电商广告的账，最后都记在毫秒上。
京东给全链路的时延预算是300毫秒，留给广告模型推理的部分只有20到100毫秒。超出这个窗口，用户滑动页面时会察觉到卡顿，电商广告成交也就无从谈起。而在最近两年，这套推荐系统正从传统深度学习迁往生成式范式：行为上下文长了10倍，模型结构大了100倍，算力需求随之涨了大约1000倍。机房扩容的周期，却是按季度计的。
当然，这不仅是京东一家所面临的挑战，而是困扰每一家涉及搜广推的互联网企业的共性难题。而在解决这个难点的路上，京东已经先行启程，为行业提供了可借鉴的解题思路。
算力的账：两位数增长与千倍挑战
先看生意本身。京东覆盖超7亿年度活跃用户，平台上70%的商家同时也是广告主，每月帮助顶级品牌触达超4000万核心受众，作为京东集团旗下的营销推广平台，面向B端用户的京准通与C端流量双向协同驱动增长。
京东零售广告产研部算法应用相关负责人对搜广推业务的定位很直接：“搜广推这个业务，是京东在同时撮合广告主，甚至是商家、卖家，跟我们京东消费者进行最重要的匹配，这件事情的用户体验的好和差，很大程度上直接决定京东业务发展的体量。”换句话说，搜广推不只是推荐系统的一个子集，而是京东商业效率的核心撮合层。它连接两端：一端是广告主和商家，另一端是消费者；匹配得好不好，直接体现在成交、转化和平台收入上。
生意的底层是算法，算法的底层是算力。电商广告成交=曝光×点击率×转化率，过去十几年，整个互联网广告行业都在沿着这条路上迭代：从规则投放，到机器学习，再到深度学习。而京东广告系统的演进轴上，最近的两级是生成式和Agentic AI。
需要说明的是，传统推荐系统一般被称作“判别式推荐”，并不全是深度学习，其核心是基于统计关联和手工特征工程的传统机器学习方法；深度学习是2015年之后才逐渐融入推荐系统的技术路线。而近两年生成式技术落到广告场景，带来的是两位数级别的收入提升，但这一切的前提是：具备先进的算法和充裕的算力。
基于此，京东广告在过去一两年从深度学习迁移到生成式推荐，用统一的序列模型把这些能力融合在一起，最终实现了行为上下文增长10倍、生成式模型结构放大100倍、算力需求提高约1000倍。而这个增长速度，远远超过了业务复杂度和业务收益的增长速度。
这些增长带来的挑战，不是单纯能靠GPU解决的。对此，京东零售广告产研部算法应用相关负责人表示，京东发现，在传统智算超节点聚焦纯GPU/NPU的背景下，想要让算法跟业务紧密结合，真正的制约是数据预处理、数据检索与业务逻辑嵌入，这些流程大量跑在通算上，CPU资源远远不够。“真正的业务流程是网状的，你依赖的算力从来不是NPU和GPU，而是能不能把前面最耗时最复杂的部分先搞定，才能给加速芯片任务。”该名负责人指出。
这个过程中，时延是一条不可谈判的红线。京东广告业务的日请求量千亿级，单次推理必须控制在几十毫秒内完成，否则收益归零。京东零售广告产研部算法应用相关负责人补充了一个来自用户侧的刻度：人对APP滑动和点击的反应有判断，超过600多毫秒就会明显感到卡顿；扣掉基站通信之后，真正留在数据中心、留给模型推理的空间很小。另一头，算法每周更新，新论文从开源到上线最短只要两周，而算力扩容以季度计。
时延主要有两部分，一部分是KV Cache。行为序列超1K，KV Cache容量需求被急剧放大；传统“片上内存—本地DDR—本地SSD—远端Redis”四级缓存存在容量小、并发弱、时延高、额外投资的矛盾——又快又大又稳定不可能同时满足。
另一部分是RPC通信。每次用户请求最少要建立几十个RPC服务调用，RPC通信耗时占整体时延20%以上，百万级RPC导致每节点TB级内存消耗。
这两笔账叠在一起，构成了整个搜广推行业在生成式时代最真实的算力账单：收入是两位数的增长空间，成本是千倍的算力需求，而时间窗口以毫秒计。
如何能把时延降低15%？
面对挑战，超节点成为当下最优解之一。先把时间拨回去年5月的鲲鹏开发者峰会。当时京东与鲲鹏团队确认了多个合作重点项目，判断的是三件事：这个事值不值得做、双方分别能做什么、做成的概率和收益空间有多大。“可能都不是以超节点的概念出来的，更多是我们遇到一些问题，一步步逼着大家要解决。”京东零售广告产研部算法应用相关负责人回忆。
既然“又快、又大、又稳定”在传统架构里是不可能三角，京东广告选的方向是做分级存储并尽量压缩层级。四级存储逐级命中调度本身就是负担，于是提出：把第四级（远端数据库）干掉，把第三级也干掉，只留两级（第一级片上内存，第二级一块足够大的DDR）。去年5月，华为架构师给出的方案是通过灵衢协议把几十台、几百台甚至更多机器连起来形成一个大池子，访问这个池子的内存与访问本地没有区别。
同时还有第二个动因：模型跑得更快、预估更准，依赖大量数据预处理。传统做法是在大数据集群上离线处理再读进来，但很多用户行为刷新极快，离线回流太慢。京东广告希望直接用推理集群的算力做处理，拿最新鲜的行为数据，数万起步的预处理任务如果按传统分布式方式拆到很多节点，必然出现木桶效应，所以更合适的方式是让CPU与数据近端处理，靠拓扑感知调度把任务打到最可能需要这份数据的节点上。
具体来看，华为鲲鹏与京东广告的合作主要集中在三个层面：
第一是创新组网。基于灵衢（UnifiedBus）互联实现4K节点ScaleUp组网，单超节点支持150万+vCPU，每节点3.2Tb互联带宽，具备大规模、大带宽、不收敛、高可用四大特性。“不收敛”这四个字对广告业务至关重要——跨层带宽不收敛、扩容时带宽不下降，才扛得住广告模型和数据的持续膨胀。
第二是KV Cache两级缓存。通过灵衢协议把几十台、几百台机器连成巨大的内存池，干掉传统的第四级远端数据库和第三级，只留片上内存+DDR两级，形成300TB共享内存池。结果是一组相当陡峭的数字：缓存间时延下降3个量级，带宽达TCP的16倍；8MB并发请求TP99时延下降85%以上（最高92.6%），64MB序列长度TP99降至1/16。“以往来说，容量跟时延是二选一的题，现在有了鲲鹏超节点合作，可以都要。”京东零售广告产研部相关负责人在华为全联接大会2026——鲲鹏计算产业峰会上说。对京东广告而言，这意味着可以继续扩大用户的行为序列作为推理的输入，提高推荐精度。
第三是RPC通信改造。跳过传统网络协议栈（Socket/TCP/IP），UB统一协议直连URMA，时延开销降低38%；RPC通道复用使通信用内存降低100倍；多路径聚合Tb级带宽替代百Gb单一通路，通信稳定性提升40%，1MB数据包时延下降84%，P99整体优化40%以上。
正是通过这三个环节的改变，京东广告把端到端推理时延降低了15%。但落地的过程也并不是一帆风顺的。“灵衢是一个非常好的能力，或者说一个工具、一个构想，但是在一个构想落地的过程中，一定会面临现实的拷打。”京东零售广告产研部算法应用相关负责人坦言。
首先，机房空间就是最大的挑战。要连的机器足够多，但机房空间不够，华为给出的反应很快，既然机器摆不开，就做液冷服务器提高摆放密度，同时线缆又不能太长；其次，扩展性与稳定性也是不小的挑战。灵衢的扩展能力、以及灵衢协议外联时的泛化问题，无一例外都需要稳定性；第三，系统风险被放大。如果业务强依赖一个巨大无比的内存池，一旦核心链路出现断网或抖动，影响范围比传统架构大得多。
这些问题的解法分散在多个层面。华为今年发布了多款NPO硬件解决光传输链路上的问题，操作系统的调度在同步升级，灵衢协议也在持续打磨，新增能力分别落在协议层、交换链路、操作系统以及CPU/NPU侧芯片上。“每一个新打磨进去的能力，我们一定都是踩过对应的坑，而这些坑的价值在于后面的人可以少踩。”京东零售广告产研部算法应用相关负责人如是说。
当新的架构搭建完成后，验证选在的电商业务负荷的最高峰之一的“618”大促期间。京东零售广告产研部算法应用相关负责人告诉我们，这套软硬件方案要至少能达到三个效果：系统不能崩、要在相对稳定的情况下服务更多客户、延迟不能因为压力大就飙升导致降级或熔断。
最终，这套方案承受住了“618”的“大考”，实现广告推理端到端时延降低15%，局部KV Cache命中率相关的重复计算降低百分之八九十，核心业务场景5%的流量已在鲲鹏集群上运行，经受住了洪峰、稳定性与时延的“考题”。
更重要的是这笔账的商业属性。“延迟省下来之后我们就会在这个固定的延迟内上更大的模型。更大的模型，更好的效果，更多的钱。”京东零售广告产研部算法应用相关负责人表示，在京东广告内部有一个观点：时延就是收益，这不只是降本，更是增收。省下的时延预算可以用来跑更大的模型，更大模型带来更好效果，进而带来更多广告收入。
当然，这一切还只是合作的开始，京东零售广告产研部算法应用相关负责人与我们分享了接下来的计划，他表示，目前还不是全量的搜广推业务都在鲲鹏上，一方面产能还需要提升，另一方面进度也没有那么快，所以策略是拿核心业务优先上新型集群，先在核心业务场景发挥作用。据披露，京东广告目前正在稳步推进“双十一”，将更大规模的业务搬上鲲鹏。
从SuperCache到Agent沙箱，下一程在哪？
随着AI Agent的发展，算力的需求还会不断增加，而双方的合作也将继续围绕让算力更普惠、更好用展开。
首先是“打好地基”。KV Cache和数据处理这两块在鲲鹏上已经做了大量工作，目前正处在大规模上量过程中，核心效果业务先上新型集群，会坚定继续推进。京东广告也意识到，除了已经实现的部分，过程中还发现了更多需要改造和改进的地方，这部分会与华为在系统、调度、机房、优化等各团队继续合作。
其次，在原有方案的基础上，除了广告推广业务，还将进一步将能力外溢。京东零售广告产研部算法应用相关负责人表示，同样的解法不只对广告场景有用，对搜推这类不需要考虑广告主的业务同样适用。
除此之外，京东之外的其他友商也在跟进做相关业务，这套方案对他们也是一个范例。该负责人还预判，随着模型越做越大、对算力和延迟的苛求越来越激进，改造一定会从广告相关推理向更上游和更下游延展。
这与华为侧的开放姿态是呼应的：灵衢希望吸引两类伙伴，一类是硬件厂商，基于其IP开发产品，可以是服务器厂商，也可以是做SSD协议的存储厂商；另一类是软件层客户，基于灵衢协议做架构和软件方案层面的创新。京东与鲲鹏共同踩过的坑、打磨进协议层与芯片侧的每一项能力，最终都将成为后来者少走弯路的路标。
第三是围绕其他前沿技术进行探索，比如共享SuperCache（L2、L3扩容，通算与智算统一互联）、Agent高密沙箱（秒级并发十万级沙箱，资源5倍超分，跨节点自由调度）、以及多模态融合检索等方向。
最后，Agentic将是接下来的重点。智能体适合的是决策复杂、次数有限且需要做最优化决策的场景；规则极其明确的场景反而没有Agent化的必要。京东零售广告产研部算法应用相关负责人表示，广告和搜推天然存在这类场景，京东广告希望把决策空间很大、次数有限且需要做最优化决策的场景都Agent化。但这位负责人指出，这件事的前提有四个：数据是否足够多、算力是否充沛、基础模型能力是否足够强、能不能支撑模型的循环迭代；再往下一层，落到算力和通信上，挑战比现在会更大。
在这个背景下，双方对算力底座还有一层更深远的判断。华为在鲲鹏计算产业峰会上给出过一个趋势：进入Agentic AI时代，智能体将大量依靠CPU完成调度编排、记忆处理、沙箱运行等工作，CPU与NPU的配比将从过去的1:4历史性地走向1:1。而京东从业务侧看到的更激进——1:1根本不够，从业务的复杂度出发，看到的可能是4:1。原因有两个：前面要处理的数据越来越多，同时单卡性能本身也在上涨。“超节点可能是一个硬件的概念，但事实上远远不是一个硬件的概念。”这意味着操作系统、调度、中间件都会跟着目标改变硬件形态。
让我们再回到文章开头提到的那笔电商的账，“如果可以优化下来一毫秒，我们就要有几千万的提升。”在广告这个行业，对于延迟的优化是无止境的。端到端推理时延降低15%只是开始，Agentic AI的星辰大海，才是双方未来合作聚焦的重点。（本文首发于钛媒体APP）
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
The bills of the electrician ads are finally recorded on milliseconds.
The latency budget for the entire link is 300 ms, leaving only 20 to 100 ms for advertising model reasoning. Beyond this window, users can see Cardon when they slide the page, and there's no way to talk about an electrical commercial. In the last two years, the referral system is moving from traditional in-depth learning to the production paradigm: behavior context is 10 times longer and model structure is 100 times larger, with a corresponding increase in computing demand of about 1,000 times. The machine room is expanded on a quarterly basis.
Of course, it is not only a challenge for the Kyoto family, but rather a common problem for every Internet enterprise involved in the search for widespread Internet access. On the way to resolving this dilemma, Kyoto has already set out to provide industry with a useful solution.
Powerful: double-digit growth versus a thousand times the challenge
Look at the business itself. Kyoto covers over 700 million year-old active users, 70 per cent of the vendors on the platform are also advertisers, helping top brands reach more than 40 million core audiences per month, serving as a marketing outreach platform under the Kyoto Group ' s umbrella, and twin-track growth driven by the Yinkyo and C traffic for B end users.
The director of the application of algorithms in the Ministry of retail advertising and research in Kyoto has been very direct in locating the search and promotion business: “The search and promotion business is the most important match between Kyoto's combination of advertisers, even merchants and sellers, and our Kyoto consumers. In other words, excavation is not just a subset of the recommended system, but a core conglomerate of commercial efficiency in Kyoto. It connects to two ends: one to advertisers and merchants and the other to consumers; and the match is directly reflected in transaction, transformation and platform income.
The bottom of the business is algorithm. The bottom of the algorithm is algorithm. Electronics commercialized = exposure x click x conversion rate. Over the past decade, the entire Internet advertising industry has been following this path over time: from rules to machines to in-depth learning. On the developmental axis of the Kyoto advertising system, the most recent two levels are Generating and Agenic AI.
It should be noted that the traditional referral system is generally referred to as the “judgmental referral” and is not all about in-depth learning, the core of which is the traditional machine learning method based on statistical linkages and manual characterization engineering; in-depth learning is the technical route of gradual integration into the referral system after 2015. In the last two years, however, the generation technology has reached the advertising scene, resulting in a double-digit increase in income, all of which presupposes advanced algorithms and sufficient computing power.
As a result, over the past year or two, the Jyeongdong ad has moved from deep learning to generating recommendations, combining these capabilities with a unified sequence model, resulting in a tenfold increase in behavior context, a 100fold increase in the generation model structure, and a 1,000fold increase in computing needs. This growth rate, in turn, far exceeds the level of operational complexity and gains.
The challenges posed by these growths cannot be solved by GPU alone. In response, the director of the application of algorithms in the Ministry of retail advertising and research in Kyoto stated that, in the context of the traditional super-centre focus on GPU/NPU, the real constraint in linking algorithms to business was pre-processing, data retrieval and business logic, and that CPU resources were far from sufficient. “The real business process is web-based, and the calculations you rely on are never NPU and GPU, but whether the most time-consuming and complex part of the front can be addressed before you can speed up the chip mission.” The head pointed out.
In the process, latency is an unnegotiable red line. The daily requests for the Kyodong advertising business are hundreds of millions of degrees, and the single reasoning must be contained in dozens of milliseconds, otherwise the proceeds will be zero. A user-side scale was added by the director of the application of the algorithm of the Ministry of retail advertising and research in Kyoto: the response of a person to the APP slides and clicks was judged to be more than 600 milliseconds of the day, which would clearly be felt in Cardon; and the real space left in the data centre after the loss of base station communications was removed, leaving little room for model reasoning. On the other hand, algorithms are updated on a weekly basis, with new papers starting from open source to online for a minimum of two weeks, with a quarterly build-up.
latency consists mainly of two parts, part KV Cache. Behavior sequences over 1K, KV Cache capacity needs dramatically amplified; traditional "film memory" - Local DDR - Local SSD-Land End cache has small and weak capacity, high latency, additional investment conflicts - fast, large and stable cannot be satisfied simultaneously.
The other part is RPC communication. A minimum of dozens of RPC service calls are established each time a user requests, with RPC communications taking more than 20 per cent of the total latency and a million-degree RPC contributing to the consumption of TB level memory per node.
The combination of these two accounts constitutes the most realistic estimate in the generation age of the entire search and promotion industry: Income is a double-digit growth space with a cost of thousands of times the calculus demand, while the time window is in milliseconds.
How can latency be reduced by 15%?
Faced with the challenge, the hypernodes became one of the best solutions at the moment. The time has first been set aside for the summit of developmentists in May. At that time, the team identified a number of priority projects for cooperation, judging by three things: the value of the matter was not worth doing, what the parties could do separately, the probability of doing it, and the scope for profit. “may not have come out of the concept of hypernodes, but more so that we are confronted with problems that are pushing us to solve.” The director of the Ministry of Commercial Advertising and Commercial Research in Kyoto remembers the application of algorithms.
Since “quick, large and stable” cannot be triangulated in the traditional architecture, the orientation of the Kyoto advertisement is to store it in a hierarchical manner and to compress it as much as possible. The fourth level of storage-by-step movement was a burden in itself, and it was suggested that phase 4 (distant database) should be eliminated and level 3 should be eliminated, leaving only two levels (first-stage memory, second-stage sufficiently large DDR). In May of last year, the Chinese proposal for architects was to connect dozens, hundreds or even more of machines into a large pool, with no distinction between access to the memory and local access.
There is also a second driving factor: faster, more predictable models relying on extensive data pre-processing. The traditional approach is to re-read offline processing in large data clusters, but many users are very quick to refresh their behaviour and offline returns too slowly. The Kyoto Advertisement wants to be processed directly with the reasoning cluster, with the most fresh behavioral data, and tens of thousands of start-up pre-processing tasks, if broken down to many nodes in the traditional distribution pattern, will inevitably have a barrel effect, so it is more appropriate to have CPU and the data processed at a close end, and to hit the task at the node where the data are most needed by increasing sensory control.
In particular, the collaboration between China and Kyoto advertising centres on three levels:
The first is the Innovation Cluster Network. The 4K node ScaleUp network, based on the UnifiedBus, supports 1.5 million +vCPU with 3.2Tb bandwidth per node, with four main features: large, large, non-condensed and highly usable. The word “non-repression” is essential for advertising operations — the continued expansion of advertising models and data is carried out with a wide range of layers of bandwidth that is not constricted or reduced at the time of expansion.
The second is KV Cache two levels cache. Through spiritual agreements, dozens of machines, hundreds of them, are connected to large memory pools, the traditional level 4 remote database and level 3 are eliminated, and only the +DR two are left, creating a 300TB shared memory pool. The result was a rather steep set of figures: latency, between cache, fell by 3 orders of magnitude, 16 times the width of TCP; 8MB, together with a request for TP99Z102Z, fell by more than 85 per cent (up to 92.6 per cent) and 64MB, by TP99, to 1/16. “In the past, capacity and latency were two-choice issues, and now there are supernodes that can be taken.” The director of the Ministry of retail advertising and research in Kyoto spoke at the China General Assembly 2026-Clause Industry Summit. For the Kyoto advertisement, this means that the user ' s behavioural sequence can continue to be expanded as an input for reasoning and to improve the accuracy of the recommendation.
The third is the RPC communications retrofit. Skipping the traditional Internet protocol bar (Socket/TCP/IP), UB UTA direct UMA UMA, latency cost reduction of 38 per cent; re-use of the RPC channel reduces the memory of communications by 100 times; multi-routing Tb-bandwidth replaces the 100 Gb single route, with a 40 per cent increase in communication stability, a 84 per cent decrease in 1MB package Z102Z and an overall optimization of P99 by more than 40 per cent.
It was through these changes that the Kyoto ad reduced the end-to-end reasoning latency by 15 per cent. But the landing process was not easy." Spirituality is a very good ability, a tool, a vision, but in the process of a vision's landing, there will be real torture.” The head of the Ministry for the Study of the Use of Algorithms in Retailing Advertising in Kyoto has spoken openly.
First, room space is the greatest challenge. There are enough machines to run, but there is not enough room for room, and the response is quick. Since the machine can't swing, the liquid-cool servers are set up to increase density, while cables cannot be too long; and, secondly, expansion and stability are not too small a challenge. Vibrant capacity for expansion, as well as broader issues of outreach, require stability without exception; and third, systemic risks are magnified. If operations rely heavily on a large and incomparable memory pool, the impact is much greater than the traditional architecture once the core links are broken or shaken.
The solution to these problems is scattered on many levels. For this year's publish, China's multiple NPO hardware solves the problem of light transmission chain, the movement of operating systems is being upgraded simultaneously, and psychic protocols continue to grind, with additional capabilities falling on the protocol layer, the exchange chain, the operating system and the CPU/NPU side chip. “Every new ability to grind in, we must have stepped on the corresponding pits, the value of which lies in the fact that those behind can do less.” If so, the person in charge of the application of algorithms in the Ministry of retail advertising research in Kyoto.
When the new architecture is completed, the verification takes place during the “618” peak of one of the peaks of the electrician's operational load. The director of algorithm application at the Ministry of retail advertising and research in Kyoto told us that the software and hardware package would have at least three effects: the system would not collapse, more clients would need to be served under relatively stable conditions, and delays would not be reduced or melted as pressure surged.
Ultimately, the package was able to withstand the “618” “matric examination”, which resulted in a 15 per cent reduction in advertising reasoning to the end of latency, an 80 per cent reduction in local KV Cache-related double counting, and a 5 per cent flow of core business scenes that were already operating in clusters and weathered the “tests” of flood peaks, stability and latency.
More important is the commercial attributes of the account. "When the delay is saved, we'll have a bigger model within this fixed delay. Larger models, better effects, more money." According to the director of the application of the algorithm of the Ministry of retail advertising and research in Kyoto, there is a view within the Kyoto advertising that latency is a gain, which is not just a downfall, but a gain. The savings from the latency budget can be used to run a larger model, with a better impact from LLM, which in turn generates more advertising revenue.
Of course, this is only the beginning of a cooperative effort, and the director of the Ministry of Commercial and Commercial Studies, who shared his plans for the next step, said that not all of the search-and-dissemination operations were under way, that capacity needed to be improved on the one hand, and progress on the other, on the other, that the strategy was to prioritize core operations over new clusters, and to play a role in the core business landscape on the other. It was revealed that the Kyoto Advertisement is now steadily moving “two elevens” to move even larger operations to the top.
From SuperCache to Agent Sandbox, where's the next ride?
As AI Agent develops, the need for computing will continue to grow, and cooperation will continue around making it more inclusive and useful.
First of all, it is “building the foundations”. The KV Cache and data processing components have already done a great deal of work and are now in the process of being scaled up on a large scale, with core impact operations ahead of the new cluster and moving forward firmly. The Kyodong Advertisement also recognizes that, in addition to what has already been achieved, more needs to be done in the course of the process, which will continue to work with the Chinese teams in systems, movement control, aircraft rooms and optimization.
Second, building on the previous programme, there will be a further spillover of capacity beyond advertising. According to the director of the application of the algorithm of the Ministry of retail advertising research and development in Kyoto, the same solution is not only useful for advertising scenes, but applies equally to searches that do not need to consider the business of the advertiser.
In addition to this, other friends outside Kyoto are following up on the relevant business, and the package is an example for them. The person in charge also predicted that, as the model grew larger, the pressure on calculus and delays became more radical, the retrofit would necessarily move from advertising-related reasoning to more upstream and downstream.
This echoes China’s openness: It is hoped to attract two types of partners, hardware manufacturers, based on their IP development products, which can be servers or storage plants for SSD protocols, and software-level clients, which can create architecture and software programme-level innovations based on their protocols. Each of the capabilities of the pits that Kyouta and Zeng have stepped on together, grinding into the layers of the agreement and the edge of the chip, will eventually become the pointer for laterers to make less detour.
The third is to explore other cutting-edge technologies, such as sharing SuperCache (L2, L3 amplification, interconnection of computing and intelligence), Agent ' s heavy sandbox (second class and 100,000 class sandboxes, five-fold overscore of resources, free movement across nodes), and multi-modular integration search.
Finally, Agenic will be the next focus. AI agent is suited to the complex, limited number of decisions and the need to optimize them; the extremely well-defined scenario is not necessary for Agentization. According to the director of the application of algorithms in the Ministry of retail advertising and research in Kyoto, advertising and the search for natural scenes exist, and the Kyoto ads seek to make all the scenes where decision-making space is large and limited and where optimal decision-making is required. However, the Officer-in-Charge noted that there were four prerequisites to this: whether the data were enough, whether the calculus was good, whether the base model capacity was strong enough, and whether it could support the model ' s cycling trajectories; and whether going down to the bottom level, to the level of computing and communication, the challenge would be greater than now.
Against this background, there is a much more far-reaching judgement on the bottom stage. In order to provide a trend at the Summit: entering Agenic AI, AI agent will rely heavily on CPU for scheduling, memory processing, sandboxing, etc., and the CPU-NPU ratio will move from 1:4 to 1:1. The more radical one in Kyoto from the operational side — 1:1 — is simply not enough. From the complexity of the operation, it may be 4:1. There are two reasons for this: the number of data to be processed is increasing, and single card performance itself is increasing. “Supernodes may be a hardware concept, but they are in fact far from it.” This means that operating systems, movement plans, and intermediates change hardware patterns with the target.
Let's go back to the account of the electrician mentioned at the beginning of the article, "If we could optimize a millisecond, we would have to raise tens of millions." In advertising, the optimization of delays is endless. The end-to-end reasoning of a 15 per cent reduction of latency is only the beginning, and the star sea of Agenic AI is the focus of future cooperation between the two sides. (This article was launched on titanium medium APP)
Auto-fetched and translated by AI Atlas. All rights belong to the original author.