文 | 最话FunTalk，作者 | 王芳洁
第一颗芯片回片那天，新芯航途（XHEART）的一位同事发现，自己桌上的日历上写着“今天有好事发生。”他赶紧把这句话拍下来，发到了公司群里。
果然，不到24小时内，这颗芯片被点亮了。
整个事情当中，最引人注目的当然是那颗芯片，但还有一个人不该被忽略，当他看到好事，首先想到的是公司，然后福至心灵。
这真是一家幸运的公司，无论在事上，还是在人身上。
成立不过三年时间里，XHEART就已经推出了第一代大模型芯片X7系列，并实现了规模化量产上车。
10月29日， 这家年轻的公司召开发布会，宣布推出新一代大模型芯片——
X7 Ultra，一款拥有700 TOPS大模型算力的车规级芯片。并且通过架构与工具链协同创新，X7 Ultra支持可部署模型参数量实现10倍提升。
和X7系列一样，X7 Ultra仍然是一款由大模型定义的芯片，芯片与模型的适配被前置到了研发链路的最前端。
并且，XHEART也仍然坚持基于真实量产智驾大模型进行评测。因为在XHEART CEO李宗霖看来，通常行业用实验室里的基础测试评估芯片性能，但实验室结果与量产模型的实际体验往往相去甚远。
根据真实量产模型的评测，X7 Ultra量化部署后模型输出与训练后FP32浮点模型的相似度达到了99.99%。
四个 9的概念并不令人陌生，人们经常在黄金领域看到：纯度 99.99% 的黄金一旦被炼造出来，就会被称为足金。某种程度上，打造一款芯片也很像炼金术，蚀刻硅片，然后得到改变人类世界的科技文明。
在X7 Ultra背后，的确有一群炼芯术士，他们有的在XHEART内部，有的在产业链上下游，他们都相信被大模型定义的芯片，能够充分释放模型能力，支持中国智驾实现L4级别自动驾驶。
“今天有好事发生”，是他们的咒语。
01 几年后，市场需要什么样的芯片？
2023年，OPPO关闭了芯片设计公司哲库之后，时任哲库COO的李宗霖决定带着核心团队一起创业。他是中国芯片行业非常资深的专家，在哲库之前，李宗霖是联发科无线通信事业部总经理，主导过 Helio 处理器系列和天玑芯片系列的开发。
但哪怕在芯片行业做了快二十年，李宗霖仍然无法回避芯片行业的一大痛点：模型迭代速度很快，但芯片的研发周期以年计，那么等芯片做出来，还能满足模型的需要吗？
而在同一时间，身为Momenta的创始人，曹旭东也有自己的疑问。按照他的判断，2028 年将是 L4 级别自动驾驶的爆发点，届时模型规模将非常庞大，需要至少10B、大概率 100B 参数的模型，而现有通用芯片无法支撑——车上的功耗、车规的可靠性和成本，是一道"不可能三角"。
如果强行用通用芯片堆算力，相当于在车上放一台巨大的 GPU 服务器；向功耗和成本妥协，模型规模又做不上去。该怎么办？
机缘巧合下，两个抱有困惑的人碰到了一起。他们深聊了几次，发现这两个问题都有同一个解法——
由大模型来定义芯片。
很快，曹旭东与李宗霖决定联袂，一起做芯片。他们互相需要。
李宗霖团队需要和 Momenta 合作，让模型在最早期就介入到研发当中，先理解模型如何运行，再倒推硬件架构与软件工具链。
Momenta 同样需要李宗霖团队，因为要支撑 L4 级别的目标，它需要一款能支持模型规模的高性能芯片，而这款芯片得能上车量产，所以还得同时满足能效和成本问题，打破不可能三角。
同年，XHEART成立了, 从成立的第一天起，XHEART便确立了“大模型定义芯片”的技术路线：从真实应用中的模型需求出发，倒推芯片架构，重新组织计算、存储和数据流转。团队从芯片定义阶段就与算法团队共同研究模型架构和关键算子，为未来几年的模型演进提前做好准备。
按照这条技术路线做出来的芯片，不会被时间打败。
以X7 Ultra为例，其超大核 NPU 把模型的数据交互从传统架构的百 GB 每秒级，聚合到片内 TB 每秒级；
针对 Transformer 模型中 Linear、Softmax、Transpose 等关键算子，X7 Ultra专门设计了硬件加速引擎，且原生支持浮点高精度向量计算；
通过底层软硬协同，X7 Ultra 的计算利用率达到 98.7%；
通过AI工具链重构，内存带宽消耗相较行业旗舰模型下降 80%，能效提升 100%。
如此性能的芯片被做出来后，就不光能支持现有模型了，它还会反向推动模型的迭代。
据曹旭东介绍，Momenta即将推出的R7 Ultra世界模型，基于X7 Ultra芯片，模型参数的规模扩大到了原来的5-10倍，安全的性能表现到现在也有5倍到10倍以上的提升。
同时，推演万物的能力开始涌现。
02 从点亮到上车
作为兄弟公司，XHEART和Momenta独立运作，股权上没有直接隶属关系，但因为大模型定义芯片技术路线，两家公司的很多工作从一开始就深度协同。
2023 年国庆假期，Momenta的模型团队和XHEART的芯片团队没有放假，他们一起围着持续演进的模型架构，一项一项分析计算时间花在哪里，再倒推硬件应该为未来做什么准备。
注意这个时间，那时XHEART几乎还是一个“新生儿”。不难看出，传统芯片行业的工作流被颠覆了，李宗霖和SoC 设计负责人讨论最多的一件事，就是怎么把对大模型的判断真正做进芯片架构里。
时间很快来到2025年，有好事发生的那一天，第一颗芯片回到了公司，几乎所有的研发人员都围了上来。
有人在实验室的白板上记下了这个时刻：15:30，芯片到达。随后，20多个部门都先后在白板上签下了自己负责的工作，完成一个就写一个，标注好时间，精确到分钟。
不到 24 个小时后，第一颗 X7 被点亮了。
这几乎是一个创造性的周期。芯片点亮是一个环节非常多的过程，按行业经验，大型芯片点亮顺利的话要几周或一两个月，不顺利的话，拖上半年也是常有的事。
点亮后的上车测试，进行得也异常顺利——但"顺利"的前提是，团队先给自己加了一道难题。
按照行业通用做法，芯片的上车测试通常会选择在封闭道路上进行。但李宗霖觉得不行，他希望交付的是一颗真的能让真实车辆在道路上稳定可靠工作的芯片。
所以，XHEART决定用更严格、更接近量产的标准来重新定义"回片上车"这个节点。
它要验证的不是芯片单项性能，而是由芯片、底层软件和智驾模型组成的完整系统，能不能在真实的道路上跑出接近量产的表现——细致到每一个路口的转弯，模型规划出来的行驶轨迹都必须与量产标准保持一致。
这个标准，对一颗刚回来的芯片是非常有挑战的。经过团队的日夜攻坚，整个系统从芯片到底层软件再到算法，磨合得越来越稳定。
首台测试车在短短几天后就成功上路；17 天后，X7 整车闭环完成，测试车可以在区域的城区道路上连续稳定行驶数小时，全程无接管；无论是白天夜晚，还是那段时间经常出现的暴雨天气，系统都经受住了考验；第 21 天，搭载 X7 的测试车驶入真实车流，向客户开放体验。
按照行业规律，通用芯片从拿到手到"跑通、跑稳、跑对、跑好"，顺利的话要三到六个月，不顺利要六到十二个月，甚至更久；而 X7 完成这一切，用了不到一个月。
为什么XHEART可以这么快？发布会上，李宗霖解释了原因——
因为量产模型和实车数据提前进入了硅前验证，从芯片的设计，从模型性能的验证，在流片之前都已经做完了。
03 只买单程票的旅程
在上汽集团总裁贾健旭的记忆里，也有一个瞬间，是一辆车汇入了车流。那辆车上搭载了 Momenta 智驾系统，曹旭东陪他一起试驾。
在车企当中，上汽是 Momenta 最坚定的同路人。早在 2021 年开始，上汽集团就多次投资 Momenta，是后者最大的机构投资者。智己、上汽通用、上汽奥迪、上汽大众均与 Momenta 达成了战略合作，在多款量产车型上搭载其智驾系统。
这种深度合作中，上汽投入的不仅是资金和订单，还有信任——而当那辆测试车灵活地避开所有路障、稳稳将贾健旭送达目的地时，这种信任得到了交付。
所以，那一路，贾健旭有些感动。他和曹旭东聊了很多。突然想到这家做软件的公司，有没有做硬件的计划，以实现双轮驱动？
曹颇为神秘地说，已经在路上了。
当时正值X7芯片的研发周期，显然，曹旭东相信事以密成。
今年上半年，作为首款搭载 X7 芯片的量产车型，上汽大众 ID.ERA 9X 也上路了。陆续地，更多量产车汇入真实车流。
到今天为止，X7 系列芯片已经有了 10 款量产车型，获得超过 40 款车型的订单。
这 40 款车型背后，包括上汽在内的所有厂商，都向XHEART这样一家年轻的芯片公司投入了信任成本。和当年上汽选择 Momenta 时一样，XHEART当时也没有可以证明的成功案例，它同样要在路上证明自己。
这是一趟只买单程票的旅程。
在XHEART，这不是一个比喻，而是一个真实的故事：在某款重磅车型上市前，公司一位同事准备去客户现场，出发前 HR 问什么时候回来，他说："我也不知道，就先买单程票吧。"
在这家公司，归期是交付客户信任的那一天。
但客户的信任怎么会有被完全交付的那一天？值得注意的是，汽车厂商与智驾解决方案供应商之间的关系非常紧密，协同的深度也远超一般零部件采购。
正如李宗霖所言，把芯片做出来之后，他们还会进一步提供丰富的解决方案，协助客户降低开发难度，让项目更快地走向量产。
所以，XHEART手上拿的，也是一张单程票。好吧，现在，这是一个比喻。
这个比喻不仅适用于它和客户之间，也适用于它和供应商之间。车规芯片是一个特殊行业，它处于汽车产业链中游，在 3 年这么短的时间内实现量产，不是靠XHEART一家公司就能完成的，它是产业链上下游共同做功的结果。
是的，XHEART还得成为上游的“同路人”。
首先，需要确认的事实是，在 X7 之前，XHEART的产品序列是一张白纸，它的兄弟公司 Momenta 成立时间也不长，并且是一家软件公司。但台积电成为了XHEART的供应商。
众所周知，当下想要拿到台积电的产能是非常难的，创业公司更是难上加难。3 年前，在和曹旭东深聊之后，台积电副总经理陈平决定接下XHEART的单子，尽管当时他觉得对方说的 L4 级别自动驾驶像是科幻片。
陈平坦言，如果换个人来讲，他是会怀疑的，但眼前的曹旭东让他选择了相信，同时李宗霖也加码了这份信任，"在我们行业里，李宗霖一直被公认是最可靠、最靠谱的人"。
3 年后的今天，XHEART量产已经 8 个月，累计出货超过 20 万颗。
发布会前，一位媒体人给曹旭东发来了两段短片：一辆搭载 X7 芯片和 Momenta 智驾系统的车开在复杂车流里，复杂路况下倒车避让，多把掉头，车表现得比老司机还老司机。
这其实就是曹旭东当年和陈平聊的东西。今天，那个几乎无声、快速自主旋转的方向盘，兑现了当初的相信。
在商业合作当中，信任关系不是一种闭环结构，而是会随着一次次兑现loop起来。
例如当智己采用Momenta智驾系统获得令消费者满意的结果后，上汽的更多车型也开始采购Momenta，当上汽大众ID.ERA 9X量产上市后，MG 07、荣威家越07也开始搭载X7芯片。
10月29日，李宗霖还披露了X9芯片的进展，其大模型算力达到3000TOPS,内存带宽达 700GB/s的芯片。
这些参数就足以说明，在下一代模型问世之前，XHEART先做好了准备。这家公司的技术战略是“量产一代、开发一代、预研一代。”
它的供应商也做好了准备。
陈平在发布会上留下伏笔：目前最先进的车规工艺是 3 纳米的 N3AA，而 L3、L4 对算力、功耗和面积的约束，决定了"非用 3 纳米不可"。
台积电的承诺是："我们的工艺已经准备好了，你们把最好的创造力、最好的设计送过来，我们一起为世界创造最美好的东西。"
他们的确都想让好事发生，曹旭东、李宗霖、陈平、贾健旭……，以及每个炼芯术士。
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。
"FunTalk," by Wang Fong Jie.
On the first chip return, a colleague of XHEART found that the calendar on his desk stated, "Good things happen today." He quickly filmed it and sent it to the company.
It's been lit in less than 24 hours.
Of course, the most striking thing about this whole thing is the chip, but there's another person who shouldn't be ignored when he sees good things, first thinking about the company and then the soul.
It's a lucky company, in matters and in people.
In only three years of its existence, XHEART had launched the first generation of LLM chips X7 series, and had achieved size-based capacity to board.
On October 29th, this young company held a publish conference to announce the launch of a new generation of LLM chips.
X7 Ultra, a car code chip with 700 TOPSLLM algorithms. And X7 Ultra supports a 10-fold upgrade of the deployable model parameter count by synergizing the architecture with the tool chain.
Like the X7 series, X7 Ultra remains a chip defined by LLM, and the suitability of chips and models is pre-positioned to the front end of the R & D chain.
Moreover, XHEART continues to insist on evaluation based on real volumes of intellectual driving LLM. Because, according to XHEART CHEO Lee, the industry usually uses basic tests in laboratories to assess the chip performance, but laboratory results are often far from the actual experience of the quantum model.
The X7 Ultraquantization post-deployment model output and the post-training FP32 floating point model were measured to be 99.99 per cent.
The concept of four nines is not strange, and is often seen in the gold sector: 99.99 per cent of pure gold, once refined, is called gold. To some extent, building a chip is also like alchemy, carving silicones, and then getting the technological civilization that changes the human world.
Behind the X7 Ultra is a group of core magicians, either inside the XHEART or up and down the industrial chain, who believe that the chips defined by LLM can fully release model capabilities and support Chinese intellectual driving to achieve L4-level self-pilot.
"Good things happen today" is their spell.
What kind of chip does the market need in a few years?
In 2023, after OPPO closed down the philosophy bank of the chip design firm, Li Jong-hwan of the then Chul-Kook Coo decided to start a business with a core team. He was a very senior expert in the Chinese chip industry and before Chulku, Li was the General Manager of the Radio Communication Department of UNDGC, leading the development of the Helio Processor series and the Skychip Chip series.
But even after nearly 20 years of work in the chip industry, Li Zongjun can't escape the core. One of the major pains of the sector is that the models are highly iterative, but the development cycles of the chips are measured year by year, so can they meet the needs of the models?
At the same time, as the founder of Momenta, Cao Xudong had his own doubts. According to his judgement, 2028 would be an L4-level self-driving flash point, when the model would be very large and would require a model of at least 10B, with an approximate rate of 100B parameters, and the existing generic chip would not be able to support it — the power of the vehicle, the reliability of its rules and the cost of it — is a “no triangle”.
It would be equivalent to putting a large GPU server in a car if it were to be compulsorily calculated using a general chip stack; it would be impossible to reconcile the scale of the model with the cost of effort. What do we do?
By chance, two confused people met together. They talked a few times, and they found that both problems were solved in the same way.
LLM defines the chip.
Soon, Cao and Li Zhongjun decided to make chips together. They need each other.
The Li Zongjun team needs to work with Momenta to get the model involved at the earliest stage in the development process, first to understand how the model works and then to reverse the hardware architecture and the software tool chain.
Momenta also needs a team from Li Jong-ho, because to support the L4 level, it needs a high-performance chip that supports the size of the model, which has to be able to produce in a car, so it also has to cover energy efficiency and cost problems and break the impossible triangle.
In the same year, XHEART was established and, from the first day of its existence, it established a technical route for “LLM defined chip”: Rewind the chip structure, reorganise computing, storage and data flow, starting from the needs of models in real applications. The team worked with the algorithm team from the chip definition stage to study model structures and key algorithms in advance of model evolution in the coming years.
The chips that follow this technical route will not be defeated by time.
In the case of X7 Ultra, its supernuclear NPU interactives the data of the model from the 100 GB per second of the traditional structure, and aggregates it into a single TB per second;
For key algorithms such as Linear, Softmax and Transpose in the Transformer model, X7 Ultra is specially designed for hardware acceleration engines, with primary support for high-precision vector calculations;
The calculation of X7 Ultra reached 98.7 per cent through soft and hard substrate synergies;
Re-engineered through the AI tool chain, memory bandwidth consumption decreased by 80 per cent compared to industry flagship models, and energy efficiency increased by 100 per cent.
When a chip of this performance is produced, it will not only support the existing model, but will also reverse the iterative model.
According to Cho Xudong, the forthcoming R7 Ultra World Model of Momenta, based on the X7 Ultra chip, has expanded the size of the model parameters to 5-10 times, and the safety performance has now increased by 5 to 10 times.
At the same time, the ability to develop everything began to emerge.
02, from light to the car.
As a brother company, XHEART and Momenta operate independently and do not have a direct reporting relationship on equity, but many of the work of the two companies has been well coordinated from the outset because LLM defines the technical route of the chip.
On the national holiday of 2023, Momenta's model team and XHEART's chip team did not take leave, and together they surrounded the evolving model structure, one analysis of where time was spent, and then what should be prepared for the future.
Look at this time, when XHEART was almost a "newborn child." It's not hard to see how the workflow in the traditional chip industry has been subverted, and one of the things that Li Jong-soon and SoC designers have discussed most is how to really put the judgement on LLM into the chip architecture.
Soon after 2025, the day when the good thing happened, the first chip returned to the company and almost all the researchers and developers were surrounded.
The moment was recorded on the whiteboard of the lab: 15:30, the chip arrived. Subsequently, more than 20 departments signed their own jobs on the whiteboard, completing one, marking the time and precise minutes.
Less than 24 hours later, the first X7 was lit.
It's almost a creative cycle. Chip-lighting is a very many-step process, and in industry experience it takes weeks or months or months for a large chip to be lighted and, if not, the first half of the year.
It's a very good time to get on the train and run it. But "success" presupposes that the team adds a problem to itself.
In accordance with industry-wide practice, the testing of chips on board is usually carried out on closed roads. However, he did not think it would be possible to deliver a chip that would actually allow real vehicles to stabilize their work on the road.
So, XHEART decided to redefine the node of "back-to-back" with more stringent and closer standards of mass production.
It's not about the individual performance of the chip, it's about a complete system of chips, bottom software and smart driving models, and it's about to get out of the real road near mass production -- – To the point of turn of every intersection, the trajectories planned by the model must be consistent with quantitative standards.
This standard is very challenging for a new chip. The entire system has become more stable, from chips to bottom software to algorithms, as a result of the team's day and night effort.
The first test vehicle was successfully on the road in just a few days; 17 days later, the X7 vehicle was closed, and the test vehicle was able to travel on the region ' s urban roads for several hours without taking over; the system was tested, both during the day and night, and during the frequent heavy rain weather; and 21 days, the X7 test vehicle entered the real traffic and was open to clients.
According to industry patterns, it takes three to six months for a common chip to get its hands on "run, run, run, run well" and six months, six to twelve months, if not longer; and it takes less than one month for X7 to do all this.
Why is XHEART so fast? At the publish meeting, Lee Jong-soon explained why.
Because volume models and actual data are pre-silicized, the design of the chip, the validation of model performance, was completed before the stream.
03, one-way trip.
In the memory of Ja Kenwu, the president of the upper-air group, there was also a moment when a car remitted into the traffic. The car was loaded with Momenta's smart driving system and Cao was with him for a test drive.
Among the cars, uplifting was the strongest of the Momenta's companions. As early as 2021, the Upper Gas Group invested several times in Momenta, the largest institutional investor in the latter. As a result, Momenta has received strategic cooperation from intellectuals, the public, the public, and the public.
In this deep cooperation, not only money and orders, but also trust — a trust that was delivered when the test vehicle managed to avoid all roadblocks with flexibility and steady delivery to its destination.
So, that's the way Ja Ken Wing moved. He talked to Cao Xudong a lot. Did the software company suddenly have any plans to make hardware to achieve double-wheel drive?
Cao is on his way.
It was during the research and development cycle of the X7 chip, and it was clear that Cao Xudong believed in secrecy.
In the first half of this year, as the first mass-production model to carry the X7 chip, the crowd ID.ERA 9X is on its way. Over time, more large volumes are being sent into real traffic.
As of today, the X7 series of chips has 10 model cars and orders over 40 models.
All the manufacturers behind the 40 cars, including the gas, invested in trust costs for a young chip company like XHEART. As in the case of Momenta, in the last year, XHEART had no proven success, and it had to prove itself on the road.
This is a one-way trip.
In XHEART, it's not a metaphor, but a true story: before a heavy-pound car comes on the market, a colleague from the company is going to go to the client's site, and before leaving HR asked when he would be back, he said, "I don't know, buy a one-way ticket."
In this company, the deadline is the day of delivery of the customer ' s trust.
But how could the trust of the client be fully delivered? It is noteworthy that there is a strong relationship between automobile manufacturers and suppliers of smart driving solutions, and that the level of synergy is far greater than the procurement of spare parts in general.
As Li Zongjun has said, once the chips are made, they will provide further rich solutions to help clients reduce their development difficulties and move the project faster towards production.
So XHEART has a one-way ticket. Well, now, that's a metaphor.
This analogy applies not only between it and the customer, but also between it and the supplier. Carchips are a special industry that is in the middle of the chain of the automobile industry and that achieves mass production in such a short period of three years, not by a company based in XHEART, but by working together up and down the chain.
Yes, XHEART will have to be a “sympathetic” upstream.
First of all, it needs to be established that before X7, XHEART's product series was a piece of paper, and its brother company, Momenta, was not long in existence and was a software company. But it became a supplier of XHEART.
It is well known that it is very difficult to obtain the capacity to build electricity at this time, especially for start-ups. Three years ago, after a deep conversation with Cao Xudong, Chen Ping, Deputy Director General of Power, decided to take the XHEART list, even though at the time he thought that what he said about the L4 level autopilot was a science fiction film.
Chen Ping said that he would be suspicious if he were someone else, but he was given the choice to believe, and Lee added that trust, "in our profession, Lee Zong Chong has always been recognized as the most reliable and reliable person."
Three years later, the XHEART production has been in place for eight months, cumulatively delivering more than 200,000.
Before publish, a mediaman sent two clips to Cao Xudong: a car carrying X7 chips and Momenta ' s smart driving system was in a complex traffic stream, with complex traffic backsliding, turning around, and the car was more experienced than the old driver.
That's what Cao Xudong was talking about with Chen Ping. Today, that almost silent, self-sustained wheel of rotation fulfils its original belief.
In commercial cooperation, trust relationships are not a closed-ring structure, but will follow the flow of loop.
For example, after the introduction of the Momenta Smart Driving System to achieve consumer satisfaction, more car models of uplifting have started to buy Momenta, and when the public ID.ERA 9X volume was put on the market, the MG 07 and the Young Wai family also started carrying X7 chips.
On 29 October, Li also disclosed the progress of the X9 chip, with its LLM calculator up to 3,000 TOPS and its memory bandwidth up to 700 GB/s.
These parameters are sufficient to show that XHEART was ready before the next generation model. The company's technology strategy is to produce, develop and prepare a generation.
Its suppliers are also ready.
Chen Ping left a pen at the publish meeting: the state-of-the-art car procedure is 3 nanonap N3AA, while L3 and L4 have determined "no use of 3 nanonap" in terms of calculus, effort and area.
The promise is, "Our process is ready, you bring the best of creativity, the best of design, and together we make the best of the world."
Indeed they all want good things to happen: Cao Xudong, Li Zongcheng, Chen Ping, Jia Kenwu... and every charlatan.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.