AGX Orin
Search documents
对标英伟达 天数智芯公布芯片四代架构路线图
2 1 Shi Ji Jing Ji Bao Dao· 2026-01-29 13:19
21世纪经济报道记者倪雨晴 国内GPU公司正抓紧产业窗口期密集出招。 近日,上市不久的天数智芯发布了芯片的四代架构路线图,并预计于2027年超越英伟达Rubin架构。 同时,天数智芯还披露新一代边端算力产品"彤央"系列,公司称在多项实测场景中性能优于英伟达AGX Orin。 当前,天数智芯正在构建其"AI++算力系统"新范式。而对标英伟达、比肩乃至超越,正成为国产GPU企业在新一轮算力竞赛中 的共同雄心。 四代架构、四款算力产品 天数智芯AI与加速计算技术负责人单天逸系统性介绍了芯片架构推进路径,这份路线图明确了四代GPU架构的时间表及竞争坐 标。 按照规划,2025年推出的"天数天枢"架构,超越了英伟达的Hopper架构;2026年的"天数天璇"架构对标Blackwell,"天数天玑"架 构则超越Blackwell;2027年,"天数天权"架构目标是超越英伟达最新的Rubin架构,并在此之后转向更具突破性的计算架构设 计。 Hopper、Blackwell与Rubin,是英伟达近几年最核心的架构平台,也是其高端GPU产品线的算力基石和主要收入来源。国产GPU 企业若要真正参与全球竞争,绕不开与这些架构的正面 ...
算力通胀终结者!凭一招把大模型Token成本砍到1/2
创业邦· 2026-01-28 12:58
如果在两年前问一家大模型公司最需要什么?答案是 "有没有卡"。但如果今天再问同样的问题,答 算力通胀 之下 都用不起了 "我们正在制造大量的垃圾算力。" 一位负责大模型训练集群的架构师曾这样抱怨。他的焦虑并不是没有风声。过去十年是算力野蛮增长 的十年,规模的快速扩张确实带来了阶段性的产业繁荣。但繁荣背后,是难以忽视的效率困局。 为了追赶 GPT-4 乃至 GPT-5 的能力,国内企业陷入了一场疯狂的参数竞赛。数以万计的 GPU 被 高度集成化塞进数据中心,它们日夜轰鸣,但产出的智能效益却并未如预期般线性增长。 这是一种典型的 "算力通胀"。行业习惯用芯片的理论峰值( Peak Performance )来衡量价值,但 在现实的复杂的训练任务中,这些昂贵的芯片往往"有力使不出"。 数据显示,在许多大规模训练集群中,算力 的有效利用率( MFU )仅能维持在 40% 左右,而在推 理场景下,大量的算力更是处于闲置状态,利用率甚至不足 20% 。 算法迭代与硬件僵化之间的错位也在加剧这种浪费,模型架构每六个月就发生一次巨变,从 案也许会变成"好不好用"。 天数智芯给出的架构路线图: 2025 年的天数天枢架构,超越 ...
研报 | 英伟达发布开源Isaac GR00T N1,预计2028年全球人形机器人市场产值达40亿美元
TrendForce集邦· 2025-03-19 08:41
Core Insights - NVIDIA's GTC 2025 introduced the Isaac GR00T N1 humanoid robot model, which features a comprehensive dataset, multi-modal input, and open-source characteristics, aimed at accelerating research and development in specific applications, potentially leading to a global humanoid robot market value nearing $4 billion by 2028 [1][4]. Group 1: Product Development - The GR00T N1 model allows researchers to conduct post-training using real and synthetic data, thereby shortening the time to market for humanoid robots [1]. - Competitors like Figure AI and Google have launched their own AI models, Helix and Gemini Robotics, respectively, which also focus on enhancing robot decision-making capabilities, but NVIDIA's model is noted for its superior data completeness and deployment flexibility [2][3]. Group 2: Technical Specifications - The GR00T N1 integrates visual language and DiT model optimization, while Gemini Robotics emphasizes reasoning capabilities, and Helix features a dual-system approach for rapid response and understanding tasks [3]. - NVIDIA's hardware solutions, such as Jetson Orin NX and AGX Orin, are designed for various humanoid robot applications, with the AGX Orin offering a peak computing power of 275 TOPS [3]. Group 3: Ecosystem and Market Position - NVIDIA has established a comprehensive ecosystem in the robotics field, including the Omniverse for training simulations and digital twins, which is utilized by both Western and Chinese robot manufacturers [4]. - Despite the open-source nature of GR00T N1, it is expected that developers will still rely on NVIDIA's hardware solutions to maximize efficiency, thereby creating a "moat" around NVIDIA's market position in humanoid robotics [4].