NPU is nothing new, in fact this component It started to be used years ago in the mobile sector. Apple was the first company to popularize this element under the name Neural Engine, which literally means neural engine. It debuted in 2017 and was integrated into the Apple A11 SoC, the chip that gave life to the iPhone 8, iPhone 8 Plus and iPhone X.

When we do the math, we’ll see It’s been more than six years since this component began to become popular, although its arrival in the PC sector is something relatively new. AMD was the first to move in this regard with the Ryzen 7040 APU, which, as many of our readers know, has a first-generation NPU.

A refresh of said generation was recently confirmed with the Ryzen 8000 APUs, which use the same CPU architecture and differ mainly in that much more powerful NPU, although the real development in this regard will come with the next-generation solution known as XDNA 2, which is scheduled for release in 2024.

Knowing how to see the potential of the NPU and the pressure of AI, Intel also decided to bet on the use of this specialized hardware with new processors Intel Core Ultrawhich fall under the Intel Meteor Lake generation and have already started making their way into a variety of mainstream consumer notebooks of various sizes, including ultralights like the ASUS Zenbook OLED 14 to 16-inch models like the MSI Prestige series.

What is NPU and what does it do?

There’s no doubt that NPU is becoming popular in the PC sector and that it’s the big guys’ answer to the rise of AI, but what exactly is it? Short for neural processing unit and as the name suggests, it is a chip or processor. specializing in neural network operationssuch as inference, deep learning and artificial intelligence.

This unit can work with different algorithms and offers tremendous versatility in everything related to AI workloads, but cannot be used for general computing tasks, which means it must be accompanied by a CPU in those computers intended for the general consumer market. This explains why both Intel and AMD, as well as Apple, Samsung and all the big players in the sector They implemented the NPU together with the CPU and GPU.

NPU works with operations specific to artificial intelligence optimized to offer an excellent level of performance at this level while maintaining low power consumption. In terms of performance, an NPU integrated into an Intel Core Ultra processor or Ryzen 8000 APU will not offer more performance than a dedicated GPU, but its power-to-consumption ratio will be so good that will be exactly where he resides its main value.

This specialized AI processor uses a a parallel computing architecture that is data-driven, which makes it an excellent option for massive processing of various data, including multimedia content such as videos or images. It can work with multiple threads in parallel, uses a readily available caching system, and its processor cores have a simplified design because, as I said before, they are not intended for general computing.

Right there is one of the most important keys to a good relationship between performance and consumption. As their cores are specialized and optimized for AI, they offer superior performance and have a simpler design that requires less power to function properly. They are optimized to work with them algorithms with low precision (INT8 and INT4), and its performance level is normally measured in TOP, short for trillion (Anglo-Saxon) operations per second.

The NPU includes multiplication and addition blocks, activation functions, and can perform 2D data and decompression operations, among others. This allows you to perform addition and subtraction and multiplication operations applied to matrices as well can work with convolutional networks and scalar operation.

Why is NPU important?

As I explained earlier, the NPU can work with deep learning algorithms and neural networks, and those in turn they can be trained to offer different types of solutions characteristic of artificial intelligence, the complexity and breadth of which can vary depending on the level of training they have received.

I can give you many examples, think about the neural network it is used to choose the best frames in NVIDIA’s DLSS upscaling and reconstruction technology or in technology that uses Tesla’s autopilot to identify in real time the different elements of the streetincluding everything from pedestrians and cars to road signs, location and distance.

What NPU does is therefore use models already trained neural networks use them through the process derivation. The acceleration it provides thanks to its specialized hardware allows it to offer optimal performance, although its greatest strength is not in raw power, but as I said in the value it offers in relation to consumption-performance, because in this sense it surpasses both. CPU as GPU and also has much lower cost.

With the arrival of NPU in the PC world access to AI is democratized, because this component will make it possible to take advantage of advanced AI features that really deliver significant value without making major sacrifices in terms of cost or efficiency. Its evolution will be constant in the coming years and we can expect an important evolution in terms of power.

Cover image: Freepik.