GPU and NPU IP is a ready-made design for a graphics engine or an AI engine, which chip companies license and build into their own chips. A GPU, or graphics processing unit, is the block of a chip that draws images on a screen. It has many small cores that do the same arithmetic on many pixels at once. An NPU, or neural processing unit, is a block built to run AI models, which need huge numbers of multiplications and additions. IP stands for intellectual property: the chip company buys the right to use the design, not a physical part.
Why license instead of design? A graphics engine is hard to build well. It also needs a large body of software. It needs a driver, the program that lets the operating system use it, and support for the standard graphics libraries, such as OpenGL ES and Vulkan, that app and game makers write their code for. A chip company that licenses the engine receives that software too. Building both from scratch would take years.
Consider a design team in Pune building a chip for a car's digital dashboard and its centre screen. Such chips are automotive processors. The dashboard screen has 1,920 × 720 pixels, and it must be redrawn 60 times a second. That is about 8.3 crore pixels every second, and the GPU colours every one. The speedometer must never freeze, even if the music app on the centre screen crashes. So the GPU must keep the dashboard's work separate from the other screen's work. The chip also has a camera that watches the driver's eyes for signs of sleep. The NPU runs that AI model about 30 times a second.
The team licenses both engines. It receives them as design files, which it loads into its electronic design automation (EDA) software, together with the drivers. It connects the engines to the chip's processor and memory. It pays a fee at the start. Later it pays a royalty, a small amount for every chip shipped. The concept page on semiconductor IP licensing explains these two payments.
The same kinds of engine go into phone chips. In a system-on-chip (SoC), the GPU draws the home screen and games. The NPU runs features such as photo enhancement and voice typing on the phone itself, without sending the data to a server.
A newer approach builds these engines on RISC-V, the open instruction set that anyone may use without paying a fee. An instruction set is the fixed list of basic commands that a processor understands. A graphics or AI engine built on RISC-V can be programmed with the same open tools that chip companies already use for their RISC-V processor cores.
How is this different from processor IP licensing? A processor core runs general programs, one instruction after another. A GPU or NPU does one kind of arithmetic on thousands of numbers at the same time. Most phone and car chips license both kinds of block.