Minutes ago, Nvidia's new AI computer sent a shocking alert that has every other computer on edge! This beast, armed with Hopper architecture, is turning all older tech into relics. Giants like Microsoft and Google are the only ones who've touched this level of tech, and Nvidia's value has skyrocketed to two trillion dollars overnight. What exactly is this alert about, and how will it change the game for everyone else? Let us dive into what this urgent message from Nvidia's AI could mean for the future of technology. How Nvidia's Hopper Redefines Computing Nvidia's newest AI processor is creating a lot of excitement in the tech world, breaking through limits that were once thought to be impossible. This isn't just a small improvement—it's a major shift that is changing how we see technology. The buzz is all about the Hopper architecture,
which has reshaped the competitive landscape. With performance leaps that dwarf its predecessors, the Hopper architecture is setting a new standard, leaving other processors struggling to keep up. But what if this is just the beginning? Could this unprecedented power reshape industries overnight, rendering entire technologies obsolete? Stay tuned, because the future of computing is about to take an unexpected turn.
This AI processor is a game-changer, challenging even some of the basic principles of physics. Only four companies in the world have managed to create something like this: Microsoft, Apple, Google, and the company based in California, which saw its market value skyrocket from one trillion to two trillion dollars in just eight months. This huge jump came from the high demand for its cutting-edge technology, which is leading today's AI revolution. It's amazing to think that the company, which started in one thousand nine hundred ninety-three to make video game graphics better, has now become a major player in the AI world in the twenty-first century. In March of two thousand twenty-two, the Hopper architecture, designed especially for data centers to support AI work, was revealed.
This launch created a lot of excitement in the AI community and led to strong demand. But the real surprise came in two thousand twenty-three, during the AI boom, when prices for these products shot up due to shortages and heavy demand. People who ordered H100-based servers had to wait between thirty-six and fifty-two weeks to receive them. Despite these delays, the company still managed to sell five hundred thousand H100 accelerators in just the third quarter of two thousand twenty-three. The strong position in the AI market and the
success of its Hopper products played a big part in boosting the company’s market value. However, this wasn’t the end of the big moves. Looking to the future, the Blackwell architecture, named after the famous American mathematician David Blackwell, was introduced. Blackwell made
groundbreaking contributions to fields like game theory, probability, and statistics, and his work has had a huge impact on AI models and how they're trained. Interestingly, he was also the first African-American to be inducted into the National Academy of Sciences. In October twenty twenty-three, the updated plans for data center technology at an investor event were revealed. They introduced the B one hundred and B forty accelerators, part of the new Blackwell architecture. This was a change from their earlier plans, which called the next step
after Hopper just Hopper-Next. Later, on March eighteenth, twenty twenty-four, Blackwell was officially introduced at the Graphic Technology Conference (GTC). Over eleven thousand people attended the event, including software developers, industry experts, and investors. It was held over four days at the Pro Hockey Arena in San Jose, and the main speaker was the CEO, Jensen Huang. Huang explained that Blackwell is more than just a chip—it's a full platform. The company is famous for its GPUs, but Blackwell takes things a step further. At the core of this technology is Hopper,
which is currently the best GPU technology with an incredible twenty-eight billion transistors. What makes it unique is its architecture, which, for the first time, combines two dies into one chip that communicates at a mind-blowing speed of ten terabytes per second. It does this without facing memory or cache problems, acting as one large chip. Some people were doubtful about reaching such ambitious goals with Blackwell. But the company pushed forward and created a chip that fits perfectly into two different systems. One system works easily with Hopper for smooth upgrades, while the other, shown on a prototype board, demonstrates its powerful features and future possibilities. Imagine a system with two Blackwell chips and four Blackwell arrays,
all connected to a Grace CPU through super-fast connections. This setup could change the world of computing. Let's see how this new technology packs powerful computing into a small space. A Glimpse into Next-Gen AI Computing Huang highlighted that this system is groundbreaking because it packs so much computing power into a small space. However, the company didn’t stop there. To truly push the boundaries, they added new features that challenge the limits of physics. They introduced the fifth-generation MV Link inside the new Transformer engine, which is twice as fast as Hopper and enables computing within the network itself. This is
important because when several GPUs work together, they need to share and sync data efficiently. With this innovation, the company is setting a new bar in technology. The new AI supercomputer is built with incredibly fast connections that allow it to handle data right inside the network, making it much more powerful. Even though it's officially rated at one point eight terabytes per second, it actually performs even better than that, making it faster than the Hopper model. This new chip improves training speed by two and a half times compared to Hopper, thanks to its new FP6 format. It also comes with FP4, which makes tasks like quick responses and predictions happen twice as fast. These improvements are not just about speed. They help save energy,
reduce the amount of data that needs to travel through the network, and save time, which is becoming more important as AI technology grows. The company calls this phase Generative AI because it represents a big shift in how technology works. The newest processor is designed for this change, using FP4 to quickly create content. It’s part of the AI
supercomputer and can produce five times more output than the older Hopper model. But that’s not the most impressive part. The company is already working on an even bigger and stronger GPU that will go beyond what’s possible now. This new chip, which includes the MVLink switch, has fifty billion transistors, almost as many as the Hopper model. It also has four MV links, each running at one point eight terabytes per second,
allowing all the connected GPUs to work together at their top speed. This is a huge step forward, pushing the boundaries of what computers can do today. If you look back six years, the first DGX1 could handle one hundred seventy teraflops, or zero point one seven petaflops. Fast forward to now, and the company is aiming for seven
hundred twenty petaflops, getting closer to reaching one exaflop for training. This is a massive achievement, making it the world’s first exaflop machine, all in one system. To give you some background, only a few machines in the world today can reach exaflop levels of computing. This DGX isn’t just another AI tool; it’s a powerhouse, all neatly packed into a single, sleek rack. But what makes this possible? It’s thanks to the MVLink backbone, which provides an incredible bandwidth of one hundred thirty terabytes per second. To put that in perspective, this speed is faster than the entire Internet combined. And it
does this without needing expensive optics or transceivers, which also helps save a lot of energy—about twenty kilowatts in a system that usually uses one hundred twenty kilowatts. The system keeps a steady temperature of twenty-five degrees Celsius while it’s working, with the help of air conditioning. The water that comes out is around forty-five degrees Celsius, similar to the temperature of a hot tub, and flows at a rate of two liters per second. This
setup ensures everything stays cool and runs smoothly. Now, let’s talk about an important part of this system: the GPU. While some might see it as just another component, for Nvidia, it’s a game-changer. The days of clunky GPUs are over;
modern GPUs are incredibly complex, made up of about six hundred thousand parts and weighing around three thousand pounds. That’s about one and a half tons—pretty amazing, right? Now, let's see how these tech upgrades make training AI faster and more efficient. Powering AI with Less Training a GPT model with one point eight trillion parameters is a huge challenge. Not long ago, this process could take several months and use a lot of energy. But thanks to
advances in technology, especially in GPU architectures like the Hopper and more recently, the Blackwell architecture, things have changed dramatically. With the Hopper architecture, training models of this size required eight thousand GPUs and about fifteen megawatts of power, and the whole process took about ninety days. This was already a big improvement over older setups that needed even more resources. But then came the Blackwell architecture, which took things even further. Now, the same task can be done with just
two thousand GPUs and only four megawatts of power, making everything much more efficient. The Blackwell architecture has several upgrades that focus on reducing energy use while boosting performance. It includes a second-generation Transformer Engine, which optimizes processing by using as little as four bits per neuron in the neural network. This doubles the compute bandwidth, allowing faster processing of large language models without increasing energy consumption. But this wasn't the most impressive part. Blackwell's NVLink switch also improves GPU communication. It can handle up to one point eight terabytes of traffic in each direction and
can support up to five hundred seventy-six GPUs, compared to Hopper's limit of two hundred fifty-six. The raw power of the Blackwell B200 model is also noteworthy, delivering up to eighteen PFLOPS for certain tasks and boasting a memory bandwidth of eight terabytes per second, making data transfer and processing lightning fast. These GPUs aren’t just about raw power, though. They are designed to handle complex scientific computing and simulations. For example,
Blackwell GPUs can run simulations for things like heat, airflow, and power use in virtual models of data centers. This ability can speed up such simulations by up to thirty times compared to traditional CPUs, making everything more energy-efficient and sustainable. And there’s more. The high-bandwidth memory and advanced tensor cores in these GPUs allow them to handle even the toughest AI training and inference tasks. This makes them suitable for a wide variety of applications, from scientific research to enterprise AI solutions. The real breakthrough here is not just in cutting down the resources needed to train large models, but also in showing how quickly AI hardware is evolving to keep up with the growing complexity of AI tasks.
At the center of the Groot Project is Isaac LAB, a unique platform created by the company for training robots. Using Omniverse Isaac Sim, which allows for simulations, Isaac LAB gives robots a virtual space to practice and improve their skills, helping them get better at handling real-world challenges. Working alongside Isaac LAB, Blackwell is set to become the company’s biggest product launch yet. As we explore the world of Robotics, it’s clear this technology is about to make a huge impact. The company is stepping into the fascinating world of physical AI, where machines can interact with the real world. Up until now, AI has mainly worked in the
digital space, limited to systems like DGX. But imagine a future where AI goes beyond that, where robots have their intelligence and can move around the physical world on their own. This is what the company calls the Robotics ChatGPT Moment. The company has been hard at work developing advanced robotics systems, including AI training from DGX and AGX, the world’s first robotic processor designed to handle high-speed sensor data in an energy-efficient way. Next, we'll explore how these advancements are shaking up the world of robotics.
Bridging Virtual Realities and Robotics The company's Omniverse is a platform that connects the virtual and real worlds using Azure cloud services. Imagine a warehouse that runs on its own, where people and machines work together smoothly. In this setup, the warehouse works like a traffic controller in the sky, making sure everything moves safely and in order. What’s even better? You can interact with this modern warehouse right away. Each machine has its own robotic system, making the whole process more efficient. Thanks to Blackwell’s leadership, it feels like robotics are closer than ever. But the journey doesn’t end here.
Let’s take a closer look at the progress being made in robotics. We are getting closer to the age of humanoid robots, and it’s expected that robots will soon be taking over different industries. They can offer a safer and more efficient way of working. One industry that’s about to see a big change is the car industry. Next year, the company plans to team up with Mercedes, and later with JLR, to introduce its latest technology. The company’s CEO,
Jensen Huang, also announced that BYD, the world’s largest electric vehicle maker, will use the company’s newest creation: Thor. Thor is a powerful computer system designed for advanced machines. This could change the world of robotics and might even lead us closer to having humanoid robots. But this isn’t the most exciting part yet.
Nvidia’s Groot Project is one of the most exciting projects in the world of robotics. It aims to change how robots learn and interact with their surroundings. The company isn’t just thinking about these ideas—they are making them a reality. The Groot Project is a huge step forward,
giving robots the ability to follow complex instructions and learn from their past experiences. With advanced algorithms, these robots can decide what to do next on their own, making the connection between human instructions and robotic actions even better. But that’s not the end of it. There’s still more to come. Osmo is a new platform that helps organize and run training sessions and simulations more efficiently. It uses powerful DGX and OVX systems to ensure everything goes smoothly. One of the most impressive things about the Groot Project is how it can learn with very little help from people. By watching just a few examples, robots equipped with Groot technology can perform everyday tasks and copy human actions with great accuracy. This is made possible by advanced technology, which uses complex systems
to understand what people do and turn those actions into tasks that robots can perform. But that's not all. The technology doesn't just make robots move; it also helps them understand and respond to spoken commands, making them even more useful and easy to interact with. Whether it’s simple gestures or more complicated tasks, robots using Groot technology show a lot of intelligence and flexibility. This is thanks to
Jetson Thor robotic chips, which were designed specifically to power these advanced robots. The Groot Project's impact goes beyond cool features. It is leading the way in robotics, bringing us closer to a future where robots are a normal part of daily life, transforming industries and how we use technology. The commitment to cutting-edge technology promises a bright future for humanoid robots with endless possibilities. Looking back, the journey started in the nineteen-nineties, when technology was quite different. Personal computers were just starting to become popular, but graphics were very basic,
often limited to text and simple images unless used for special purposes like animation or engineering. For better graphics, people relied on video game consoles. However, three engineers from California—Jensen Huang, Chris Malachowski, and Curtis Priem—had a different vision. They wanted to create a special chip that could handle more complex graphics on personal computers. Let's step back and see how it all started with a game-changing idea in GPU tech.
The Birth of Nvidia and the GPU After many long brainstorming sessions, often at a local diner, they realized that while regular computer processors (CPUs) were good at handling one task at a time, creating 3D graphics for games needed something that could handle many tasks at once. Their solution was a new type of chip that could process tasks in parallel, which changed the world of computing. This chip, later known as the GPU, wasn’t meant to replace the CPU but to work alongside it, especially for graphics-heavy tasks. But that wasn’t the hardest challenge. In its early days, the focus was on making PC gaming
better, a market that was quickly growing. The company was founded in 1993 in a small condo in Fremont, California, and its main goal was to bring parallel processing GPUs into regular household computers. The name 'Nvidia' comes from 'NV', short for 'next version', and 'Invidia', the Latin word for 'envy'. The green in their logo was chosen because their powerful chips were meant to make others envious.
The company was started by three skilled engineers. Jensen Huang, a Taiwanese-American electrical engineer, had a lot of experience from working as the Director of CoreWare at LSI Logic and designing microprocessors at AMD. Chris Malachowski brought in valuable engineering skills from his time at HP and Sun Microsystems. Curtis Priem, on the other hand, had designed graphic chips at IBM and Sun Microsystems. Even with all their expertise, starting a new company wasn't easy.
In nineteen ninety-three, the three founders reached a point where they weren't sure how to move forward. To help them navigate the legal side of things, they decided to hire a lawyer. Jensen Huang only had two hundred dollars at the time, but he chose to invest this small amount to officially start the company. This initial investment not only got the company incorporated, but it also gave Huang a twenty percent stake in it. However, this was just the beginning of their journey. The next big challenge was getting enough money to turn their ideas into reality. Convincing investors was not easy because many venture capitalists preferred to back founders who already had successful businesses and a clear, exciting vision. Still, the company eventually caught the attention
of Sequoia Capital and Sutter Hill Ventures, securing twenty million dollars in funding. Looking back, it’s clear how important those early investments were. Huang’s connection with the CEO of LSI Logic came in handy, helping them secure a meeting with Sequoia Capital, the same firm that had also invested in LSI Logic. Although Sequoia was initially unsure, they eventually recognized the potential in the graphics card market and decided to invest.
They first put in two million dollars, which was later increased by an additional eighteen million, putting the company on the path to success. But this wasn’t the hardest part yet. Back then, investing was seen as risky. Out of eighty-nine companies with similar goals, only AMD and the company survived the intense competition. By the time it went public in nineteen ninety-nine, its value shot up to six hundred million dollars, proving the early investors had made a smart bet. But we’re getting ahead of the story—the
company was still a small group of engineers working hard to launch their first product. With the funding secured, it took two more years to build a team and create their first product, the NV1, which came out in nineteen ninety-five. During this time, the company made a deal with Sega to produce chips for their gaming consoles. These chips powered popular games like Virtual Fighter and Daytona. Interestingly, the NV1 chips were also compatible with PCs, allowing players to enjoy Sega Saturn games on their computers—an exciting concept for its time. The company took a bold step in how they designed the NV1 chips, using a rendering method based on quadrilaterals instead of the more common triangles. The goal was to speed up
the rendering process by lightening the load on the CPU. In theory, this would use fewer polygons and better capture rounded shapes, giving game designers more creative freedom. But as technology advanced and memory became cheaper, this method became less effective and even caused compatibility issues since it didn’t work with OpenGL. Could quadrilateral rendering make a comeback with today's tech? Like, comment, and don't forget to subscribe for more discussions!
2024-11-18