NVIDIA presented a lot of news yesterday taking advantage of the GTC 2023 scenarioa very important event that has become a reference in the technological sector at a professional level.

As the list of novelties is very long, and it would be counterproductive to collect them all In a huge article, we are going to focus on the four most important announcements, and we are going to detail them so that you know all their keys.

In general terms, the most interesting novelties that NVIDIA has presented this year are focused above all on the hardware, and are not limited to the traditional model but rather extend to quantum computing. It is precisely with this that we open this article, although we must not forget that the green giant has also confirmed an important boost to AI to be able to take it to any industry, and has strengthened its collaborations with the main giants of the sector.

NVIDIA DGX Quantum

Its about first GPU-accelerated quantum computing systemand represents a major inflection point as it shapes a next-generation computing solution powered by NVIDIA’s Grace Hopper superchip and the open source CUDA Quantum programming model, coupled with the world’s most advanced quantum control platform. world, OPX, from Quantum Machines.

With this system it is possible to unite the best of both worlds, classical computing and quantum computing, and this will allow researchers to have all the potential they need to create extraordinary applications, and without having to give up advanced features that are key in quantum computingsuch as calibration, control, quantum error correction, and hybrid algorithms.

At the hardware level, the NVIDIA DGX Quantum has a Grace Hopper super chip, as we said at the beginning, and is connected via a PCIe interface to a Quantum Machines OPX+ system. This allows reduce latency to less than a microsecondand allows near-instant communications between graphics cores and quantum processing units.

Five new GPUs for professional laptops

NVIDIA has also confirmed the release of five new professional laptop GPUs, based on the Ada Lovelace architecture and designed to shape more powerful, efficient and lighter workstations. The RTX 5000 is the most powerful of five, since it has 9,728 shaders, it has 76 RT cores, 304 tensor cores, comes with 16 GB of GDDR6 over a 256-bit bus, and has a configurable TGP of between 80 and 175 watts.

The RTX 4000 is a step behind with its configuration of 7,424 shaders, it has 56 RT cores, it has 232 tensor cores, it has 12 GB of GDDR6 on a 192-bit bus and its TGP can be configured between 60 and 175 watts. The RTX 3500 fits in the upper-middle range with its 5,120 shaders, 40 RT cores and 160 tensor cores, it has 12 GB of GDDR6 on a 192-bit bus and its TGP can be configured between 60 and 140 watts.

The RTX 3000 is a mid-range model with 4,608 shaders, 36 RT cores, 144 tensor cores, 128-bit bus, 8 GB of GDDR6, and a configurable TGP from 35 to 140 watts. Finally we have the RTX 2000, which has 3,072 shaders, 24 RT cores, 96 tensor cores, 128-bit bus, and 8 GB of GDDR6. Its TGP can be set between 35 and 140 watts. NVIDIA has also announced a RTX 4000 ADA SFF for Desktopwhich is nothing more than a small version equipped with 6,144 shaders.

NVIDIA H100 NVL, a dual GPU specialized in Large Language Models (LLM)

The green giant has also surprised us with a Hopper-based dual GPU setup that is specifically designed for work with Large Language Models (LLMs), large language models in a direct translation. It may not sound familiar to some of you and you may have even raised an eyebrow, but if I tell you that Chat-GPT is one of them, you will surely understand the importance of this new hardware.

The acronym NVL refers to NVLink, an interconnection technology that is what allows the two GPUs that make up this accelerator to be joined, and a total of three Gen4 bridges are used. At the specification level, we are facing an impressive solution, since it offers a power of up to 68 TFLOPs in FP64reaches the 134 TFLOPs in FP64 with their tensor cores, and can offer up to 1,979 TFLOPs in FP32 with such cores.

In FP16 the power with the tensor cores is 3,958 TFLOPs, and at INT8 we have a power of 7,916 TFLOPs. The numbers speak for themselves, and yes, they are truly impressive. The NVIDIA H100 NVL has a 6144-bit bus, features 188GB of 5.1GHz HBM3 memory, and offers up to 7.8TB/s. LLM models need large blocks of memory and high bandwidth, so NVIDIA hasn’t cut a thread.

In recent days we have heard the expression Generative Artificial Intelligence, or generative AI, everywhere. The blame lies, above all, with the latest developments in OpenAI in its GPT model, Microsoft’s improvements in its products or systems like Claude. But few have stopped to explain in detail what Artificial Generative Intelligence implies. Or its bases. Despite this, it is on more and more systems, from ChatGPT to DALL-E.

What is Generative Artificial Intelligence

This type of Artificial Intelligence is known as generative, because it is able to create something that does not exist. It is its main difference with respect to discriminative Artificial Intelligence, which is dedicated to making distinctions between different types of inputs, and tries to answer questions whose answer implies identifying something from a question that implies a choice.

For example, one of these AIs will be able to answer a question about an image, to answer if it is one thing or another. But you won’t be able to create an image from simple instructions. Something that generative AI can do.

Despite having made a lot of noise lately, generative AI has been around for quite some time. It goes back, in fact, to the appearance of Eliza, a chatbot that was quite popular years ago, and which pretended to be a therapist to talk to. They created it at MIT, and it was launched in 1966. It was a revolution then, despite being quite rudimentary. And years of work and research have evolved generative AI so much that Eliza now looks like something made by beginners.

The arrival of DALL-E, stable Diffusion, and above all, of Chat GPT, have turned Artificial Intelligence upside down, as well as the perception that the general public has of it. The first two allow the generation of realistic images from simple instructions.

The third is even capable of engaging in a text conversation with humans, and providing certain types of information. It is even likely that soon it will be multimodal, thanks to the evolution of its model, GPT, to its version 4. For now, it only allows you to respond with text, but in the future you may also be able to work with multimedia elements.

We usually refer to these systems, and others like them, as models. This denomination has not been made at random. It has occurred because all three are capable of representing an attempt to simulate or model some aspect of the real world based on a set, sometimes very large, of information about it.

How does Generative Artificial Intelligence work?

This type of Artificial Intelligence uses machine learning to process large amounts of data images or text. Most of this information is taken from the Internet. After processing it, it is able to determine which things are most likely to appear along with others that have already appeared. That is, they generate text by predicting which word is most likely to come after other words you have already created.

Most of the programming work of generative AI is dedicated to creating algorithms that can distinguish the things that interest the creators of the AI ​​they want to develop. In the case of ChatGPT, it will be words and phrases. In the case of DALL-E, graphics and visual elements.

But above all, it must be taken into account that this type of AI generates responses and outputs based on the assessment of a huge set of data, which have been used to train it. With them, it responds to requests and instructions with images or phrases that, based on what’s in that dataset, the generative AI suggests is likely to be appropriate.

The chip giant has launched the Intel Agilex 7a new generation of FPGA solutions (Field Programmable Gate Array, or matrix of logic gates programmable in the field in Spanish), with which the Santa Clara company has once again opted to maintain high versatility and a high level of performance.

According to Intel, the Agilex 7s are equipped with the Field Programmable Gate Array (FPGA) transceivers. fastest available on the market today, and have been designed to help different user profiles address challenges in the most bandwidth-demanding areas of the world, which are typically found in data-intensive industries, one way or another. The two clearest examples would be data centers and high-speed networks.

These new FPGAs have been built with cloud, network, and embedded systems in mind, and are F-Tile-enabled. Taking a look at the performance data we find very interesting data, since They are capable of reaching a speed of up to 1.6 Tbpsthey can work with 25/50G passive optical networks and also with different broadcast standards.

Shannon Poulin, Intel corporate vice president and general manager of the Programmable Solutions Group, commented:

“Intel’s Agilex 7 with F-Tile is loaded with transceivers that offer more flexibility, bandwidth and performance data rate than any other FPGA on the market today. Coupled with Intel manufacturing and our supply chain resilience, we are offering multiple industry-leading products and capabilities that our customers and the industry require to address a wide variety of critical business needs.”

Intel Agilex 7 FPGAs have been manufactured under the intel 10nm node, also known as Intel 7, which means they use a state-of-the-art process and are optimized to offer good value in terms of performance and power consumption. Compared to the previous generation from Intel these new FPGAs have achieved double the bandwidth per channel, reducing power and maintaining a compact and functional form factor. This was precisely one of Intel’s most important innovations at MWC 2023.