At the dawn of an IT revolution: From instruction-based to intention-based computing

NVIDIA’s core competence lies in accelerated computing. We develop specialized computing platforms to drastically accelerate computationally intensive tasks and make them more efficient. Data centers around the world are switching to accelerated computing, as most modern applications are created in this environment. We are at the beginning of an IT revolution, the likes of which we have not seen in the last 60 years: the transition from instruction-based to intention-based computing.

To drive this transformation, we have spent the last 25 years developing our own computing platform with graphics processing units (GPUs), an ARM-based central processing unit (CPU) and a data processing unit (DPU/network accelerator). We have developed our DGX systems and DGX SuperPOD clusters – supercomputers for high-performance computing and artificial intelligence – on this technological basis. Together with device manufacturers and hyperscalers, we offer vendor-specific solutions to ensure maximum flexibility for our customers.

But hardware alone is not enough. In order to operate artificial intelligence efficiently and cost-effectively, a comprehensive software solution is required. That is why we have developed the NVIDIA AI enterprise software stack, which includes everything required for the development and high-performance operation of AI. The stack is independent of the deployment – so it doesn’t matter whether the solution is operated on-prem, in the cloud or in a hybrid infrastructure – and supports all the relevant major open-source language models, for which we also offer licensing support. Today, NVIDIA technology is integrated into almost every application that uses artificial intelligence.
 

The rapid development of AI in recent months came as no surprise to us – but rather, to our customers. Years ago, we were already talking about generative AI and language models as the next big topic, although many companies did not yet have it on their radar. This was because we had already been supporting the development of generative AI in research for several years. With more than 25 years of experience in accelerated computing, during this time we developed the concept of an “AI factory”. The efficient operation of AI requires not just a single chip, but a holistic full-stack approach. This includes GPUs, CPUs, network processors and the corresponding software. All these components must be integrated into a server that is optimized for AI. This also includes industry-specific software – for example, for the pharmaceutical or financial industry – which facilitates the development of end applications.

In 2016, we sold our first AI supercomputer, the DGX-1, to a then still unknown start-up and supported it in using large language models. It took another six years before OpenAI went public with ChatGPT and this technology could suddenly be used worldwide. That was AI’s “iPhone moment.”

However, we are currently observing a clear trend towards open-source and community models. Our customers have realized that they can operate large language models easily and efficiently in a controlled manner in their private cloud or their own data center and are not solely dependent on external services. This enables them to integrate confidential data into language models, improve models, use different models and provide access to company-internal or even confidential information.
 

Our current strategic focus is on democratizing artificial intelligence and making it accessible to everyone. Until now, open-source models for generative AI required an in-depth understanding of the models and the software stack, which represented a significant hurdle for many companies. To meet these challenges, we have developed our NVIDIA Inferencing Microservices (NIM). These are AI models that are provided in containers, optimized for the hardware platform used, and equipped with all the necessary software to operate the model efficiently.

These microservices are accessed via an API, which makes it possible to develop applications independently of the underlying language model. In addition, language models can be easily updated as soon as new and improved versions are available. If we consider 2022 to be the “iPhone moment” of AI, then NIM represents the “iPhone app” in this context, so to speak.

However, our development work goes even further. We are already working on AI agents and blueprints that essentially combine workflows with various AI models and queries to internal systems. In principle, we are creating a digital employee for simple applications that automates processes and increases efficiency.
 

First of all, I would like to congratulate ti&m on its impressive success over the past 20 years. It is remarkable how much the company has grown. I wish you continued success in the future. I look forward to supporting you in your transformation into the age of artificial intelligence and continuing our successful collaboration.