Intel has let slip the existence of a new type of processor – the “versatile processing unit” (VPU) – built to accelerate AI inference workloads.
Although no formal announcement has been made, Intel recently published a new Linux driver complete with written materials that refer explicitly to the specialist chip.
According to the documentation, the VPU will feature inside Intel’s 14th Gen Core CPUs (also known as Meteor Lake) and will improve inference performance across “computer vision and deep learning applications”.
TechRadar Pro has asked Intel for comment.
With rivals like Nvidia fighting to establish themselves as the leading chip maker of the AI era, Intel will be thinking hard about how to demonstrate its own credentials in the space.
Broadly speaking, there are two types of AI workload: training and inference. The former refers to the use of large-scale datasets to develop AI applications with specific capabilities, while the latter refers to the feeding of new data into these systems to generate a result.
As a product of the amount of data involved, training is both highly compute-intensive and highly time-consuming. By contrast, inference needs to be executed almost instantaneously, which means the compute requirements differ significantly.
Although details are obviously scant, Intel’s new VPU appears to be designed to accelerate interference workloads across client devices. It’s unclear whether the company’s line of server CPUs will benefit from similar technology in future.
With Meteor Lake expected to land in late 2023, the new VPU won’t see the light of day any time soon, but the existence of the chip offers some additional insight into the broader strategy at Intel.
With some of the world’s largest companies (including the likes of Apple) finding that custom silicon designed specifically for in-house configurations can provide marked performance improvements over general-purpose chips from Intel or AMD, the chipmakers need to find a way to remain competitive.
One option would be to broaden the portfolio with specialist SKUs packing additional accelerators and co-processors designed to boost performance across target workloads.