With artificial intelligence (AI) emerging as a hot topic and causing much debate these days, ChatGPT, which implements artificial intelligence, is receiving great attention and praise.

The GPT (generative pre-trained transformer) applied in ChatGPT is a type of large language model (LLM) that requires a huge amount of computing power.

As a leading supplier of computing and other semiconductor technologies related to AI, Infineon’s perspective will be helpful to many designers who want to become familiar with AI and leverage it.

This article is intended to provide background information and insights on AI, explaining the timeline, risks, current status, and future prospects for AI.

■ Background on ChatGPT and LLM

Machine learning (ML) initially found success in applications specific to a particular domain.

In 1997, the Deep Blue supercomputer beat then-world chess champion Kasparov.

In 2010, IBM Watson, a question-answering computer system, won the quiz show Jeopardy!

Following the emergence of Siri and Alexa, natural language processing (NLP) models began to gain popularity.In 2015, deep learning models were shown to be able to outperform humans at image classification tasks, such as diagnosing skin cancer and interpreting MRI (magnetic resonance imaging) scans.

More recently, these models are being applied to autonomous driving programs in self-driving cars.

After developing all these specific products, researchers turned to integration, wanting to develop one single model that could solve all problems for everyone.

To do this, these researchers developed a device that learns from all information on the web, including Wikipedia and books.

These are called giant language models (LLMs), and the best known LLM is OpenAI's ChatGPT (3.0 and 3.5).

According to experts, LLMs will soon surpass human intelligence.

There are currently more than 30 LLMs available, including Google's PaLM, Meta AI's LLaMa, and OpenAI's GPT-4.

You can ask ChatGPT any question you want, for example, ask it to plan a trip to Paris.

More specifically, you can ask, “Recommend five tourist attractions in Paris.” This question can easily be changed to recommend ten places. Just say, “Recommend ten places.” Then, instead of repeating the previous answer, you simply add to it, just like when we have a normal conversation. ChatGPT summarizes the results at a glance, so you can use it to plan your trip.

The first answer suggested a 5-day trip, and when the user adds a date, it suggests an itinerary accordingly. New users of ChatGPT will be amazed at how easy it is to interact with it and how great the responses are. It feels like you are talking to a friend or colleague.

■ Historical background and rapid growth

The timing of LLM is based on the availability of extensive computing power. As shown in Figure 1, the floating point operations per second (FLOPS) required to train transformers has increased by 105x since 2017.

Model sizes have increased 1,600-fold in just the past few years.

It takes 32 hours to get an answer or inference with GPT-3 on a desktop CPU (central processing unit), but it takes only 1 second on an Nvidia A100-class GPU (graphics processing unit).

Also, it would take 30 years to train GPT-3 on a single GPU like this.

Therefore, training is typically performed using a large cluster of 1,000 to 4,000 GPUs, takes from two weeks to as little as four days, and the training cost ranges from $500,000 to $4.6 million for a single training iteration.

The entities that have access to these massive computing resources have been very limited, and their popularization in this regard is relatively recent.


The importance of LLM can be seen in the speed at which ChatGPT is being adopted. As shown in Figure 2, it took Instagram and Spotify 75 and 150 days, respectively, to reach 1 million users, while ChatGPT reached this point in just 5 days. Compared to Facebook, which took 4.5 years to reach 100 million users, ChatGPT reached this point in just two months. This is an unprecedented growth rate in the technology industry.


■ How ChatGPT works

ChatGPT can be viewed as a probabilistic sentence generator, that is, a generator structured to guess the most common sentences.

It guesses the most likely next word in a sentence, and the context can change based on the amount of information provided.

Based on what you said above, answer (guess) the following word.

For example, GPT3 looks at the previous 2,000 words (about 4 pages) to decide about the next word.

In comparison, GPT3.5 reviews about 8 pages, and GPT4 reviews about 64 pages. This greatly improves its ability to guess the next word. A brief summary of the LLM process is as follows:

▶Step 1: Learn a basic model that contains detailed information on various topics.
▶Step 2: The LLM owner or someone else provides a pre-prompt. This defines the ‘personality’ of the chatbot (friendly, helpful…) and defines the harmful domains to avoid.
▶Step 3: Provide additional input or value from a specific perspective through detailed adjustments.
▶Step 4: Filter out some answers through RLHF (Reinforcement Learning with Human Feedback).

To provide answers, tools like ChatGPT capture what the user types and, in many cases, call an application programming interface (API) to quickly respond. The API usually accesses information from another server and provides the answer, usually through a licensing fee. This process can be repeated additional times.

LLM tools do not deserve to be trusted unconditionally by users, as there are risks associated with using LLM tools.

These tools can provide very well-packaged answers to many questions, but those answers can be wrong, and sometimes very wrong. Because users cannot tell which filters have been applied and whether their answers have been distorted or even misleading.

■ Possess system expertise for AI

Infineon, a leading supplier of semiconductor system solutions, provides solutions for safer and more efficient vehicles, more efficient and environmentally friendly energy conversion, connected and secure IoT systems, power management, sensing and data transmission.

For example, the Infineon power system is part of a DC-DC solution that can supply 1,000 amps to the GPU.

It also provides various system-on-chip (SoC) solutions for memory, advanced automotive, and other applications.

With its recent acquisition of Imagimob, a leading platform provider of machine learning solutions for edge devices, Infineon strengthens its position in the fast-growing TinyML (Tiny Machine Learning) and AutoML (Automated Machine Learning) markets, offering a comprehensive development platform for machine learning on edge devices.

Clients can leverage this experience across a range of applications, including LLM.

■ AI Outlook

Although there may be some controversy, artificial intelligence is already here and its use is likely to continue to increase in the future.

As designers become more familiar with AI and want to leverage it, they need solid system solution companies like Infineon.

As a market leader in a variety of fields including AI/ML, Infineon enables customers to bring their products to market quickly through a broad portfolio of advanced sensors and IoT solutions. Do it.


※ author
Bernard Berg, Director AI and Data Science, Infineon Technologies