XHEART New Core Voyage Releases 700 TOPS Automotive-Grade Chip X7 Ultra

Building chips for large models: XHEART's three-year breakthrough

Digital Intelligence Frontline, October 30, 2026, 13:53:06 Updated October 30, 2026, 15:19:33

With only one chance, they bet on large models defining chips.

"In the era of large models, what kind of chip do we actually need?" On the afternoon of October 29, Li Zonglin, CEO of XHEART, opened with the calm, unhurried tone of a veteran chip industry figure. On the large screen behind him, "XHEART" glowed against a tech-blue background; at the entrance to the venue, crowds gathered around a row of production vehicles equipped with XHEART chips.

Over the past few years, smart-driving chips competed on peak computing power, climbing from 200 TOPS and 500 TOPS all the way to over 1000 TOPS. But after end-to-end large models and world models were deployed in vehicles, the models did not necessarily run faster or better. More troublesome still, models iterate every few months, while an automotive-grade chip takes years from design to mass production.

XHEART proposed "large models defining chips": first figure out what the models need, then work backward to define the chip architecture and toolchain. The company, founded less than three years ago, made its first appearance under its brand identity that day, unveiling its new chip, the X7 Ultra. Its previous-generation X7 series has already shipped more than 200,000 units cumulatively, is deployed in 10 production vehicle models, and has won orders for more than 40 vehicle models.

NVIDIA and Qualcomm dominate the main market, and automakers are also building their own chips. XHEART decided to enter the game with a different technological approach.

01

A question from three years ago

In 2023, Li Zonglin and several colleagues who had spent ten to twenty years in the chip industry decided to start a new venture. "Starting over actually made everyone more cautious," he said. This time, what exactly should they do, what customer problems could they solve, and when the chip came out several years later, would it be what the market needed? It was also during that period that he met Momenta CEO Cao Xudong.

After multiple discussions, the two formed a judgment: large models would change intelligent driving, and would also change the way chips are designed. In the past, smart-driving systems were assembled from multiple small models, and the ceiling of system capability depended on the rules engineers wrote down. After the rise of Transformer-based end-to-end large models, smart driving began to be truly data-driven, which also posed three questions for chips: Can computing power be brought out in real systems? Can the chip designed today keep up with tomorrow's models? Can the required capability be achieved at reasonable energy consumption?

In Li Zonglin's view, these three questions must be answered together: starting from real large-model requirements, rethink how computation, storage, and data flow are organized. This is "large models defining chips."

This also became the starting point for XHEART's deep cooperation with Momenta. XHEART was founded in December 2023, and its chairman, Li Jun, is the former head of chips at Momenta.

Cao Xudong later revealed in his speech that Momenta had formed the judgment of "large models defining chips" at the end of 2022 and the beginning of 2023, and had been putting it into action with XHEART since 2023. "Today is the first time we are speaking about it publicly."

He gave two reasons: chips defined by large models can achieve 10 times the performance of general-purpose chips; and the scale of L4 requires at least 10B-parameter models, and in all likelihood 100B-parameter models. Using general-purpose chips for this would be equivalent to putting a giant GPU server in the car, and power consumption, automotive-grade reliability, and cost cannot all be satisfied at the same time—what he called the "impossible triangle" of general-purpose chips.

He also gave a timetable. Cao Xudong called the pattern of smart-driving safety and performance improving tenfold every two years the "Moore's Law of smart driving." From 2022 to 2026, the industry has basically moved forward at this pace; next, large-model chips are expected to accelerate it to tenfold every year. He observed that over the past four years, the scale of vehicle-side models has grown roughly 5 to 10 times, and with large-model chips, there is an opportunity to grow 10 to 100 times in the next two years.

Cao Xudong revealed that Momenta's cumulative mass production has exceeded 1.4 million vehicles, with 20 billion kilometers driven, and the latest 100,000 units were delivered in less than a month. The continuously accumulated real-world data and iterating models have also become an important foundation for XHEART to judge next-generation chip requirements in advance.

The demands of automakers have also formed a consensus. Zu Sijie, vice president and chief engineer of SAIC Motor, said consumers "will definitely pay only for experience"; computing-power figures can be a technical benchmark, but if the user experience is not delivered, the technology is meaningless. In his view, cloud models can be very large, and cooling can be done at any cost; but in the car, models must be distilled and compressed, and space and energy consumption are limited, so "it must be model-chip co-design."

Gao Xinhua, executive vice president and CTO of Chery Automobile, admitted frankly, "We will not look at just one computing-power number." He believes whether a chip is good must be judged in a real vehicle system, with real models and real usage scenarios; to build chips, one must anticipate the requirements of the next-generation models and leave room for them. In his view, algorithm companies best understand how models will change next, and chip companies best understand the underlying architecture. "In the future, it will not be that everyone perfects their own thing and then shakes hands; rather, we must walk forward side by side from the very beginning."

Algorithm companies, chip companies, and automakers have come together on the basis of this concept, with chips defined from the outset around real models and vehicle requirements.

02

700 TOPS: How does XHEART improve effective computing power?

Regarding the newly launched large-model chip X7 Ultra, Li Zonglin asked the audience on stage to first remember one number: 700 TOPS, which is the computing power of the X7 Ultra chip.

He particularly emphasized the test conditions. The industry typically uses laboratory benchmarks to evaluate chips, but laboratory results often differ greatly from the actual performance of production models. XHEART uses real production intelligent-driving models, and requires that after model compression and deployment, the output similarity to the original model be above 99.99%, that is, "four nines."

At the roundtable forum, Turing Award winner Jack Dongarra discussed this issue from another angle. The LINPACK benchmark he developed later became the industry standard for measuring supercomputers, and he described himself as an "accidental benchmarker." He said this test suite was created 30 years ago, when computing itself was expensive; today, machines' computing power is long since in surplus, and what is truly scarce is moving data within chips. He believes AI chips will face the same problem.

Behind 700 TOPS, Li Zonglin introduced six architectural designs of the X7 Ultra. In the current high-end intelligent-driving chip market, 700 TOPS alone is no longer enough to constitute a competitive barrier. What XHEART focuses on more is how, while ensuring model accuracy, to reduce the overhead caused by data transmission and computation waiting, so that more computing power is truly used for model inference.

The first is a super-large-core NPU designed specifically for large models. Li Zonglin used an analogy: in the past, collaborators were scattered across different offices and materials had to be passed back and forth; the super-large core is like putting the team into one spacious workspace. Data interaction thus rises from the hundreds of GB/s level between small cores to the TB/s level inside the super-large core. The NPU also has high-precision dedicated acceleration engines for key operators such as LayerNorm, Softmax, and Transpose, making operator execution fast and accurate.

The second is unlocking hardware parallelism, enabling more computing units to work simultaneously, with computation and weight loading also proceeding in parallel. Based on actual testing with production intelligent-driving large models, NPU computing utilization reaches 98.7%.

The third is reconstructing the AI toolchain. XHEART proposes unifying the orchestration of the entire model's computation through network-level fusion, reducing repeated reading and writing of intermediate results to external memory. Compared with flagship industry products, memory bandwidth consumption is reduced by 80% and energy efficiency is improved by 100%.

The fourth is China's first mass-produced Armv9 automotive-grade CPU, which handles complex control and system scheduling and supports continuously evolving intelligent-driving systems.

The fifth is deep heterogeneous computing, with local tasks processed locally by dedicated computing units; LiDAR processing bandwidth reaches up to 20 million points per second, and sensor preprocessing stability is improved by 100%.

The sixth is security capability combining software and hardware. The X7 Ultra meets ISO 26262 ASIL-B functional safety requirements, and the built-in MCU is ASIL-D certified. Li Zonglin said, "Hardware safety plus driving safety is what truly constitutes intelligent-driving safety."

Together, the six capabilities increase the number of deployable model parameters by 10 times.

Dr. Chen Ping, Deputy General Manager of TSMC, said that advanced processes provide the computing foundation for large models by continuously improving performance, energy efficiency, and integration. TSMC has provided process support for the X7 Ultra and said it will continue to support XHEART innovation with advanced processes: "Our process is ready."

03

"Only one chance"

"We can only make one choice, and we have only one chance." Li Zonglin recalled the situation when the company was founded in 2023: the capital market was sluggish and resources were limited.

There were two paths before them: one was to make chips with a general architecture that could run no matter how models changed; this path was safe but performance would not be very good. The other was to customize around large models, which carried risk and uncertainty. They had extensive discussions with Cao Xudong and chose the latter path.

The working model changed accordingly. The industry's usual process is serial: design the chip, tape it out, get it back, then integrate it with algorithms, then adapt it to the customer's domain controller, and problems often only surface later. XHEART moved all these steps earlier, before tape-out.

Collaboration with Momenta began at the architecture definition stage. In 2023, the two sides discussed architecture and operator requirements around the continuously evolving mainline models, judging which computations would become increasingly important in the coming years. "The models tell us where computation time is spent," Li Zonglin said. The division of labor was also clear: Momenta provided the pain points of models on chips and future directions, while XHEART decided the chip's microarchitecture, bandwidth, operators, and toolchain itself.

English text is being prepared.

Customers were also brought in. The team of Tong Xiaofeng, Senior Director of Intelligent Driving and Chassis R&D at SAIC Volkswagen, maintained a high degree of transparency with XHEART; before the chip was taped out, the peripheral systems on Volkswagen's domain controller had already been integrated for interface verification.

These upfront efforts paid off at several milestones after the chip came back.

In less than 24 hours, the chip successfully lit up. At the launch event, laboratory footage of the first X7 lighting up was played: engineers confirmed current and voltage item by item, staring at the indicator lights, until someone shouted "It passed" and "Success." When the video ended and Li Zonglin mentioned this moment, he choked up for a moment.

On Day 17, the chip, low-level software, and autonomous driving model achieved a full-vehicle closed loop. But the team decided to redefine "chip-on-board" using standards closer to mass production: the trajectory planned at every intersection had to match mass-production standards. The system was tested on real roads in daytime, nighttime, and heavy rain. It was around this time that Tong Xiaofeng was invited to test the car. At that point XHEART had not yet officially opened its experience program to outside parties, and he was the first customer executive to get in the car. "At that time, you truly had everything ready except for this chip," Tong Xiaofeng said.

On Day 21, real-vehicle experience was opened to more customers to verify the actual performance of the chip and algorithm working together.

Cao Xudong summed up the reasons for the "speed." General-purpose chips are designed for all possible applications; the designers do not know where the chip will ultimately be used or what the algorithm should run like, so after users receive the chip, they often encounter various problems and mismatches. A capable and lucky team that spends 3 to 6 months getting the chip running, stable, correct, and performing well is already "thanking its lucky stars"; sometimes it takes 6 to 12 months or even longer. This is also why automakers are all "unwilling to be the first to eat crab" when choosing chips.

The key to X7 shortening this process is that a large amount of software-hardware integration and verification work had already been completed in advance.

From the company's founding to mass production, XHEART took two years.

Three years ago, Li Zonglin and Cao Xudong approached TSMC Vice President Chen Ping to explain a chip architecture based on large models, the vision for L3 and L4, and this plan. Chen Ping admitted: "If someone else had come to explain it, I would have had doubts."

Today, X7 has entered mass production. For XHEART, the architectural judgment made three years ago is beginning to face verification by real vehicle models, customer orders, and large-scale delivery.

04

From one car to forty vehicle models

The X7 series entered mass production in March this year, and its first customer was SAIC Volkswagen. Jia Jianxu, President of SAIC Group, recalled that when planning the first model, the ID. ERA 9X, he asked Cao Xudong whether he had a chip strategy. Cao Xudong told him the chip was "on the way" and then asked in return: Does Volkswagen dare to use it?

Jia Jianxu did not hesitate: "Let's just do it." In March this year, the ID. ERA 9X went on sale and became the first mass-production model to launch the X7 series. Jia Jianxu said that a car bearing the SAIC brand has "not only Chinese software, but also Chinese chips," and this is "a kind of pride from the bottom of my heart."

Since the first model, X7 has been installed in 10 mass-production models and has received orders for more than 40 models. At the launch event, these numbers were presented more concretely: outside the venue, 14 cars were lined up in sequence, including the all-new IM LS6, the all-new Cadillac XT5 PHEV, as well as the Chery Fengyun A9L, Fengyun T11, Roewe Jiayue 07, and the SAIC Volkswagen ID.ERA series and SAIC MG07. TSMC Vice President Chen Ping said that when he saw these cars, he could hardly believe it: "If you had told me three years ago, I would not have believed it no matter what."

Most of the display cars use the integrated software-hardware combination of the X7 series plus the Momenta R7 world model, but X7 does not serve only this one solution: the all-new IM LS6 is equipped with IM Motors' IM AD advanced driver assistance system. Li Zonglin emphasized that X7 is not tied to Momenta's algorithms; the chip is sold independently, and automakers decide whose chip and whose algorithm to use.

X7 is also entering a more mainstream price range. After the RMB 300,000-class ID. ERA 9X, the MG 07 and Roewe Jiayue 07 have brought this integrated software-hardware solution into the RMB 100,000 to 150,000 market. Roewe disclosed that pre-sale orders for the Jiayue 07 exceeded 50,000, and among test-drive users who had experienced advanced driver assistance, 77% placed orders on the spot.

X7 is also going overseas. Li Zonglin revealed that XHEART has obtained design wins from several international brand automakers. On October 22, Shenlong Technology also confirmed that it would adopt XHEART chips, and the two sides will cooperate on intelligent models for brands such as Peugeot and Jeep®, expanding into China, Europe, and other global markets.

Beyond chips, the models are also being upgraded. Cao Xudong announced that based on X7 Ultra, Momenta will soon release the R7 Ultra world model, with parameter scale and safety performance both increasing by 5 to 10 times, and it is beginning to show the ability to "simulate everything." In the on-site video, when meeting another vehicle on a narrow road, the car actively reversed to yield, passing through with only a few centimeters of space, and was also able to make multiple adjustments to complete a U-turn at a narrow intersection.

Cao Xudong also revealed that the same world model has been used in Robotaxi, Robovan, and Robotruck. Momenta has obtained test operation permits in Suzhou, Shanghai, Shenzhen, Abu Dhabi, Munich, and other places, and its joint venture with an Emirati sovereign fund plans to deploy 5,000 to 10,000 vehicles in 2028. He judged that Robotaxi will enter the scale-up stage next year, become the tipping point for L3 and L4 in 2028, and achieve large-scale deployment in 2030; by the end of 2028, leading companies' autonomous driving safety has a chance to reach 10 times that of human driving.

To support such models, the chips must also continue to advance. At the launch event, XHEART also announced the R&D goals for its next-generation chip X9: 3,000 TOPS large-model computing power and 700GB/s single-chip memory bandwidth, planned for launch in 2027. On Cao Xudong's presentation slides, between the R7 Ultra world model supported by X7 Ultra and the next-generation world model supported by X9, model parameters increase by 10 to 100 times.

Li Zonglin said the company will maintain the rhythm of "mass-producing one generation, developing one generation, and pre-researching one generation." "Autonomous driving is the starting point, robotics is the next stop, and physical AI across all scenarios is the distant place we want to go." But he also reminded the team to have "a strong sense of reverence" for chips.

As the launch event ended, a large screen displayed an "X" against a tech-blue background.