Chip manufacturing is an “best utility” for NVIDIA accelerated and AI computing, NVIDIA founder and CEO Jensen Huang stated Tuesday.
Detailing how the most recent developments in computing are accelerating “the world’s most essential business,” Huang spoke at ITF World 2023 semiconductor convention in Antwerp, Belgium.
Huang delivered his remarks by way of video to a gathering of leaders from throughout the semiconductor, expertise and communications industries.
“I’m thrilled to see NVIDIA accelerated computing and AI in service of the world’s chipmaking business,” Huang stated as he detailed how developments in accelerated computing, AI and semiconductor manufacturing intersect.
AI, Accelerated Computing Step Up
The exponential efficiency enhance of the CPU has been the governing dynamic of the expertise business for practically 4 a long time, Huang stated.
However over the previous few years CPU design has matured, he stated. The speed at which semiconductors turn out to be extra highly effective and environment friendly is slowing, whilst demand for computing functionality soars.
“Consequently, world demand for cloud computing is inflicting knowledge middle energy consumption to skyrocket,” Huang stated.
Huang stated that striving for web zero whereas supporting the “invaluable advantages” of extra computing energy requires a brand new strategy.
The problem is a pure match for NVIDIA, which pioneered accelerated computing, coupling the parallel processing capabilities of GPUs with CPUs.
This acceleration, in flip, sparked the AI revolution. A decade in the past, deep studying researchers similar to Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton found that GPUs could possibly be cost-effective supercomputers.
Since then, NVIDIA reinvented its computing stack for deep studying, opening up “multi trillion-dollar alternatives in robotics, autonomous automobiles and manufacturing,” Huang stated.
By offloading and accelerating compute-intensive algorithms, NVIDIA routinely quickens functions by 10-100x whereas lowering energy and value by an order of magnitude, Huang defined.
Collectively, AI and accelerated computing are remodeling the expertise business. “We’re experiencing two simultaneous platform transitions — accelerated computing and generative AI,” Huang stated.
AI, Accelerated Computing Come to Chip Manufacturing
Huang defined that superior chip manufacturing requires over 1,000 steps, producing options the scale of a biomolecule. Every step have to be practically good to yield purposeful output.
“Refined computational sciences are carried out at each stage to compute the options to be patterned and to do defect detection for in-line course of management,” Huang stated. “Chip manufacturing is a perfect utility for NVIDIA accelerated and AI computing.”
Huang outlined a number of examples of how NVIDIA GPUs have gotten more and more integral to chip manufacturing.
Firms like D2S, IMS Nanofabrication, and NuFlare construct masks writers — machines that create photomasks, stencils that switch patterns onto wafers — utilizing electron beams. NVIDIA GPUs speed up the computationally demanding duties of sample rendering and masks course of correction for these masks writers.
Semiconductor producer TSMC and gear suppliers KLA and Lasertech use excessive ultraviolet mild, often known as EUV, and deep ultraviolet mild, or DUV, for masks inspection. NVIDIA GPUs play a vital function right here, too, in processing classical physics modeling and deep studying to generate artificial reference photos and detect defects.
KLA, Utilized Supplies, and Hitachi Excessive-Tech use NVIDIA GPUs of their e-beam and optical wafer inspection and assessment techniques.
And in March, NVIDIA introduced that it’s working with TSMC, ASML and Synopsys to speed up computational lithography.
Computational lithography simulates Maxwell’s equations of sunshine conduct passing by way of optics and interacting with photoresists, Huang defined.
Computational lithography is the most important computational workload in chip design and manufacturing, consuming tens of billions of CPU hours yearly. Large knowledge facilities run 24/7 to create reticles for brand new chips.
Launched in March, NVIDIA cuLitho is a software program library with optimized instruments and algorithms for GPU-accelerated computational lithography.
“We have now already accelerated the processing by 50 occasions,” Huang stated. “Tens of hundreds of CPU servers may be changed by a couple of hundred NVIDIA DGX techniques, lowering energy and value by an order of magnitude.”
The financial savings will scale back carbon emissions or allow new algorithms to push past 2 nanometers, Huang stated.
What’s Subsequent?
What’s the subsequent wave of AI? Huang described a brand new type of AI — “embodied AI,” or clever techniques that may perceive, cause about and work together with the bodily world.
He stated examples embrace robotics, autonomous automobiles and even chatbots which are smarter as a result of they perceive the bodily world.
Huang supplied his viewers a take a look at NVIDIA VIMA, a multimodal embodied AI. VIMA, Huang stated, can carry out duties from visible textual content prompts, similar to “rearranging objects to match this scene.”
It may possibly be taught ideas and act accordingly, similar to “It is a widget,” “That’s a factor” after which “Put this widget in that factor.” It may possibly additionally be taught from demonstrations and keep inside specified boundaries, Huang stated.
VIMA runs on NVIDIA AI, and its digital twin runs in NVIDIA Omniverse, a 3D growth and simulation platform. Huang stated that physics-informed AI may be taught to emulate physics and make predictions that obey bodily legal guidelines.
Researchers are constructing techniques that mesh data from actual and digital worlds on an enormous scale.
NVIDIA is constructing a digital twin of our planet, known as Earth-2, which is able to first predict the climate, then long-range climate, and ultimately local weather. NVIDIA’s Earth-2 staff has created FourCastNet, a physics-AI mannequin that emulates world climate patterns 50-100,000x sooner.
FourCastNet runs on NVIDIA AI, and the Earth-2 digital twin is inbuilt NVIDIA Omniverse.
Such techniques promise to handle the best problem of our time, similar to the necessity for reasonable, clear power.
For instance, researchers on the U.Ok.’s Atomic Power Authority and the College of Manchester are making a digital twin of their fusion reactor, utilizing physics-AI to emulate plasma physics and robotics to manage the reactions and maintain the burning plasma.
Huang stated scientists may discover hypotheses by testing them within the digital twin earlier than activating the bodily reactor, enhancing power yield, predictive upkeep and lowering downtime. “The reactor plasma physics-AI runs on NVIDIA AI, and its digital twin runs in NVIDIA Omniverse,“ Huang stated.
Such techniques maintain promise for additional developments within the semiconductor business. “I stay up for physics-AI, robotics and Omniverse-based digital twins serving to to advance the way forward for chip manufacturing,” Huang stated.