文 | 数智奔流，作者 | 白芨，编辑 | 沈校
10月29日，XHEART在苏州发布智驾芯片X7 Ultra，CEO李宗霖透露，X7 Ultra具有700 TOPS大模型算力，下一代X9则要冲击3000TOPS。
作为公司首款大模型定义芯片，在量产智驾模型实测中，X7 Ultra跑出98.7%NPU计算利用率、内存带宽消耗下降80%、推理能效提升100%的成绩。媒体报道显示，测试模型即Momenta的R7世界模型。
就在XHEART发布会第二天，地平线创始人余凯发了一条微博，把矛头对准行业的等效算力折算口径：
“现在我们业界竟然又突破了宇宙尺度，标称算力甚至可以吹到6-7倍的稠密算力，号称等效算力。”
余凯拿自家征程6P举例，NV算力口径下有560 TOPS，如果用所谓的等效算力算法，同一块芯片就能写成1960 TOPS。
两方分歧在于口径，余凯要求统一折算方法，杜绝注水；李宗霖要求拿量产模型实测，强调实际利用率。双方侧重点不同，却在一件事上达成一致，即唯TOPS论不可取。
更大的行业背景是，算法和硬件正在互相靠拢。大模型改变了车端的主要计算任务，算法从大量人工规则转向数据驱动，芯片重心从峰值算力转向真实模型下的能效。
就在六天前，千里科技也发布了一款面向AI大模型定制的毫米波雷达。团队在雷达里加了一个10 TOPS的NPU，既可以在雷达端完成目标识别，也可以向中央计算平台传输原始探测数据，让雷达定位从采集传感器，转向有独立判断能力的感知单元。
二者都把算法提前放进硬件设计，减少对标准硬件的被动适配。千里希望自己决定雷达保留哪些信号，XHEART希望让模型决定芯片算力花在哪。
对智驾企业而言，纯软件供应商和纯芯片供应商的边界都在消失，将模型需求前移到硬件定义阶段，更符合行业趋势。
但这条路有物理天花板。算法嵌入硬件无法让摆脱硬件物理上限的局限性。基于Momenta定制的XHEART芯片，以更窄的适配性换取算力效率，但同时又失去了更大的开放和兼容性。
通用芯片算力利用率低，在业内已经有共识了。
英伟达Orin 254 TOPS的标称算力有个条件。这是INT8精度（低精度整数运算）下的最高值，其中87 TOPS来自深度学习加速器DLA。模型只要有一段运算交不到DLA手里，就无法调用全部的254TOPS算力。
何小鹏曾批评，通用GPU芯片的算力利用率只有30%到40%，即便在更乐观的估算下也只有70%到80%。
小鹏通用智能中心负责人刘先明在2026年CVPR上披露的三组实测数据显示，开源模型配通用芯片，计算利用率为22.8%，推理耗时800毫秒；开源模型换上小鹏自己的图灵芯片，利用率升到35.1%，耗时降到300毫秒；图灵芯片再配上小鹏自研的第二代VLA模型，数据变成82.5%和80毫秒。
理想走过另一条路。前理想自动驾驶高级算法专家詹锟透露，2024年理想为了在Orin上部署大模型，团队曾与英伟达协作，重写了PTX底层指令，并开发自研引擎，最终把推理延迟压了一个数量级。
XHEART则在理想的基础上更进一步。
从芯片定义阶段就由Momenta介入开发，将传统NPU的小核切分改为超大核，同时为高频低体量任务配置专用加速引擎，与硬件配套的编译工具可以查看整张网络，安排哪些步骤连续执行，哪些结果值得留在芯上。
XHEART的底气来自Momenta。XHEART脱胎于Momenta芯片项目，二者同源分拆、无直接股权关系、业务深度绑定，现任XHEART董事长李俊是前Momenta芯片负责人，Momenta CEO曹旭东通过XHEART的控股股东间接持有其股份。
按曹旭东10月底披露的数据，Momenta的智驾搭载规模已达约140万辆，相关车辆累计行驶超过200亿公里。这让XHEART能提前拿到长尾场景数据，了解芯片应该预留多少冗余。对于重资产投入，架构确定后无法修改的智驾芯片而言，减少了极大试错成本。
但X7 Ultra也面临难题，即模型与芯片之间的时间差。
芯片从架构到量产需要多年，模型可能几个月就迭代一次，该为未来预留多少空间，很难被量化。
如千里科技的AI毫米波雷达，通过把雷达从固定规格模式改成可OTA升级来延长有效寿命，端侧模型可通过OTA持续更新，下一代模型如果只需要更好的目标分类和轨迹预测，OTA可以延长雷达的使用周期；但涉及采样率、输入带宽等硬件维度的提升，OTA就无能为力。
对XHEART而言，X7 Ultra留给缓存的面积难再调整。在下一轮融资的窗口期，X7 Ultra能否经受住至少一次模型换代检验，这至关重要。
“Momenta搞绑定芯片，本质上是技术路线的收束，没必要多方下注了，可以给自己下重注。”有智驾行业人士对「数智奔流」表示。
作为智驾算法企业，此前智己汽车开发基于Thor的量产智驾方案时，合作名单里也有Momenta。到了上汽大众ID. ERA 9X，Momenta R7世界模型的首发车型用上了XHEART X7。
早期的硬件设计需要明确的模型和场景，芯片需要真实车辆数据验证，车规认证需要稳定的量产项目，整个链条需要数据和现金流持续回流，这时候跨平台适配的安卓模式效果较好。
这是Momenta相较地平线的优势所在，不需要押注任何一颗芯片，可以无视硬件格局的变化去跑模型。在路线不确定的阶段，把筹码分散在各个芯片平台上下注，风险较低。
但车企价格战正在压低单车软件授权费用，留给软件供应商的议价空间在减少。2026年半年报业绩会议上，Momenta提到价格战正在冲击单车软件许可售价，但公司不会无底线降价。
曹旭东把芯片筹码押给自己，有一条清晰的时间线。
2023年，Momenta将两段式端到端方案带入量产，同期确认大模型定义芯片的判断并开始推进芯片协同；2024年完成基于模仿学习的一段式端到端技术；2025年推出R6，提升强化学习地位；到2026年R7世界模型量产首发，技术路线从分段式端到端方案转向一段式端到端驾驶模型，加世界模型。
如今曹旭东在发布会上明确，端到端大模型、世界模型已经成为确定的智驾技术主线，智驾从规则驱动转向数据驱动。沿着这条路走，模型需要怎样的硬件计算，已经可以交付芯片团队讨论。
跨平台这笔生意，Momenta也没放。除了英伟达Orin和Thor，它和高通公布的合作横跨骁龙Ride的多个计算平台，X7 Ultra只是给Momenta多了一套参与原生设计的选择，模型本身仍保留跨平台能力作兜底。
换言之，Momenta算法正从完全开放转向分层开放，底层提供跨品牌芯片适配；顶层提供性能最优、体验一致性最好的X7 Ultra+R7组合。
按照李宗霖的说法，X7 Ultra会单独销售，和Momenta的算法没有排他关系，车企可以分别采购，并拿智己LS6搭载自研IM AD系统做反例。这款车型中，智驾算法上与Momenta深度绑定，但系统使用智己自己的IM AD品牌，芯片则选择英伟达Orin X。
对比地平线模式，企业正从芯片侧向算法和整套智驾方案延伸。
从征程2到征程5，它向客户提供芯片、算法授权和硬件参考设计，到征程6P又下场做全套HSD软件，随后推出的HSD Together允许合作伙伴获得基座模型授权和算法适配服务，车企与供应商可以自己做车型适配、功能定制和系统集成。
两家公司都在谈开放，在软硬一体里找平衡点，区别在于，余凯要用一套已经跑通的软件，降低客户选征程芯片的门槛；曹旭东要用一颗提前参与设计的芯片，降低自家模型上车的难度。
要挑战英伟达，XHEART面对的不只是一颗Orin。
理想的工程师在Orin上改写PTX指令，证明了通用芯片的优化空白，但也能证明Orin的生态优势。
工程师可以借CUDA写自己的计算程序，TensorRT负责优化推理执行，DRIVE OS提供车端软件基础。一个算子不适合交给专用加速器，开发者可以评估是否转给GPU，再评估产生的额外开销。
这个办法可能耗时、费电，却让工程师有机会在原芯片上交付新功能。车企明知Orin的算力存在大量浪费仍然买，也在确保智驾算法能以更快速度、更低不确定性落地。
当车企要为未来OTA功能升级预埋空间，冗余算力就成为对冲不确定性的方式。
当前智驾功能迭代极快，几个月不上车就可能被对手甩开，通用GPU架构能最大限度保持对不同算法的适应性，CUDA生态已经完成工具绑定和开发者积累，能确保算法团队的训练流程稳步推进，这是车企选择的重点所在。
一部分冗余算力也难以完全解释为浪费。GPU为了保持什么都能算的通用性，须预留大量调度和冗余逻辑，这在智驾场景下本身就在消耗资源。另一些余量，则是为了在多个摄像头同时送来数据、泊车任务与障碍物检测争用资源时，最慢的那几帧仍然按时完成。
同时，系统要为故障检测、冗余计算和降级接管预留资源。这些平时闲置的资源，构成了智驾系统的安全冗余。
这是X7 Ultra的真正难题，专用芯片再高效，如果换上去要花时间适配、重写工具链，省下的算力很快会被工程成本吞掉。
对X7 Ultra来说，它要把单一模型的深度优化转成多模型可迁移能力，开放编译器、算子库和SDK，降低第三方算法团队的适配成本，公开同条件下的延迟、功耗、模型覆盖和持续运行数据，证明架构可以跨模型复用。
当前XHEART已进入量产节奏。公司披露X7系列量产八个月，累计出货超过20万颗，已搭载10款量产车型，并获超过40款车型订单。
雷诺、奇瑞、长鑫等产业投资者入场，台积电和上汽出现在发布会，表明XHEART的产业位置初步稳固。而能否迈出这一步，将决定XHEART仍是Momenta的芯片部门，还是真正独立的芯片企业。
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。
♪ We've got a lot of ideas ♪
On 29 October, XHEART at publish Smart Driving Chip X7 Ultra in Suzhou, CEO Lee Jong-soon revealed that X7 Ultra had 700 TOPSLLM algorithms and that the next generation X9 would impact 3,000 TOPS.
As the company's initial LLM definition chip, X7 Ultra ran 98.7 per cent of NPU calculated utilization, 80 per cent reduction in memory bandwidth consumption, and 100 per cent performance in reasoning energy efficiency improvement in quantitative driving models. Media reports indicate that the test model is the R7 World Model of Momenta.
On the second day of the XHEARTpublish meeting, the founder of the horizon, Yu Kai, sent a tweet that aimed at the equivalent calculator of the industry:
“Now that our industry has broken the universe's scale again, it can be called a dense, six-to-seven-fold calculator, called the equivalent.”
The example of Yu Cai for 6P in the home journey is 560 TOPS under the NV calculator, which can be written into 1960 TOPS with the so-called equivalence algorithm.
The difference between the two sides was in calibre, and Yu Cai called for a uniform method of conversion to eliminate water injection; Li Zongcheng asked for a quantitative model to be measured, emphasizing actual utilization. There is a difference in focus, but there is agreement on one thing, namely, that TOPS is not desirable.
The larger industry context is that algorithms and hardware are coming closer to each other. LLM changed the main computational tasks of the end of the car, with algorithms shifting from a large number of manual rules to data-driven, and the core of the chip moving from peak to real model energy efficiency.
Just six days ago, publish, a millimetre radar designed for AZLLM. The team added a 10-TOPS NPU in radar, which can either complete target identification at the radar end or transmit raw detection data to the central computing platform, allowing radar positioning to move from collecting sensors to sensory units with independent judgement capability.
Both place algorithms in advance of hardware design and reduce passivity to standard hardware. Thousands of miles want to decide which signals the radar keeps, and XHEART wants the model to decide where the chip is.
For smart drivers, the boundaries of pure software suppliers and pure chip providers are disappearing, moving model demand ahead to the hardware definition stage, more in line with industry trends.
But this road has a physical ceiling. Algorithms embedded in hardware do not allow them to escape the physical limit of hardware. The XHEART chip, tailored to Momenta, offers a narrower adaptability in exchange for more efficient calculation, while losing greater openness and compatibility.
Common chips are under-utilized and there is consensus within the industry.
We've got a condition for the nominal calculus of Orin 254 TOPS. This is the highest value under INT8 precision (low accuracy integer) of which 87 TOPS comes from the DLA deep learning accelerator. As long as the model does not have a single calculation in the hands of DLA, it will not be able to access all 254TOPS.
He Xiao Peng has criticized the use of only 30 to 40 per cent of the computing power of the generic GPU chip and, even under more optimistic estimates, between 70 and 80 per cent.
The three sets of empirical data revealed by Liu Xianmin, the head of the General Intelligence Centre for Plum, on CVPR in 2026, showed that the open source model had a common chip utilization rate of 22.8 per cent and that the reasoning took 800 milliseconds; the open source model had been replaced by a peng's own Turin chip, which had increased to 35.1 per cent and had been reduced to 300 milliseconds in time; and the Turin chip had been supplemented by a second generation VLA model developed by Peng itself, which had become 82.5 per cent and 80 milliseconds.
Ideals go another way. The former expert in advanced algorithms, Zang Zhuang, who was the first to operate LLM on Orin, revealed that in 2024 the team, in collaboration with Inverda, had rewritten the PTX ground-level directive and developed a self-study engine, ultimately delaying reasoning by one order of magnitude.
XHEART builds on the ideal.
From the chip definition stage, Momenta has been involved in the development of small nuclei for traditional NPUs to supernucles, with dedicated accelerators for high frequency low volume missions, and hardware-compilation tools to view the entire network, arrange for steps to be carried out successively and for results to be retained on cores.
XHEART's base air is from Momenta. XHEART was born from the Momenta Chip project, which was split from one source to the other, without a direct equity relationship, and tied to the depth of business. The current XHEART Chairman, Lee Joon, was the former head of the Momenta Chip, and Momenta CEO Zhoudong held his shares indirectly through the controlling shareholders of XHEART.
According to data released at the end of October, Momenta ' s smart-riding capacity reached approximately 1.4 million, with the associated vehicles travelling over 20 billion kilometres. This allows XHEART to get long-end scene data in advance and to know how much redundancy the chip should have. For a heavy asset input, a smart drive chip that cannot be modified after the architecture has been established has reduced the cost of a significant trial error.
However, X7 Ultra also faces the challenge of the time difference between the model and the chip.
The chip takes years from architecture to mass production, and the model may be repeated once in a few months, leaving much space for the future and it is difficult to get quantization.
For example, in the case of the 1,000-mile technology AImm wave radar, the active life can be extended by converting the radar from a fixed specification model to an upgradeable OTA, the end-side model can be continuously updated through the OTA, which can extend the duration of the radar if the next generation model requires only a better target classification and trajectory prediction; but with regard to the upgrading of hardware dimensions such as sampling rates, input bandwidth, etc., the OTA is powerless.
For XHEART, the area of X7 Ultra left to cache is difficult to adjust. Whether X7 Ultra can withstand at least one model replacement test during the next round of financing window is critical.
“Momenta binds the chip, essentially a technical route, so it is not necessary to bet more than one.
As an intellectual driving business, Momenta was also included in the list of cooperations in the former development of the Thor-based mass-driving programme. To the upper steam mass ID. ERA 9X, the Momenta R7 world model's starter model uses XHEART X7.
Early hardware designs require clear models and scenarios, chips require real vehicular data validation, vehicle certification requires stable volume projects, and the entire chain requires data and a continuous flow of cash flows, at a time when the Anjo model, which is suitable across the platform, is working better.
This is the advantage of the Momenta parallel horizon, which does not require a bet on any chip and can run models regardless of changes in hardware patterns. At an uncertain stage in the course of the route, the chips are spread over the chips platforms, with a lower risk.
However, price battles for motor vehicles are reducing the authorized cost of cycling software and the bargaining space available to software providers is decreasing. At the semi-annual performance conference in 2026, Momenta mentioned that price wars were impacting on the licence price of a bicycle software, but that companies would not reduce the price without a bottom line.
Cao Xudong took the chips to himself with a clear timeline.
In 2023, Momenta brought in two end-to-end programmes, confirmed the judgement of the LLM defined chip during the same period and began to advance chip synergy; completed a one-end-to-end technology based on imitation learning in 2024; introduced R6 in 2025 to enhance the status of enhanced learning; and by 2026, the R7 World Model was launched, with a technology route moving from a segment-to-end programme to a one-end-to-end driving model, plus a world model.
Today, at the publish meeting, Cao Xudong made it clear that the end-to-end LLM, the world model, has become the established master of smart driving, moving from a rule-driven to a data-driven drive. Following this path, what kind of hardware the model requires is already available for discussion by the chip team.
The cross-platform business didn't go on either. In addition to Young Weidar Orin and Thor, it works across the board with Takato. DragonRide ' s multiple computing platforms, X7 Ultra, merely gave Momenta a set of options to participate in the original design, and the model itself retains a cross-platform capability to do bottom-up.
In other words, the Momenta algorithm is moving from full openness to stratification, with the bottom providing cross-brand chip matching; the top layer providing the best performance and the best experience of a consistent combination of X7 Ultra+R7.
According to Li, X7 Ultra would sell separately and there would be no exclusionary relationship between the algorithms of Momenta, so that the car company could purchase separately and carry the intellectual body LS6 on the IMAD system as an example. In this model, smart driving algorithms are tied deep to Momenta, but the system uses its own IM AD brand, and the chip chooses the British Orin X.
In contrast to the horizon model, businesses are extending from the side of the chip to algorithms and the whole set of smart driving programmes.
From the 2nd to the 5th journeys, it provides the customer with chips, algorithmic authorization and hardware reference design, to the 6th expedition, where the complete HSD software is developed, and the subsequent HSD Together, which allows partners to obtain base model authorization and algorithmic adaptation services, allows companies and suppliers to adapt, customize and integrate their own systems.
Both companies are talking about opening up and finding a balance in a soft and hard one, the difference being that Yucai is going to use a run-through software to lower the threshold for client selection of the entrance chip; Cao Xudong is going to use an early participatory chip to make it easier for his own model to get in the car.
To challenge Yvette, XHEART is not just an Orin.
The ideal engineer rewrites the PTX directive on Orin, proving that there is an optimisation gap in the generic chip, but also the ecological advantage of Orin.
Engineers can write their own calculations using CUDA, TensorRT is responsible for optimizing reasoning, and DRIVE OS provides the end-of-pipe software base. A algorithm is not suitable for handing over to a dedicated accelerator, and the developer can assess whether it is transferred to GPU and then evaluate the additional costs incurred.
This approach, which may be time-consuming and costly, gives engineers the opportunity to deliver new functions on the original chip. Knowing that Orin's calculus is wasteful and still being bought, car drivers are also ensuring that smart driving algorithms can land at faster and lower uncertainty.
When companies are planning to upgrade space for future OTA functionality, redundancy is the way to hedge uncertainty.
The current intellectual driving function is extremely rapid, with several months of non-carriage likely to be lost to rivals. The universal GPU architecture can maximize adaptation to different algorithms. CUDA ecology has completed the binding of tools and the accumulation of developers. It ensures that the training process of the algorithm team is steadily advancing, which is the focus of the car company ' s choice.
It is also difficult to fully interpret redundancy as waste. In order for GPU to remain universal for all purposes, it is necessary to set aside significant movement control and redundancy logic, which in itself consumes resources in the context of intellectual driving. The other remaining amounts were intended to be completed on time while multiple cameras were being sent together to compete for resources for data, parking assignments and barrier detection.
At the same time, the system has to set aside resources for failure detection, redundancy calculations and downgrading. These resources, which are normally idle, constitute a security redundancy in the intellectual driving system.
This is the real challenge of X7 Ultra, where dedicated chips are more efficient and, if replaced with a time-appropriate, rewrite tool chain, the savings will soon be swallowed up by the cost of the project.
For X7 Ultra, it will convert the depth optimization of a single model to a multi-model portability capability, open compilers, algorithms and SDK, reduce the fit costs for third-party algorithm teams, and demonstrate that structures can be reused across models by delayed, functional, model-covered and ongoing data under open conditions.
XHEART is now entering the quantum rhythm. The company disclosed that the X7 series produced over eight months, cumulatively delivered more than 200,000 goods, carried 10 car types and received over 40 car-type orders.
The entry of industrial investors such as Renault, Chire and Changju, and the presence of electricity build-ups and up-to-date gas at publish, indicate that XHEART ' s industry is initially well established. Whether or not this step can be taken will determine whether XHEART remains the chip department of Momenta or a truly independent chip business.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.