Graphics Processing Unit (GPU) Evolution

The unsung hero that transformed GPUs into the hardware powering modern AI. In 2003, Ian Buck was a phD student at Stanford’s Computer Graphics Laboratory. During his PhD, he published the paper “Brook for GPUs: Stream Computing on Graphics Hardware”, the programming language that turned GPUs into affordable supercomputers for general‑purpose applications, not just games. At the time, whoever wanted to use a GPU for non-graphic tasks had to "trick" the hardware by writing OpenGL code. Brook's greatest innovation was allowing developers to use a familiar C-like syntax instead. This didn't become a hit on release, but it caught the eye of exactly the one person who mattered: Jensen Huang. He hired Buck and redirected R&D resources towards his vision. Inside NVIDIA, Buck teamed up with John Nickolls and a group of ex‑Silicon Graphics engineers. And 2 years later, in 2006, they released the first version of what they called Compute Unified Device Architecture (CUDA). It was free software, but locked to NVIDIA hardware. Downloads were in the thousands, but it slowly started gaining adoption in scientific communities. It was used for complex modeling in physics, fluid mechanics, chemistry, and biology. After 2012, with the publication of Alexnet, the R&D started to pay-off. Neural networks were previously trained on CPUs, and the transition to CUDA opened a way to train them in days instead of months. The rest is history. 🔗 Interview with Ian Buck at GTC 2008: https://lnkd.in/esM5PQH2 🔗 Ian’s CUDA tutorial from 2009: https://lnkd.in/e8WdgBYJ 🔗 The original Brook: https://lnkd.in/eeMRJp9p

Evolving thoughts on Traditional GPU Depreciation and useful life Models Eight months ago I argued that Rubin and its predecessor would make the B200 feel obsolete almost as soon as it arrived. The bigger shift, though, is not just performance. It is how these platforms break our assumptions about useful life and depreciation of AI hardware. Under traditional GAAP treatment, accelerated compute often sits on a six year schedule. In a world of compounding hardware and software gains, that is already wrong. For the most constrained operators, the true economic life of frontier GPUs and XPUs is closer to three to four years at best, and in some cases approaches zero for production use. The performance curve tells the story. On equivalent models H100 to B200 delivered around 5x tokens per second with B300hitting 8x over the H100 and Rubin, with dual logic chips and HBM4, speeds this trend even further. Even Jensen CEO of Nvidia is famous in saying "when Blackwell starts shipping in volume, you couldn't give Hoppers away". These are step changes driven by new numerical formats like FP4 and architecture level shifts that older hardware cannot simply inherit. Now add power and access constraints. Frontier cloud providers and large model labs are constrained on both top end GPUs and the power envelopes to run them. Every megawatt wants to be filled with the most tokens per joule available. If the combined software and hardware curve keeps delivering effective order of magnitude gains per generation, keeping previous generations in production for state of the art training or high volume inference becomes economically negative. The opportunity cost of not swapping in the latest parts dwarfs residual book value. That does not make the hardware worthless. It makes the value cohort specific. At the frontier, useful life can collapse once the next generation lands and can be powered at scale. For enterprises, governments, and regional providers without day one access, previous generation GPUs and XPUs are still a massive step up from their baseline. This is where a secondary market and NEOcloud type providers matter. Hyperscalers and specialized AI clouds can run very short economic lives on the newest hardware for their largest tenants, then cascade those systems into second and third tier customers once initial commitments burn down. The first cohort lives on ultra short depreciation, the next cohorts can reasonably sit in a three to four year band, and only the broad enterprise base looks anything like the old five to six year cycles. In other words, depreciation curves are no longer just an accounting choice. They reflect where you sit in the access hierarchy for GPUs, power, and data, and whether you are training the next state of the art or building durable inference platforms on top of it. KamiwazaAI #1trillionInferences Keith Townsend Ryan Shrout Matthew Wallace

Whats’s a GPU and why it’s so important for computers 💻 Think of your computer like a kitchen. The CPU (Central Processing Unit) is like a master chef—it handles all kinds of tasks but can only do a few at a time. It’s designed for general-purpose computing, like running software, browsing the internet, or managing files. CPUs are powerful but not the best at handling repetitive, high-volume tasks quickly. The GPU (Graphics Processing Unit), on the other hand, is like a team of line cooks in a restaurant. Instead of focusing on one or two complex tasks, it can handle thousands of small tasks simultaneously. Originally, GPUs were designed for rendering graphics in video games, processing huge amounts of pixels and colors at once. Why Was NVIDIA So Important? NVIDIA realized that GPUs could do more than just make video games look good. Their key innovation was figuring out that GPUs are perfect for parallel processing, meaning they can solve many small calculations at the same time. This made GPUs ideal for artificial intelligence (AI), cryptocurrency mining, scientific research, and machine learning. Instead of just drawing graphics, NVIDIA’s GPUs started being used for training AI models, simulating physics, and even helping self-driving cars process vast amounts of sensor data in real time. The Impact • AI Boom: GPUs became the backbone of AI and machine learning, helping companies like OpenAI (ChatGPT), Google, and Tesla. • Cryptocurrency Explosion: Bitcoin and Ethereum mining took off because GPUs could crunch numbers better than CPUs. • Scientific Breakthroughs: GPUs are used in medicine, weather forecasting, and deep space research. • Gaming & Media Revolution: NVIDIA made high-quality gaming possible with ultra-realistic graphics. In short, NVIDIA didn’t just make graphics cards—they redefined computing by proving that GPUs could handle workloads far beyond gaming. Now, GPUs power the future of AI, automation, and even the metaverse.

In 1999, NVIDIA nearly went bankrupt making gaming chips. Now those same chips power ChatGPT and made them worth $4.5T. Here's how NVIDIA went from near-death to dominating AI: Gaming was never the endgame for NVIDIA... 1993: Jensen Huang and 2 engineers started NVIDIA in a Fremont condo. 1999: Released GeForce 256, the first official GPU after nearly going bankrupt. Key move? Went fabless, outsourcing chip production to TSMC. Smart early decision. Also their biggest modern risk. 2006: NVIDIA released CUDA: A platform that flipped GPUs from serial to parallel computing. Thousands of small processors working at once instead of one big one. Wall Street valued it at $0 for 10 years. Jensen kept investing anyway. CUDA enabled deep learning, a completely new way to write software. No market. No revenue. No proof it would work. Just Jensen building for a future nobody else could see. He kept pouring millions into it. 2012: AlexNet hit. A neural network running on NVIDIA GPUs crushed an image recognition contest. Proved deep learning actually worked. Parallel processing for graphics was perfect for AI. 10 years of patience paid off: Jensen didn't make every bet right. Tegra smartphone chips flopped in the early 2010s. A $40B Arm acquisition got blocked. But those Tegra chips? Now they power warehouse robots and early Teslas. Great founders repurpose their failures. From there, Jensen went all in on AI. Every engineering team shifted focus. Sales retrained on AI use cases. Product roadmaps rebuilt from scratch. While others stayed skeptical, NVIDIA was building what the AI revolution would run on. 2022: ChatGPT drops. Microsoft trained it with 10,000 NVIDIA A100 GPUs at $200,000 per server board. Today, NVIDIA's one of the world's top 10 most valuable companies. While AI drives the headlines, their real edge is the full ecosystem. They don't just make chips, they power autonomous driving, factory robotics, massive simulations. Data center CPUs, self-driving cars, everything in between. 30 years in, Jensen still runs the company. True builders don't exit. They evolve. Most companies would've played it safe. Jensen doubled down on CUDA with zero revenue for a decade. Wall Street called it wasted capital. He called it building the rails before anyone needed them. That’s the principle we live by at NextLink Labs. You don't build to get noticed. You build so when the wave hits, you're already the foundation it runs on. Clean execution. Systems that scale. Technical work that lasts. If you’re leading a company built for longevity, not hype, let’s talk. We build software that scales, systems that secure, leadership that delivers: NextLinkLabs.com

𝕋𝕙𝕖 𝕟𝕖𝕩𝕥 𝕘𝕖𝕟𝕖𝕣𝕒𝕥𝕚𝕠𝕟 𝕠𝕗 𝔾ℙ𝕌𝕤 𝕚𝕤 𝕒𝕣𝕣𝕚𝕧𝕚𝕟𝕘 The architecture of intelligence is shifting again. NVIDIA’s Blackwell and AMD’s Instinct Helios represent the threshold of a new era—one where computation is no longer measured in FLOPs, but in joules per token. The future of GPU design is about efficiency, not excess. It’s about collapsing distance—between memory and compute, between photonics and silicon, between the physical and the cognitive. NVIDIA’s upcoming Rubin and Kyber systems promise optical interconnects capable of sub-picojoule transfers, while AMD’s MI450 is preparing to debut on a 2 nm process with HBM4 memory that fundamentally rewrites the bandwidth equation. In the labs, wafer-scale compute, neuromorphic chiplets, and photonic cores are redefining what a “GPU” even means. The war is no longer over speed alone—it’s over energy, precision, and intelligence density. Every generation brings us closer to machines that think faster, learn deeper, and do so on a fraction of the power. We are standing at the edge of the most profound transformation in computing since the transistor. And this time, it’s not just evolution. It’s convergence—quantum, photonic, and neural—all colliding in silicon. Singularity Systems will be there at the frontier. With friends in the field like NVIDIA’s Dr. Jochen Papenbrock and AMD’s Alexey Navolokin i’m hoping to bring the next generation of artificial intelligence to the world.

Moore’s Law didn’t just power better gadgets. It built a new civilization. For half a century, the steady doubling of transistors made computing exponentially more powerful — and exponentially cheaper. That single principle turned silicon into the engine of finance, logistics, medicine, entertainment, and war. In 1965, a chip held 64 transistors. Today, the most advanced chips exceed 200 billion. Moore’s Law was the engine of the digital age. But as AI arrived, something changed. The demands of training a general intelligence quickly outpaced what classical CPU scaling could provide. Doubling transistor counts every two years was no longer enough. We needed a new paradigm — accelerated computing — and a new scaling law. GPUs, and the architecture behind them, became that answer. Under what’s now sometimes called 'Huang’s Law' in honor of Jensen, AI performance has improved not every two years, but on a cadence closer to every 6–12 months, driven by advances in parallel compute, software optimisation, networking, and systems design. Moore’s Law scaled hardware. Huang’s Law scales the ability of that hardware to create intelligence. And that marks the beginning of a new era — one where we are no longer merely scaling machines, but scaling intelligence itself. The consequences will be even more profound than the digital revolution Moore unleashed.

I was recently reading "Thinking Machines" the wonderfully written biography of Jensen Huang and NVIDIA's unlikely journey from a video game hardware startup to the world's most valuable corporation powering the AI revolution and came across this fascinating story. We take the term "GPU" for granted today, but how did this term come to life? It was as much a technical innovation as it was a marketing stroke of genius. Back in the late 1990s, the graphics card world was dominated by a company called 3dfx and their famous Voodoo cards. Think of 3dfx as the king of gaming graphics at the time - their Voodoo cards were THE choice for gamers who wanted the best visual experience. But Nvidia had other plans. In 1999, Jensen Huang and team launched what they called the "Voodoo killer" - the GeForce (short for "Geometry Force"). This wasn't just another graphics card. It could render 10 million triangles per second and handle complex 3D transformations that previously required expensive workstations costing millions. Here's the genius part: Nvidia turned to Dan Vivoli, a clever marketer who saw their limited marketing budget as an opportunity, not a problem. Instead of expensive ad campaigns, he reached out directly to hardware reviewers with a simple message: "This is the world's first graphics-processing unit - or GPU." Vivoli had literally just made up the term "GPU." But it worked. Reviewers started using the term, grouping all graphics cards under this new category. Soon, everyone was calling them GPUs. As Vivoli put it: "We invented the category so we could be the leader in it." Meanwhile, 3dfx engineers were frustrated. Not just because of Nvidia's marketing tactics, but because Nvidia had genuinely out-engineered them. The GeForce could handle four rendering pipelines on a single chip, while 3dfx's competing prototype needed four separate chips and still couldn't match the performance. Sometimes the most powerful innovations happen at the intersection of technology and storytelling. Nvidia didn’t win because they named the term 'GPU', they earned it with engineering, then ensured it stuck with a story. Great products make the future possible. Great stories make it inevitable. Breakthroughs start with engineering but they don’t travel without narrative. #NvidiaHistory #PowerOfNarrative #ThinkingMachines #AI #AILiteracy AI&Beyond Jaspreet Bindra

In 1993, NVIDIA discovered something remarkable Just 10% of code handles 99% of processing workloads. This insight sparked our obsession with parallel computing. We invested tens of billions in R&D, spent a decade in focused development, and completely rebuilt the computing stack. Many wanted quick profits. We chose long-term transformation instead. Gaming drove our parallel computing revolution. 🎮 The gaming market proved crucial: - Provided massive market volume - Funded continuous R&D - Enabled high-volume GPU production - Created global distribution channels 2012 brought our defining moment: researchers used our gaming GPU (GTX 580) to train AlexNet, achieving unprecedented computer vision accuracy. This showed us exactly where parallel computing would take AI. Our breakthroughs delivered: - CUDA: Unlocking parallel processing for millions - DGX Platform: AI supercomputing for every lab - 10,000x energy efficiency gains since 2016 - Price evolution: From $250K systems to $3K workstations These capabilities serve: - AI researchers advancing the field - Companies scaling compute infrastructure - Teams developing next-gen applications - Scientists accelerating discoveries Core principles drive technological breakthroughs. Our conviction in parallel processing reshaped computing after 65 years of CPU architecture. As Jensen says: "At some point, you have to believe something." Today we unveil our latest breakthrough: Omniverse + Cosmos fusion brings this parallel computing vision to robotics and simulation. The next computing revolution starts now. Building something that demands massive compute?

In 2003, Mark Harris hacked NVIDIA's gaming chips into simulating clouds. Jensen Huang hired him. Harris was a PhD student with a problem: weather simulations required supercomputer access he couldn't afford. So he wrote code that tricked an Nvidia gaming chip into solving physics equations. It ran 10x faster than anything he'd tried before. He published his hack. Within a year, scientists around the world were doing the same thing. Nvidia's gaming chips were secretly becoming the backbone of academic research. When Jensen Huang discovered what was happening, he had a choice: protect the gaming business or chase an unproven market of broke researchers. He chose the researchers. And he hired Mark Harris. In 2006, Nvidia launched CUDA, a platform to help developers use Nvidia chips for scientific computing. It was a billion-dollar bet: redesigning chips with features gamers would never pay for, then building programming tools and training for a market of a few hundred university researchers. Wall Street was furious. Why would a video game company sacrifice profits for academics nobody had heard of? Jensen's answer: "Accelerated computing is going to be essential in the future." For six years, it looked like a disaster. Gross margins dropped from 45% to 35%. Analysts downgraded the stock. Engineers questioned whether leadership had lost its mind. Then in 2012, a neural network built on Nvidia GPUs crushed an image recognition contest, stunning the world. The AI revolution had arrived, and Nvidia owned the only infrastructure ecosystem that mattered. In 2025, that "crazy bet" made Nvidia the first company to reach a $4 trillion market cap. The hardest part of a transformational bet isn't making it. It's holding on while everyone calls you delusional. To learn more, check out "The Nvidia Way" by Tae Kim. Link in Comments.

𝐆𝐏𝐔 = 𝐆𝐚𝐦𝐢𝐧𝐠 Isn’t that what most of us have thought for years? But did you know that these powerful chips are also driving innovation across industries you might never expect? Industries like medical imaging, financial modeling, automotive technology, and even space exploration. Computing has come a long way, and at the heart of this evolution are two key players: CPUs and GPUs. While CPUs are designed for handling sequential tasks efficiently, GPUs are the multitaskers, taking computing to a whole new level by processing thousands of tasks simultaneously. Chances are, you’ve seen this video of the Mythbusters, Adam Savage, and Jamie Hyneman, demonstrating this difference using paintball cannons. A single paintball cannon represents a CPU, shooting one paintball at a time — this is sequential processing in action. Now imagine 1,100 paintball cannons firing all at once; that’s your GPU, blasting through tasks with parallel processing. The difference is clear: while CPUs excel at detailed, step-by-step tasks, GPUs are built for scenarios that require handling massive amounts of data at lightning speed. Did you know the global GPU market size is calculated at $75.7 billion this year? And the world’s top chipmakers — AMD, Intel, and Nvidia — have already shipped 70 million GPUs in the first quarter. But how did we get here? Nvidia, the company that pioneered the modern GPU, was founded in 1993 with a vision to transform visual computing. At the time, computers were limited to sequential processing. Great for basic tasks but slow for anything requiring large-scale calculations. Nvidia's breakthrough was developing a chip capable of handling thousands of tasks at once, drastically improving computer performance, especially in gaming, scientific research, and later, AI. Fast forward to today, the company is valued at nearly $3 trillion, largely thanks to its innovations in GPUs that have drastically changed what computers can do. Over the years, we’ve seen several inflection points in computing: the explosion of the internet, the rapid growth of artificial intelligence, and the rise of GPUs. Each step brought new challenges that demanded more power, speed, and efficiency — and GPUs delivered every time. With the global GPU market projected to reach around $1,414.39 billion by 2034, it's clear that these powerful chips are becoming an essential component of our digital future. GPUs will keep us moving forward as AI and data demands keep rising. The applications are limitless: from accelerating and improving real-time healthcare diagnostics, to powering driverless cars that navigate with human-like precision, to facilitating advancements in climate change research. So, the question is no longer "what can GPUs do?" but rather "what can't they do?" Video source: NVIDIA #computing #AI #innovation #technology #GPU