SenseTimeCo-founder and chief scientist Wang Xiaogang announced on the 17th,Shang Tang Jue YingThe first in the industry to achieve nativeMultimodal large modelThe vehicle-side 8B model has a first packet delay of less than 300 milliseconds, an inference speed of 40 Tokens/second, and covers mainstream computing platforms.
SenseTime has created a computing engine called "HyperPPL" for multimodal large models. It currently expands and supports mainstream in-vehicle computing hardware, is compatible with a variety of mainstream operating systems, and adapts to the deployment platforms of multiple in-vehicle chips.
SenseTime Jueying said that HyperPPL is optimized for multi-person scenarios in vehicles, so that when there are multiple people in the vehicle at the same time, the model reasoning efficiency of the multi-modal large model on the vehicle side is not significantly reduced compared to a single person.
SenseTime previously stated that Shenzhen’s first autonomous driving bus line uses its vehicles and technology, and all driving operations do not require human intervention.
Next year, automotive chips (NVIDIA Thor) with a computing power of over 1,000 TOPS will be available. Based on a computing platform with higher computing power, SenseTime expects that the first packet delay of Jueying's multi-modal large model vehicle-side deployment solution will be greatly reduced, and the inference speed will be further improved.