Buzz
With leaked details from the Linux operating system driver, Intel has silently affirmed the presence of VPU in the 14th-generation processor - Meteor Lake. VPU, also known as Versatile Processing Unit, is an ASIC (Application-Specific Integrated Circuit) - a specialized integrated circuit designed to accelerate AI inference for deep learning applications, akin to NVIDIA's Tensor Core.There are two types of AI tasks: training and inference. During AI training, large datasets are utilized to develop AI with specific features, essentially teaching AI to master and autonomously perform a specific task. In contrast, AI inference applies trained models to unlabeled data for computation and result generation. AI inference is a critical component in all current AI application workflows. As the technology sector increasingly focuses on artificial intelligence algorithms, Intel's integration of VPU into processing units not only meets user demands but also supports developers.
Intel's VPU stems from the acquisition of Movidius - a chip manufacturer specializing in system integration and vision processing - six years ago. The VPU developed by Movidius at that time was groundbreaking, delivering remarkable performance per watt. Movidius' specialized architecture boasts up to 4 TOPS with just 1.5 W power consumption - an energy efficiency level that other companies in the industry can only dream of, including NVIDIA. After six years of continuous development, it remains uncertain what surprises the VPU in Meteor Lake will bring. With two generations of processors before the release of Meteor Lake, and initial information suggesting that VPU only existed in CPUs for end users, Intel's potential integration into server processors remains a mystery.Intel's VPU will include a memory management unit (used for translating VPU data and transferring it to DMA addresses), a RISC-based microcontroller, a Neural Compute Subsystem for auxiliary neural computations, and a Network on Chip. The current VPU core has a driver containing 8500 lines of code, while the firmware is closed-source.
