Author | Song Wen Editor | Su Peng
In the past two years, the parameter sheets for automotive chips seem to have been hit by an acceleration key.
Tens of TOPS, hundreds of TOPS, thousands of TOPS—the PPT slides at launch events get flashier with each page, and the numbers grow larger with each round. When automakers release new vehicles, chip compute power has become a mandatory metric, and "total compute" has also become the default hard indicator in media reviews.
However, a counter-intuitive question is beginning to emerge: Does higher compute power necessarily mean better driving performance?
The answer is not necessarily yes. TOPS represents theoretical peak performance, akin to the maximum horsepower listed on an engine nameplate.
Whether this potential can be realized depends on whether data can be fed to the computing units in time. Large models are characterized by heavy computation and even heavier data movement.
Parameters must be read from memory, intermediate results written back, and then read again. Often, the chip is not calculating but waiting. The industry calls this the "memory wall."
On October 29 in Suzhou, a chip company named XHEART launched its next-generation large model chip, the X7 Ultra, with a nominal 700 TOPS.
This number does not stand out today, as some in the industry have already pushed compute power to 2000 TOPS. However, what this company deliberately emphasizes is not peak performance, but another concept: compute utilization. Based on real-world testing with mass-produced intelligent driving large models, the NPU utilization of the X7 Ultra reached 98.7%.
The signal it aims to convey is clear: The era of stacking numbers is passing. Whether compute power can be truly utilized is the decisive factor in the next round of competition.
Cars Cannot House Data Centers: Compute Power Involution Hits the Physical Ceiling
Do not rush to dismiss compute growth.
End-to-end large models replacing multiple small specialized models is an architectural shift currently occurring in the intelligent driving industry. From perception to decision-making, the system increasingly operates like a whole entity "driving." Momenta CEO Cao Xudong judges that achieving scalable L4 requires models with at least tens of billions or even hundreds of billions of parameters. There is no suspense about this direction; models will indeed get larger.
The real constraint lies on the vehicle side. A car is not a data center; it cannot infinitely stack power or heat dissipation capabilities. A vehicle must leave space for the battery, cockpit, and sensors, while controlling cost, energy consumption, and thermal management.
The "impossible triangle" often cited in the industry—performance, power consumption, and automotive-grade reliability—essentially draws three red lines against "infinite compute stacking."
This means that the chip industry's past path of "process node upgrades + core stacking" will hit a wall first in automotive applications. No matter how high the paper compute power of a chip is, if it runs unstably or costs too much to cool once installed in a car, the numbers on the spec sheet remain just numbers. This is why many automakers have been releasing high-compute platforms while simultaneously "locking down compute" in mass-produced cars over the past two years—there is a gap between claimed peaks and actually released capabilities, separated by the gates of thermal management and cost.
A more subtle issue lies in the architecture.
Traditional intelligent driving chips mostly feature multi-small-core NPU designs optimized for the era of small models, with inter-core data exchange bandwidth in the range of hundreds of GB/s. As models grow larger, data movement becomes the bottleneck, and the chip spends a significant amount of time "idling" internally. Stacking more compute merely scales up the waiting period.
Turing Award winner Jack Dongarra raised a point at the launch event: Thirty years ago, floating-point operations were the most expensive resource; today, floating-point compute is relatively abundant, and data movement is becoming an increasingly critical resource.
This judgment holds especially true for automotive applications. Peak compute is important, but what matters more is whether this compute can be truly utilized under constraints of power consumption, cost, and response time.
Software Companies Entering Hardware Definition: The Auto Industry Leads the Way
An industrial signal worth noting is that software companies are diving deep into hardware.
Last October, OpenAI announced a partnership with Broadcom to develop custom AI accelerators, with OpenAI responsible for system architecture design. Anthropic has partnered with Amazon's chip team Annapurna Labs, participating in the development of subsequent generations of Trainium accelerators.
Large model companies are no longer satisfied with adapting to general-purpose chips; they want to embed their understanding of models directly into hardware.
The automotive industry actually encountered this problem earlier, with stricter constraints. Data centers can add servers and cooling systems; cars cannot.
In recent years, domestic intelligent driving chips were mostly designed by automakers or chip companies following a "general-purpose route," adapting to models only after they arrived. The pain points of this model are obvious: Chip design cycles take years, while major versions of intelligent driving models update monthly. By the time the chip is ready, the model has already iterated several times.
XHEART takes a different path.
Founded in 2023, the company originated from Momenta's chip design project. Its CEO, Li Zonglin, was formerly an executive at OPPO's chip subsidiary, Zeku. From its inception, it had direct access to Momenta's mass-produced models and the algorithm team's judgments on next-generation models.
By the end of 2022, both parties reached a consensus: First determine what the model needs, then reverse-engineer how to build the chip, rather than building a general-purpose chip first and forcing the model to fit.
Applied to the X7 Ultra, the traces of this approach are specific.
As models move from multiple small models to end-to-end architectures, computation becomes increasingly centralized. The X7 Ultra adopts ultra-large-core NPUs, consolidating data interactions that originally occurred between small cores within large cores, raising interconnect bandwidth from the hundred GB/s level to the TB/s level;
After Transformer became the mainstream architecture, dedicated hardware acceleration units were added for the most time-consuming calculations within it. The toolchain was also restructured so that intermediate results do not need to be written back to external memory and read out at every step.
This practice of "models defining chips" essentially completes algorithm adaptation and functional verification before tape-out.
The effect is reflected in the mass production rhythm: The X7 lit up less than 24 hours after silicon return, achieved full vehicle closed-loop in 17 days, and Momenta ran through the entire process in 21 days. In contrast, the adaptation cycle for general-purpose chips in the industry is typically 3 to 6 months, sometimes exceeding one year.
Automakers Willing to Wait for a Startup: Calculating a Different Kind of Ledger
The business logic of the chip industry ultimately depends on whether automakers are willing to pay.
This is the most worthy aspect of XHEART's report card to dissect: Three years since founding, eight months of mass production, cumulative shipments of the X7 series exceeding 200,000 units, deployed in 10 mass-produced models, and orders secured for over 40 models.
Vehicles such as the SAIC Volkswagen ID.ERA 9X, Roewe Jiayue 07, MG 07, and Cadillac XT5 PHEV have been launched successively. Stellantis (Dongfeng Peugeot Citroën) also announced that XHEART chips would be adopted for intelligent models targeting Peugeot and Jeep brands.
Why do automakers dare to use chips from a startup? The example of SAIC Volkswagen explains part of it.
SAIC Group President Jia Jianxu recalled that when the two sides initially discussed cooperation, the focus was still on Momenta's intelligent driving software. When talking about chips, Cao Xudong told him that chips were already being prepared and asked back whether Volkswagen dared to use them. Jia Jianxu's answer was "Let's do it."
Before the chips returned from fabrication, both sides began designing domain controllers, performing early system integration of the chip with other components.
The key to this calculation is not "daring or not," but timing. XHEART entered the market exactly when automakers were generally seeking cost-effective solutions for high-level intelligent driving platforms.
Top-tier chips continue to push compute limits, with prices and power consumption rising accordingly, while mid-to-low-priced models are extremely sensitive to cost. A chip that can run large models, maintain controllable power consumption, and avoid half-year adaptation delays is as attractive to automakers as pushing paper compute power to the next order of magnitude.
Another layer is the binding of the algorithm ecosystem.
Currently, besides running Momenta's models, some automakers have already deployed their proprietary algorithms on this chip.
XHEART and Momenta collaborate deeply on projects but sell independently commercially, allowing automakers to purchase chips and algorithms separately. This gives automakers a flexible option: They don't have to hand over their "soul" to a single chip company, yet still benefit from a hardware foundation optimized for models.
Cao Xudong offered a judgment: Chips defined by large models can improve performance by 10 times compared to general-purpose chips.
If this multiplier holds, it means that for automakers, the same power budget can run larger models; for the same model, there is lower heat generation and more stable operation. Amidst converging intelligent driving functions and increasing cost pressures, this calculation is becoming increasingly attractive.
Betting on Model Evolution: Risks and Rewards Coexist
However, this route is not without cost. The biggest risk is betting on the wrong direction of model evolution.
Once a chip is manufactured, it is difficult to modify like software; it takes two to three years, or even three to four years, to produce a new one. Meanwhile, major versions of intelligent driving models iterate monthly. The industry consensus is that chip design must foresee the direction of model evolution in advance; otherwise, the chip becomes obsolete upon release.
Cao Xudong himself admitted that if future market judgments are wrong, even without technical issues, the chips might not sell.
XHEART's bet is placed on the premise that "models will get larger and more concentrated, and Transformer remains the primary architecture."
The X9 is already on the way, planned for launch in 2027, featuring 3000 TOPS and 700GB/s memory bandwidth, targeting larger world models. This choice of parameters itself reflects a judgment: Not just boosting compute alone, but synchronously improving compute and memory access bandwidth—a logic consistent with the X7 Ultra's principle of "calculating fast while feeding adequately."
Competition is also intensifying. Players like NVIDIA and Horizon Robotics have barriers in compute and ecosystems, and the ranks of automakers developing their own chips are expanding. XHEART's advantage lies in its deep bond with Momenta, whose intelligent driving solutions are already equipped in over 120 models, with additional design wins for over240 models. This provides XHEART with a ready-made customer entry point and ensures continuous access to first-hand information on model evolution.
Conversely, dependence on a single algorithm partner means that if Momenta's route judgment deviates, chip sales will face direct pressure.
The reliability of automotive-grade chips still requires time to accumulate. Consumer electronics chips can be validated within two to three years, whereas automotive-grade chips need to log sufficient mileage and data in real road conditions and extreme environments. Shipping 200,000 units is just the starting point; there is still a long way to go before being "fully validated by the market."
The Second Half: Competing on Compute That Is Actually Used
Returning to the initial question. In the first half of intelligent driving, the competition was about who could implement features first and who had broader coverage;
In the second half, the competition shifts to a different set of capabilities: For the same model, whose power consumption is lower and cost more controllable; for the same compute, whose actual utilization rate is higher; for the same iteration cycle, whose software-hardware synergy is tighter, enabling chips to keep pace with the annual evolution of models.
Compute power is never enough. But in the era of large models, what is truly scarce is not TOPS, but TOPS that can be practically utilized by models. As intelligent driving moves from assisted driving to L3 and L4, users will not perceive numbers on a spec sheet, but rather a smooth lane change on a rainy night or a composed passage through a complex intersection.
The phrase "let large models define chips" is less a technological route than a reminder returning to essentials: The value of a chip lies not in how big it is, but in how far it allows the model to go.