Momenta now has a chip designed around its own driving models, so carmakers adopting R7 Ultra can source software and silicon as one tested package.
XHEART, a Suzhou chip designer and strategic partner of autonomous-driving company Momenta, launched the X7 Ultra, an automotive processor it says delivers 700 TOPS of computing power for large AI models. TOPS, trillions of operations per second, is the standard measure of an AI chip’s raw throughput. At the launch event on October 29, Momenta said it would release a new driving model, R7 Ultra, built to run on the chip, IT Home reported from the event.
The pairing is the point of the product. XHEART’s pitch, summed up at the event as "大模型定义芯片" (large models define the chip), is to design the processor around the driving models that will run on it, rather than build a general-purpose chip and fit models to it later. Momenta chief executive Cao Xudong attended the launch.
The performance figures are company claims, and the two companies did not publish independent test results.
XHEART claims high utilization and lower memory traffic
XHEART says the X7 Ultra reaches 98.7% NPU utilization in its own measurements and uses 80% less memory bandwidth than a comparable flagship chip. The NPU, or neural processing unit, is the part of the chip that runs AI models; utilization measures how much of its rated capacity a real model actually uses. The company also claims double the energy efficiency of comparable flagship chips and support for models with 10 times as many parameters as the previous generation, according to MyDrivers.
The chip uses a single very large NPU core with hardware acceleration for transformer operations, the building blocks of most current AI models. Its central processor uses an Armv9 core that MyDrivers describes as the first mass-produced automotive-grade Armv9 core in China. It includes a safety microcontroller rated ASIL-D, the highest automotive safety level under the ISO 26262 standard, and supports 10-gigabit Ethernet, PCIe 4.0, lidar and high-resolution cameras.
Momenta’s R7 Ultra is designed to need the extra capacity
Momenta said R7 Ultra has five times the parameters of the original R7 world model and five times its safety performance, and that it plans to reach 10 times the parameters and more than 10 times the safety performance, IT Home reported. A world model is an AI system that simulates how the physical world behaves, which a driving system can use to predict what other road users will do. Momenta says the new model can match or exceed experienced human drivers in complex situations such as multi-point turns, negotiating with oncoming traffic, yielding while reversing and extreme parking.
R7 Ultra keeps the three-layer structure of R7: pretraining on driving data, simulated physics to generate rare and extreme scenarios, and reinforcement learning to refine behavior, according to MyDrivers. Momenta draws on more than 100 million kilometers of data returned from production vehicles and uses an automated toolchain, which it calls CLA, to pick the most useful samples for further training. The model runs on the vehicle and is meant to estimate the physical properties of objects and predict how pedestrians, cyclists and other cars will interact. Momenta says it performs best in mixed traffic on narrow streets, around construction that blocks the road and in complex reversing; a demonstration at the launch showed the car reversing around obstacles in a crowded space.
Earlier this month XHEART completed a strategic funding round in which Renault, Chery and memory maker CXMT took part, as EastFrontier reported. In August, Momenta, XHEART and QNX announced a production-grade driving platform built on the first-generation X7.
A named car program will be the real test
XHEART’s advantage is its fit with Momenta’s software. That fit is also a dependency: the X7 Ultra’s value rises or falls with the number of carmakers that adopt Momenta’s driving system. The next signal will be a carmaker naming a model that uses the X7 Ultra with R7 Ultra, and a start date for production. Independent measurements of the chip’s utilization and power use would test whether the efficiency claims hold outside XHEART’s own benchmarks.