Unveiling its volume-built autonomous-driving processor in Suzhou on October 29, XHEART turned the event into an argument about how the chip industry reports performance. The X7 Ultra’s headline claim is 700 TOPS of large-model compute, and the company insists that number was obtained without loosening precision: the deployed, quantized model must reproduce the trained FP32 model at better than 99.99 percent similarity.
Chief executive Li Zonglin told the audience the figure came from testing against vehicle models that already ship, not from a lab benchmark. Alongside it, XHEART cites neural-processing-unit utilization of 98.7 percent, memory bandwidth use 80 percent below rival flagship parts it does not name, and capacity for a model with ten times as many parameters.
Deliveries tell a firmer story. Eight months of volume production have produced over 200,000 units, fitted to 10 vehicle models, with orders booked for more than 40. A two-chip compute platform aimed at higher automation has taken design wins across several programs.
Funding so far totals about RMB 5 billion, near $750 million. Renault Group came in as a new investor, Chery Automobile increased its holding, and DRAM maker CXMT acquired a stake. A successor rated at 3,000 TOPS, the X9, is slated for 2027.
Whether the methodology survives outside scrutiny may matter less than the shift it marks. Chinese automotive silicon has begun competing on how much performance a shipped model actually extracts, the quantity carmakers ultimately buy, instead of the peak rating in a specification sheet.
