The automotive chip company XHEART, established as part of the Momenta chip program, unveiled its X7 Ultra chip on October 29. The company positions this chip as a solution specifically designed to work with large control models, estimating its computational power for such tasks at 700 TOPS.
The head of XHEART noted that the development began with the observation that driver assistance software is shifting from numerous small models to a single, comprehensive large model, which requires corresponding chip adaptation. When the company was founded in 2023, the decision was made to work in reverse, starting from the models actually used in vehicles, and organizing computation, memory, and data movement around them, leaving room for future models.
The X7 Ultra is equipped with a very large NPU core. XHEART claims that data transfer within a single core occurs orders of magnitude faster, reaching terabytes per second, compared to data transfer between cores. A multi-level cache, designed for high throughput and parallelism, aims to reduce inference latency and power consumption. A specialized operator engine accelerates key Transformer operations, such as LayerNorm, Softmax, and Transpose, and also supports FP16 and FP32 vector mathematics to maintain model accuracy.
According to the company's data, production models running on the chip meet benchmark results with 99.99 percent accuracy, and compute utilization reaches 98.7 percent. Compared to products the company calls industry flagships, XHEART states an 80 percent reduction in memory bandwidth consumption and a doubling of energy efficiency. These figures are presented by the company itself and have not undergone independent testing.
The chip also includes, according to XHEART, China's first mass-produced automotive Armv9 processor, which demonstrates 20 percent higher performance and 10 percent better energy efficiency. The architecture with a large cluster improves the coordination of multi-core processors and access to L3 memory, while the new SVE2 vector engine provides higher signal processing performance compared to the previous generation NEON.
XHEART also demonstrated its next chip. The X9 model targets 3000 TOPS for large model computations and 700 GB/s memory bandwidth per chip, with its release planned for 2027. It will be developed jointly with next-generation world-class models, synchronizing computation and memory access for efficient operation of larger models.
The company reports being in a cycle of producing one generation, developing the next, and researching a third. Two years have passed from its founding until mass production. The first X7 chip was operational within 24 hours of returning from the factory, and on the 17th day, the chip, low-level software, and control model were working together in a full vehicle.