类一样行走，结果只需几个小时就能模拟出多年的数据。 Sim-to-Real ：通过将仿真中的域随机化与机器人上的高频扭矩反馈相结合，模拟训练无需额外调整即可直接转换为真实硬件。 网友们纷纷表示被惊艳到，甚至觉得像是 太空行走 。 有一说一，自从与OpenAI取消合作后，这成果输出确实又快又多。 机器人像人一样自然行走 此次推出的，是经过强化学习训练的端到端神经网络。 具体来看。 首先，利用强化学习技术，在GPU加速物理仿真中对新的行走控制器进行了全面训练，并在几个小时内收集了数年的仿真演示数据。 在模拟器中，数以千计的Figure 02机器人被并行模 ...

如果您希望可以时常见面，欢迎标星收藏哦~ Nvidia 在很大程度上拥有 AI 训练，并且如今在 AI 推理方面占有很大的份额，尤其是基础和推理 模型。所以你可能会认为路线图上没有具体信息。但 Nvidia 也让世界上很多人想知道对 AI 计算的 需求是否最终会减弱，或者至少会用更便宜的替代品来满足。此外，作为其最大客户的所有超大规 模和云构建者也在构建自己的 CPU 和 AI 加速器；公开的路线图是为了提醒他们 Nvidia 致力于构 建比他们更好的系统——并让我们都知道，这样我们就可以跟踪谁在实现他们的里程碑，谁没有。 Nvidia 的路线图非常宏大，它拥有 GPU、CPU、纵向扩展网络（用于跨 GPU 和有时 CPU 共享内 存的内存原子互连）和横向扩展网络（用于更松散地将共享内存系统相互连接）。它还有 DPU，即 具有本地化 CPU 和有时 GPU 处理的高级 NIC，以下路线图中未显示这些产品： Quantum 系列 InfiniBand 交换机的容量增长也同样不尽如人意，也没有入选。对于人工智能领域来 说，InfiniBand 的重要性越来越低，因为人工智能领域希望能够进一步扩展，而基于 Infi ...

图片来源： Unsplash Nvidia 在 AI 芯片领域的霸主地位正面临挑战，初创公司 DeepSeek 等竞争对手，正抓住 AI 计算需求 变化，试图打破其统治。 从训练到推理，AI芯片市场格局的转变 DeepSeek 的 R1 和其他推理模型，如 OpenAI 的 o3 和 Anthropic 的 Claude 3.7 ，在用户发出请求时 消耗的计算资源比之前的 AI 系统更多。 这改变了 AI 计算需求的重点，直到最近，这一需求还主要集中在模型的训练或创建上。随着个人和 企业，对超越目前聊天机器人（如 ChatGPT 或 xAI 的 Grok ）应用的需求增长， 推理预计将在技术 需求中占据更大比重。 Nvidia 的竞争对手，从 Cerebras 和 Groq 等 AI 芯片制造商初创公司，到谷歌、亚马逊、微软和 Meta 等大型科技公司，定制加速处理器——正集中力量，试图颠覆这家全球最有价值的半导体公司。 "训练让 AI 成长，而推理则是 AI 的应用。" Cerebras 的CEO Andrew Feldman 表示，" AI 的使用量 已经大幅飙升，目前打造一款在推理方面远胜于训练的芯片， ...

由于该处理器专为 AI 训练而设计（这意味着要处理大量数据），因此预计该处理器将配备 HBM3 或 HBM3E 内存。考虑到我们正在处理定制处理器，Meta 定义了其支持的数据格式和指令，以优化 芯片尺寸、功耗和性能。至于性能，该加速器必须提供与 Nvidia 最新的 AI GPU（例如 H200、 B200 以及可能的下一代B300）相媲美的每瓦性能特性。 该芯片是 Meta 的 Meta 训练和推理加速器 (MTIA) 项目的最新成员。该项目曾面临各种挫折，包括 在类似阶段停止开发。 例如，在有限的部署测试中，Meta 内部推理处理器未能达到性能和功率目标，因此停产。这一失败 导致 Meta 在 2022 年改变了战略，大量订购 Nvidia GPU，以满足其即时的 AI 处理需求。 如果您希望可以时常见面，欢迎标星收藏哦~ 来源：内容 编译自tomshardware ，谢谢。 几年前，Meta 是首批为 AI 推理打造基于 RISC-V 的芯片的公司之一，旨在降低成本并减少对 Nvidia 的依赖。据路透社报道，该公司更进一步设计了（可能是在博通的帮助下）用于 AI 训练的 内部加速器。如果该芯片能够满 ...

大模型训练成本，再砍一刀！ MoE是当前大模型的主流架构，最近大火的国产大模型DeepSeek采用的就是MoE架构。DeepSeek自研的 DeepSeekMoE作为一种创新的大规模语言模型架构，通过整合专家混合系统、改进的注意力机制和优化 的归一化策略，在模型效率与计算能力之间实现了新的平衡。 字节豆包大模型团队表示，MoE在分布式训练中存在大量跨设备通信开销，严重制约了大模型训练效率和 成本。针对这一难题，字节在内部研发了COMET计算-通信重叠技术，通过多项创新，大幅压缩了MoE专 家通信空转时间。 在此前的"开源周"活动中，DeepSeek也曾开源了团队为解决MoE通信瓶颈而采取的DualPipe+DeepEP方 案。不过，与之不同的是，COMET可以像插件一样直接接入已有的MoE训练框架，支持业界绝大部分主 流大模型，无需对训练框架进行侵入式改动，更加方便、灵活、通用。这一方法，还因其简洁性与通用性 而高分入选全球机器学习系统顶级会议 MLSys 2025，被认为"在大规模生产环境中极具应用潜力"。 不仅如此，由于在降低MoE通信开销上，COMET采用了计算-通信融合算子的优化方式，DeepSeek ...

Core Viewpoint - Dell's stock has underperformed since November due to market concerns about a slowdown in AI data center construction, but the company is positioned to benefit from the shift towards inference computing, suggesting potential upside for its stock price [1][10]. Group 1: Market Concerns and Opportunities - The market is worried about the efficiency of AI chips leading to a slowdown in GPU demand, which could impact sales growth expectations for companies like Dell [1]. - Despite concerns, key factors are shifting favorably for Dell, particularly in the inference computing space, which is expected to perform well [1][10]. - The transition from pre-training to inference computing is anticipated to happen faster than expected, with more cost-effective data centers supporting AI inference [3][10]. Group 2: Strategic Partnerships - Dell has partnered with AMD to integrate Ryzen AI PRO processors into new Dell Pro devices, marking a significant milestone in their strategic collaboration [4]. - AMD's CEO highlighted that the total cost of ownership (TCO) for AMD's inference computing solutions is significantly lower than Nvidia's, which could benefit Dell in both PC and server markets [4][9]. Group 3: Financial Performance Expectations - Dell is expected to report solid earnings and revenue growth in its upcoming Q4 financial results, with analysts predicting a 14.46% year-over-year increase in earnings per share (EPS) to $2.52 [5]. - Revenue forecasts for Q4 are set at $24.57 billion, indicating a 10.09% year-over-year growth, with a consensus among analysts on the earnings estimates [5][6]. Group 4: Valuation Metrics - Dell's non-GAAP expected price-to-earnings (P/E) ratio is 14.50, significantly lower than the industry median of 23.87, indicating a 39.26% discount [9]. - The expected price-to-sales (P/S) ratio for Dell is 0.83, which is 73.43% lower than the industry median of 3.11, suggesting strong valuation metrics [9]. Group 5: Future Growth Catalysts - Dell is projected to benefit from a $5 billion deal with Elon Musk's xAI and an anticipated $4 billion increase in AI server shipments from FY 2024 to FY 2025 [8][9]. - The shift towards inference computing is expected to catalyze Dell's next growth phase, supported by recent strategic agreements [11].

Core Insights - The article challenges the traditional belief that language models require "clean data" for effective training, suggesting that exposure to noise and imperfect data can still lead to strong language capabilities [1][2]. Group 1: Research Findings - Researchers from Peking University conducted experiments by intentionally adding random noise to training data, revealing that models could tolerate up to 20% "garbage data" with minimal impact on performance, as the Next-token Prediction (NTP) loss increased by less than 1% [2][4]. - The study utilized the OpenWebText dataset and injected random noise ranging from 1% to 20%, demonstrating that even with significant noise, the model's predictive loss remained relatively stable [3][4]. - The findings indicate a complex relationship between noise and model performance, leading to the development of a new method called "Local Gradient Matching" (LGM) to enhance model robustness in noisy environments [2][10]. Group 2: Theoretical Analysis - The research posits that the presence of noise does not significantly alter the global minimum of the NTP loss, even when noise levels are high, due to the low probability of finding meaningful text within random noise [6][7]. - The study's assumptions can be extended to multilingual models, where different languages can be viewed as noise to each other, thus not adversely affecting the performance of individual languages [9]. Group 3: Practical Implications - Despite the minor changes in pre-training loss, downstream tasks showed a decline in accuracy, highlighting a "loss-performance decoupling" phenomenon where pre-training metrics do not fully capture model capabilities [10]. - The proposed LGM method enhances the model's resistance to noise by constraining the gradient differences between original and perturbed features, thereby maintaining decision consistency under noise [10][12]. - Experimental results across various natural language understanding and visual classification datasets confirmed that LGM significantly improves performance for models affected by noise [11][12]. Group 4: Future Directions - The research opens new perspectives on large-scale pre-training, suggesting that retaining some random noise can reduce data cleaning costs, particularly for resource-constrained teams [15]. - Future work will explore the dynamic relationship between noise types and model capacity, as well as the application of LGM in other modalities [14].

以下文章来源于36氪汽车 ，作者李安琪 36氪汽车 . 看懂汽车产业新百年。36氪旗下智能电动车产业报道公号。 特斯拉FSD入华，时不我待。 文 ｜ 李安琪 编辑 ｜ 李勤 来源｜ 36氪汽车（ID：EV36kr） 封面来源 ｜ IC photo 万众瞩目的特斯拉FSD中国落地终于有了实质进展。 2月25日，特斯拉中国官方发布了FSD智能驾驶软件更新的消息。 从功能描述上看，特斯拉将向用户推送的"城市道路Autopilot自动驾驶辅助"功能，车辆能在城市道路直行、左右转、掉头、自动变道等。 （图源：特斯拉小程序） 不过特斯拉客服回应称，该功能仅为L2级辅助驾驶，尚无法实现美国FSD的完全无人驾驶功能。 这或许是特斯拉官网将"FSD 完全自动驾驶"的表述改为"FSD智能辅助驾驶功能"的原因。特斯拉强调，未来用户车辆将能够在驾驶员极少干预的情况下完成 绝大多数的驾驶任务。 （图源：特斯拉官方） 但这基本与特斯拉在北美推出的FSD主体功能一致，也与当下国内华为、小鹏、理想等车企的城区领航功能类似。也就是说，特斯拉FSD功能引进中国市场 的消息盛传一年多后，终于迈出实质性一步——开始上车交付。 FSD（全称Full ...