The NVIDIA accelerated computing platform is main supercomputing benchmarks as soon as dominated by CPUs, enabling AI, science, enterprise and computing effectivity worldwide.
Moore’s Legislation has run its course, and parallel processing is the best way ahead. With this evolution, NVIDIA GPU platforms at the moment are uniquely positioned to ship on the three scaling legal guidelines — pretraining, post-training and test-time compute — for all the pieces from next-generation recommender programs and enormous language fashions (LLMs) to AI brokers and past.
The CPU-to-GPU Transition: A Historic Shift in Computing 🔗
At SC25, NVIDIA founder and CEO Jensen Huang highlighted the shifting panorama. Throughout the TOP100, a subset of the TOP500 checklist of supercomputers, over 85% of programs use GPUs. This flip represents a historic transition from the serial‑processing paradigm of CPUs to massively parallel accelerated architectures.
Earlier than 2012, machine studying was primarily based on programmed logic. Statistical fashions have been used and ran effectively on CPUs as a corpus of hard-coded guidelines. However this all modified when AlexNet working on gaming GPUs demonstrated picture classification might be realized by examples. Its implications have been huge for the way forward for AI, with parallel processing on growing sums of information on GPUs driving a brand new wave of computing.
This flip isn’t nearly {hardware}. It’s about platforms unlocking new science. GPUs ship much more operations per watt, making exascale sensible with out untenable power calls for.
Current outcomes from the Green500, a rating of the world’s most energy-efficient supercomputers, underscore the distinction between GPUs versus CPUs. The highest 5 performers on this business normal benchmark have been all NVIDIA GPUs, delivering a mean of 70.1 gigaflops per watt. In the meantime, the highest CPU-only programs supplied 15.5 flops per watt on common. This 4.5x differential between GPUs versus CPUs on power effectivity highlights the large TCO (complete price of possession) benefit of shifting these programs to GPUs.
One other measure of the CPU-versus-GPU energy-efficiency and efficiency differential arrived with NVIDIA’s outcomes on the Graph500. NVIDIA delivered a record-breaking end result of 410 trillion traversed edges per second, putting first on the Graph500 breadth-first search checklist.
The profitable run greater than doubled the following highest rating and utilized 8,192 NVIDIA H100 GPUs to course of a graph with 2.2 trillion vertices and 35 trillion edges. That compares with the following greatest end result on the checklist, which required roughly 150,000 CPUs for this workload. {Hardware} footprint reductions of this scale save time, cash and power.
But NVIDIA showcased at SC25 that its AI supercomputing platform is way over GPUs. Networking, CUDA libraries, reminiscence, storage and orchestration are co-designed to ship a full-stack platform.
Enabled by CUDA, NVIDIA is a full-stack platform. Open-source libraries and frameworks reminiscent of these within the CUDA-X ecosystem are the place massive speedups happen. Snowflake just lately introduced an integration of NVIDIA A10 GPUs to supercharge information science workflows. Snowflake ML now comes preinstalled with NVIDIA cuML and cuDF libraries to speed up widespread ML algorithms with these GPUs.
With this native integration, Snowflake’s customers can simply speed up mannequin growth cycles with no code modifications required. NVIDIA’s benchmark runs present 5x much less time required for Random Forest and as much as 200x for HDBSCAN on NVIDIA A10 GPUs in contrast with CPUs.
The flip was the turning level. The scaling legal guidelines are the trajectory ahead. And at each stage, GPUs are the engine driving AI into its subsequent chapter.
However CUDA-X and plenty of open-source software program libraries and frameworks are the place a lot of the magic occurs. CUDA-X libraries speed up workloads throughout each business and utility — engineering, finance, information analytics, genomics, biology, chemistry, telecommunications, robotics and far more.
“The world has an enormous funding in non-AI software program. From information processing to science and engineering simulations, representing lots of of billions of {dollars} in compute cloud computing spend every year,” Huang mentioned on NVIDIA’s current incomes name.
Many purposes that after ran solely on CPUs at the moment are quickly shifting to CUDA GPUs. “Accelerated computing has reached a tipping level. AI has additionally reached a tipping level and is reworking present purposes whereas enabling fully new ones,” he mentioned.
What started as an power‑effectivity crucial has matured right into a scientific platform: simulation and AI fused at scale. The management of NVIDIA GPUs within the TOP100 is each proof of this trajectory and a sign of what comes subsequent — breakthroughs throughout each self-discipline.
In consequence, researchers can now practice trillion‑parameter fashions, simulate fusion reactors and speed up drug discovery at scales CPUs alone might by no means attain.
The Three Scaling Legal guidelines Driving AI’s Subsequent Frontier 🔗
The change from CPUs to GPUs is not only a milestone in supercomputing. It’s the muse for the three scaling legal guidelines that symbolize the roadmap for AI’s subsequent workflow: pretraining, submit‑coaching and check‑time scaling.
Pre‑coaching scaling was the primary legislation to help the business. Researchers found that as datasets, parameter counts and compute grew, mannequin efficiency improved predictably. Doubling the information or parameters meant leaps in accuracy and flexibility.
On the most recent MLPerf Coaching business benchmarks, the NVIDIA platform delivered the best efficiency on each check and was the one platform to submit on all exams. With out GPUs, the “greater is best” period of AI analysis would have stalled underneath the burden of energy budgets and time constraints.
Publish‑coaching scaling extends the story. As soon as a basis mannequin is constructed, it should be refined — tuned for industries, languages or security constraints. Methods like reinforcement studying from human suggestions, pruning and distillation require huge extra compute. In some circumstances, the calls for rival pre‑coaching itself. This is sort of a scholar enhancing after primary training. GPUs once more present the horsepower, enabling continuous high-quality‑tuning and adaptation throughout domains.
Check‑time scaling, the latest legislation, could show probably the most transformative. Trendy fashions powered by mixture-of-experts architectures can cause, plan and consider a number of options in actual time. Chain‑of‑thought reasoning, generative search and agentic AI demand dynamic, recursive compute — usually exceeding pretraining necessities. This stage will drive exponential demand for inference infrastructure — from information facilities to edge gadgets.
Collectively, these three legal guidelines clarify the demand for GPUs for brand new AI workloads. Pretraining scaling has made GPUs indispensable. Publish‑coaching scaling has bolstered their position in refinement. Check‑time scaling is guaranteeing GPUs stay important lengthy after coaching ends. That is the following chapter in accelerated computing: a lifecycle the place GPUs energy each stage of AI — from studying to reasoning to deployment.
Generative, Agentic, Bodily AI and Past 🔗
The world of AI is increasing far past primary recommenders, chatbots and textual content technology. VLMs, or imaginative and prescient language fashions, are AI programs combining laptop imaginative and prescient and pure language processing for understanding and decoding photographs and textual content. And recommender programs — the engines behind personalised buying, streaming and social feeds — are however certainly one of many examples of how the large transition from CPUs to GPUs is reshaping AI.
In the meantime, generative AI is reworking all the pieces from robotics and autonomous autos to software-as-a-service firms and represents an enormous funding in startups.
NVIDIA platforms are the one to run on the entire main generative AI fashions and deal with 1.4 million open-source fashions.
As soon as constrained by CPU architectures, recommender programs struggled to seize the complexity of person conduct at scale. With CUDA GPUs, pretraining scaling allows fashions to study from huge datasets of clicks, purchases and preferences, uncovering richer patterns. Publish‑coaching scaling high-quality‑tunes these fashions for particular domains, sharpening personalization for industries from retail to leisure. On main world on-line websites, even a 1% acquire in relevance accuracy of suggestions can yield billions extra in gross sales.
Digital commerce gross sales are anticipated to succeed in $6.4 trillion worldwide for 2025, in line with Emarketer.
The world’s hyperscalers, a trillion-dollar business, are reworking search, suggestions and content material understanding from classical machine studying to generative AI. NVIDIA CUDA excels at each and is the perfect platform for this transition driving infrastructure funding measured in lots of of billions of {dollars}.
Now, check‑time scaling is reworking inference itself: recommender engines can cause dynamically, evaluating a number of choices in actual time to ship context‑conscious strategies. The result’s a leap in precision and relevance — suggestions that really feel much less like static lists and extra like clever steering. GPUs and scaling legal guidelines are turning suggestion from a background function right into a frontline functionality of agentic AI, enabling billions of individuals to kind via trillions of issues on the web with an ease that may in any other case be unfeasible.
What started as conversational interfaces powered by LLMs is now evolving into clever, autonomous programs poised to reshape almost each sector of the worldwide financial system.
We’re experiencing a foundational shift — from AI as a digital expertise to AI getting into the bodily world. This transformation calls for nothing lower than explosive development in computing infrastructure and new types of collaboration between people and machines.
Generative AI has confirmed able to not simply creating new textual content and pictures, however code, designs and even scientific hypotheses. Now, agentic AI is arriving — programs that understand, cause, plan and act autonomously. These brokers behave much less like instruments and extra like digital colleagues, finishing up advanced, multistep duties throughout industries. From authorized analysis to logistics, agentic AI guarantees to speed up productiveness by serving as autonomous digital employees.
Maybe probably the most transformative leap is bodily AI — the embodiment of intelligence in robots of each type. Three computer systems are required to construct bodily AI-embodied robots — NVIDIA DGX GB300 to coach the reasoning vision-language motion mannequin, NVIDIA RTX PRO to simulate, check and validate the mannequin in a digital world constructed on Omniverse, and Jetson Thor to run the reasoning VLA at real-time pace.
What’s anticipated subsequent is a breakthrough second for robotics inside years, with autonomous cell robots, collaborative robots and humanoids disrupting manufacturing, logistics and healthcare. Morgan Stanley estimates there will likely be 1 billion humanoid robots with $5 trillion in income by 2050.
Signaling how deeply AI will embed into the bodily financial system, that’s only a sip of what’s on faucet.
AI is now not only a device. It performs work and stands to rework each one of many world’s $100 trillion in markets. And a virtuous cycle of AI has arrived, basically altering your complete computing stack, transitioning all computer systems into new supercomputing platforms for vastly bigger alternatives.
