What are AI tops? Understand the concept of new AI PCs

In the last two years, major technology companies have started to present the concept of AI PCs (Brazil, IA PCs) to the public. These are more focused on artificial intelligence in the middle of productive IAS’s fever, and all performances are measured on the hills. So what exactly are these tops?

Abbreviation for a long time in information processing, the concept of the upper parts is extremely new for ordinary consumer. To better understand this abbreviation Tecmundo Intel’s application engineer Yuri Daglian spoke and we will help you understand the AI ​​world better.

NPU and IA

First of all, let’s start at the beginning. Before talking to the balls, we need to understand how AI works on the new AI PCs. These notebooks and computers are only called because They had the introduction of a part called NPU (Neural Processing Unit) to the processor.

The processor or CPU is the brain of computers responsible for performing, interpreting, processing data and displaying them by us. This CPU is divided into several sections, each functional, and NPU, Intel processors, Qualcomm, AMD and so on.

The wonderful idea behind this NPU was that users could direct work loads directly from the computer, so local. This ends the need for everything to be a process for Openai for Google or Chatgpt fans for Microsoft servers for Copilot users or for those who love Gemini.

“All NPU architecture is designed from start to nervous networks. It was made to offer great performance and consume less energy. A great maximum in the calculation [NPUs]Projects for specific purposes [IAs]It will always have more performance and energy efficiency than general processing units. [CPUs]Yur Yuri Daglian explains.

The neural processing unit is a small part of the chip specializing in performing workloads focusing on more performance, more performance, and less energy spending. For this reason, young laptops are called AI PCs and this unit is usually certified as Copilot+PC, since CPUs are in the unit.

However, the performance of these NPUs and new PCs is measured on tops. But what are these tops?

What are AI tops?

Tops is the abbreviation of trillions of operations per second, an unit of precautionary unit to refer to the performance of artificial intelligence inference. As Intel Expert points out, this abbreviation says that the superiors can perform a ridiculous amount of transactions in a certain period.

Simply put, the functioning of both computers and artificial intelligence is pure mathematics. Thus, the more the chip calculates in less time, the more performance to the user. Thus, theoretically, the more a processing unit is, the more performance the machine will have in AI.

How do the tops work?

Tops basically performs these mathematical procedures in a neural network. The nervous network is a method that uses machine education to teach the computer to perform certain tasks similar tasks, but is not equal to the human brain.

The most basic nerve network used is called perptron – and no, does not come from transformers. Perptron is the most basic nerve network available, but helps us to understand IAS’s functioning. Pay attention to the picture below:

In the drawing, XYZ is fixed numbers and the letter W is a certain weight assigned to them. The combination between numbers and weights produces an output with the answer to a question.

When we ask a life chatgpt when it is 1+1, these constant numbers are added or multiplied by weights. Each number, 0 + 0, 0.5 + 0, 0 + 1, etc. It will test various combinations of different weights such as. In the end, a fixed number will find that 0 + 1 is equal to 1, and that another number will make 0 + 1 equal to a total of 1 + 1 and find the answer equal to 2.

Thus, the neural networks are trained with a huge combination of number and weight to recur these processes until they reach the expected result.As explained by Yuri Daglian. “Operations of an NPU will ensure that the matrices are added and multiplying in very complex situations,” the operations of an NPU. ” Says.

Daglian, because they are much simpler for machines, “computers like to calculate matrices” jokes. PCs calculate these matrices on a “large” scale of billions of parameters to reach a final response. These are the processes on the upper part.

When we say that a neural network is trained, the answers of developers mean that they operate billions of different weights in an intervention and error system up to the best possible type.

Daglian, “Artificial intelligence pure math. Not mystical. Magic. There is pure math,” says Daglian.

How many inspections do I need?

Theoretically, the more the upper parts of the NPU are, the greater the ability to deal with some AI loads. But, There is no definite response to the ideal amounts and this is very dependent on the activity.

As Intel’s application engineer points out, 5 tops may be sufficient for Microsoft teams to create these virtual funds without accidents at meetings or use a smart filter to remove the noise from the microphone.

If the user wants to run an application based on a SLM (Small Language Model), something should be sufficient between 15 and 20. Now, operating a much heavier and more complex LLM (great language model) may require more than 100 tops depending on optimization.

Yuri Daglian, “Today, having an interesting experience, I can say that the NPU is at least 10 over to cool applications.

CPU and GPU tops

Yes, the tops are widely used when we talk about neural processing units, but in traditional processors there are tops and special or integrated video signs. For example, Intel Core Ultra Series 2 processors exemplify this well.

For notebooks, this Intel generation has a maximum of 120 top applications. Of this 120, only 48 upper NPUs, 5 of them are on the processor and 67 are on the integrated video card. CPUs and GPUs have always had these tops, ultimately we have used for years in Google Meet filters and other features.

Generally, video cards are the top hardware parts, especially special models. For example, a NVIDIA RTX 5090 has an incredible 3,300 AI upper part, the newly published Intel Arc B580 233 is a very important amount.

However, as Yuri Daglian explained, video cards have not been designed with loyalty to perform workloads such as an NPU. “GPU has an optimization for parallel calculations, so it has many nuclei that can work at the same time to make an account. CPU is more sequential, so it makes an account first, ends, makes someone else, etc., etc.” Says the expert.

Even though video cards and processors do this job, Yuri ends by explaining that NPU is always more efficient because it consumes less energy while performing these tasks.

Did you like to learn IAS’s tops and functioning? Then continue on the Tecmundo website. If you like hardware and computer science, don’t forget to take a look at our GeForce RTX 5080 Founders Edition.