AI Wiki

Explore AI Wiki

Citation and evidence

NVIDIA cuLitho

15 min full readUpdated 14 references

This article's verification

Report a problem with this article

More

Use this article

Raw MarkdownExplore connections

Improve this page

Suggest editRevision history

Topics

AI ComputeAI HardwareNVIDIA

Cite this article

NVIDIA cuLitho is a software library from NVIDIA that runs computational lithography on GPUs instead of CPUs. Computational lithography is the simulation and optimization work that turns a chip layout into the patterns written on photomasks, and it includes optical proximity correction (OPC) and inverse lithography technology (ILT). NVIDIA introduced cuLitho at its GTC conference on March 21, 2023, together with TSMC, ASML and Synopsys, and said it could speed up computational lithography by up to 40 times.[1] TSMC said in 2024 that it was moving cuLitho into production,[3][4] Synopsys integrated it into its Proteus OPC software,[3][5] Samsung reported a 20x gain in its OPC process in 2025,[6][7] and in October 2026 Siemens announced that it had integrated cuLitho with its Calibre manufacturing software, claiming up to 12x faster full-chip OPC.[10] Siemens describes cuLitho as "a CUDA-X library for GPU-accelerated computational lithography", and NVIDIA includes it in its October 2026 CUDA-X chipmaking announcement.[10][12] The speedup figures for cuLitho come from NVIDIA or its partners, and they measure different parts of the workflow against different CPU baselines.

Background: why lithography needs so much computing

Chips are patterned by shining light through a photomask onto a light-sensitive resist on a silicon wafer. Each layer of a chip needs its own mask, and a full set for an advanced chip is large: at GTC 2023, Jensen Huang said the NVIDIA H100 used 89 masks.[14] Because modern features are smaller than the wavelength of the light used to print them, the shapes drawn by designers cannot be copied directly onto the mask. In the words of a Siemens Calibre blog post, "Diffraction and other physical effects distort the image."[11]

Optical proximity correction (OPC) compensates for this distortion by deliberately changing the edge placements of the patterns on the mask so that the printed result on the wafer matches the intended design.[11] An NVIDIA and imec paper on cuLitho describes the field as having "evolved from simple rules to physics-guided decorations", and model-based OPC works by segmenting the layout and iteratively moving its edges according to a simulation model.[8]

Inverse lithography technology (ILT) goes further. Siemens explains that instead of starting from a conventional mask shape and making local edge corrections, ILT "gives the optimizer far more freedom to determine the mask geometry needed to produce the desired wafer pattern", which can produce complex curvilinear mask shapes.[11] NVIDIA distinguishes curvilinear masks, whose shapes are represented by curves, from "Manhattan" masks, whose shapes are constrained to horizontal and vertical edges.[3] In a 2024 post, NVIDIA's Vivek Singh wrote that inverse lithography had been described in the scientific literature for two decades but that "an accurate realization at full chip scale has been largely precluded because the computation takes too long."[4]

The cost of this work grows with each process node. According to NVIDIA, computational lithography is "the most compute-intensive workload in the semiconductor manufacturing process", consuming tens of billions of CPU hours a year, and a typical mask set "could take 30 million or more hours of CPU compute time", which requires large data centers inside foundries.[3] NVIDIA's 2023 announcement also noted that a change in fab process often requires an OPC revision, which creates bottlenecks.[1] The NVIDIA and imec paper attributes the growth to rising mask complexity (tens of billions of transistors and trillions of layout polygons per chip), stricter process requirements as features approach or fall below the illumination wavelength, and the need for more accurate optical and resist models.[8] The move to High Numerical Aperture extreme ultraviolet (High-NA EUV) lithography adds further modeling burden, according to Siemens.[11]

How cuLitho works

NVIDIA's developer page describes cuLitho as "a library with optimized tools and algorithms for GPU-accelerating computational lithography."[2] Foundries and electronic design automation (EDA) vendors integrate it into their own software and workflows, including Synopsys Proteus, TSMC's internal workflows, Samsung's OPC platform and Siemens Calibre.[3][6][10]

The 2026 paper by NVIDIA and imec researchers (including cuLitho architect Vivek Singh) gives the most detailed public account of its design.[8] It says cuLitho "reinvents the core primitives" of computational lithography (diffractive optics, computational geometry, multi-variant optimization and data processing) and accelerates these building blocks by between 50x and 1,500x.[8] The authors argue that end-to-end speed is limited by Amdahl's law: earlier GPU efforts accelerated only the imaging stages, which account for roughly half of OPC run time, so "achieving speedups on the order of 50X requires porting essentially all major modules to the GPU." They also stress keeping data resident in GPU memory, because the host-to-device link is far slower than GPU memory bandwidth.[8]

Generative AI

At GTC 2024, NVIDIA added generative AI algorithms to the platform. According to NVIDIA, the AI model creates "a near-perfect inverse mask or inverse solution to account for diffraction of light", after which the final mask is derived by "traditional and physically rigorous methods". NVIDIA said this delivered an additional 2x speedup in the overall OPC process on top of cuLitho's GPU acceleration.[3][4] In the 2026 paper, the corresponding AI step is described as a "diffractive-optics-informed generative-AI model" that cuts the number of iterations by half; in that study it provided about 1.7x of additional speedup without loss of edge placement accuracy.[8]

History

DateEventSource
March 21, 2023NVIDIA announces cuLitho at GTC 2023, with TSMC and Synopsys integrating it and ASML planning GPU support across its computational lithography software[1]
March 18, 2024At GTC 2024, NVIDIA says TSMC and Synopsys are going into production with cuLitho, and introduces generative AI algorithms for the platform[3]
October 8, 2024NVIDIA blog post says cuLitho is moving to production at TSMC[4]
March 18, 2025Synopsys says Proteus, integrated with cuLitho on H100 GPUs, achieves a 15x OPC speedup and projects up to 20x on Blackwell[5]
October 30, 2025NVIDIA and Samsung announce a Samsung AI factory; Samsung reports a 20x gain in computational lithography from cuLitho in its OPC process[6][7]
2026NVIDIA and imec researchers publish silicon results for cuLitho-based OPC on imec's 3 nm-node EUV designs[8]
May 31, 2026At GTC Taipei, NVIDIA says TSMC uses cuLitho with a 20-50% improvement in cost effectiveness or cycle time[9]
October 14, 2026Siemens announces integration of cuLitho with Calibre, claiming up to 12x faster full-chip OPC; NVIDIA's announcement the same day lists Siemens among CUDA-X adopters[10][12]

Expanded article table

Launch at GTC 2023

NVIDIA announced cuLitho on March 21, 2023. It said TSMC and Synopsys were integrating the library into their software, manufacturing processes and systems for NVIDIA Hopper architecture GPUs, and that ASML was working with NVIDIA on GPUs and cuLitho.[1] NVIDIA made several performance claims in the announcement:[1]

  • cuLitho "delivers a performance leap of up to 40x beyond current lithography."
  • It "enables 500 NVIDIA DGX H100 systems to achieve the work of 40,000 CPU systems."
  • In the near term, fabs using it "could help produce each day 3-5x more photomasks" using 9x less power, and "a photomask that required two weeks can now be processed overnight."

NVIDIA's developer page restates these figures as 500 Hopper GPU systems doing the work of 40,000 CPU systems at "1/9 the power and 1/8 the space", and describes the 40x figure as a speedup of inverse lithography.[2] In his keynote, Huang said a single mask took about two weeks to compute on CPUs and could be cut to eight hours with cuLitho, and that TSMC could reduce the 40,000 CPU servers it used for computational lithography to 500 DGX H100 systems, cutting power from 35 MW to 5 MW, according to EE Times Japan's report of the keynote.[14] EE Times Japan also reported that TSMC planned to begin using cuLitho from June 2023.[14]

Partner executives were quoted in the announcement. TSMC CEO C.C. Wei said the work "opens up new possibilities for TSMC to deploy lithography solutions like inverse lithography technology and deep learning more broadly in chip manufacturing." ASML CEO Peter Wennink said ASML was "planning to integrate support for GPUs into all of our computational lithography software products" and that the benefit would be "especially true in the era of high NA extreme ultraviolet lithography." Synopsys chair and CEO Aart de Geus said running Synopsys OPC software on cuLitho "massively accelerated the performance from weeks to days."[1]

Production at TSMC and generative AI (2024)

At GTC on March 18, 2024, NVIDIA announced that TSMC and Synopsys were going into production with the cuLitho platform and would "in the future support" NVIDIA Blackwell GPUs. The release's subheadline said cuLitho accelerates the workload "by 40-60x". NVIDIA said that "350 NVIDIA H100 systems can now replace 40,000 CPU systems," a lower system count than the 500 DGX H100 systems cited in 2023.[3] In tests on shared workflows, TSMC and NVIDIA "jointly realized a 45x speedup of curvilinear flows and a nearly 60x improvement on more traditional Manhattan-style flows."[3] C.C. Wei said, "We are moving NVIDIA cuLitho into production at TSMC."[3] The same release introduced the generative AI workflow described above.[3]

On October 8, 2024, Vivek Singh published an NVIDIA blog post stating that cuLitho "is moving to production at TSMC", repeating the 350 H100 systems figure and the 2x generative AI speedup.[4]

Synopsys Proteus

Synopsys Proteus is the company's mask synthesis and OPC product line. In the March 2024 announcement, Synopsys CEO Sassine Ghazi said Proteus products had been "the production-proven choice for accelerating computational lithography" for more than two decades, and NVIDIA said Proteus OPC software running on cuLitho "significantly speeds computational workloads compared to current CPU-based methods."[3] In a March 18, 2025 press release, Synopsys said that "Today, Synopsys Proteus is optimized for NVIDIA H100 GPUs and integrated with the NVIDIA cuLitho library, achieving a 15x speed up of OPC," and that using the Blackwell platform it was "expected to accelerate computational lithography simulations by up to 20x" compared with CPU-based models.[5] The 20x Blackwell figure was a projection.[5]

Samsung (2025)

On October 30, 2025, NVIDIA announced plans with Samsung Electronics to build an AI factory powered by more than 50,000 NVIDIA GPUs. As part of that announcement, the companies said they were integrating cuLitho into Samsung's OPC platform and that the collaboration had "led to 20x greater performance."[6] Samsung's own release said that by "leveraging NVIDIA cuLitho and CUDA-X libraries for its optical proximity correction (OPC) process, Samsung has achieved a 20x gain in computational lithography performance."[7]

imec silicon study (2026)

A paper by researchers from NVIDIA and imec, "Transforming Computational Lithography with AC and AI -- Faster, More Accurate, and Energy-efficient" (arXiv:2602.15036), reported what it called "the first quantified chip-scale demonstration of the lithography benefits of AC and AI in silicon."[8] The study corrected 25 layouts of 28-36 nm pitch metal layers in imec's 3 nm node, using single-exposure low-NA (0.33) EUV patterning modeled for an ASML NXE:3400B scanner, a total of 1.6 billion polygons over 22 mm2. The correction flow used the cuLitho-accelerated Proteus OPC engine from Synopsys, run on NVIDIA H200 GPUs.[8]

Metric (as reported in the paper)Result
End-to-end speedup over production CPU flows57x (34x with GPU acceleration alone, 57x with the AI step)
Speedup for flows including through-focus correction24x (up from 14x without AI)
Depth of focus in silicon59 nm to 80 nm (35% better process window)
Through-focus wafer CD error in silicon3.7 nm to 3.0 nm (19% better)
Silicon measurements1.5 million
Modeled datacenter for 30 full-reticle layers per day353 H200 HGX systems in 118 racks vs 40,690 CPU servers in 1,403 racks
Modeled power37.6 MW (CPU) vs 2.9 MW (GPU)

Expanded article table

The authors reinvested part of the freed-up compute in through-focus correction, which they say is "prohibitively expensive on CPUs"; the CPU baseline runs used only on-focus correction. The datacenter comparison assumed a CPU implementation on AMD Milan processors.[8] In its October 2026 semiconductor announcement, NVIDIA described imec's use of cuLitho for process-window OPC as achieving "a 30% improvement in the lithography process window", a different figure from the 35% in the paper.[12]

TSMC at GTC Taipei (2026)

On May 31, 2026, at GTC Taipei, NVIDIA announced that TSMC was using NVIDIA accelerated computing across lithography, transistor and process simulation, process control and fab operations. On lithography, NVIDIA said cuLitho "delivers a 20-50% improvement in cost effectiveness or cycle time compared with CPU-based computational lithography, while maintaining the same cost of ownership."[9] This framing (a percentage improvement in cost or cycle time in production) differs from the "up to 40x" figure NVIDIA gave at launch in 2023 and the 45x to nearly 60x speedups it reported from TSMC tests in 2024.[1][3]

Siemens Calibre (2026)

On October 14, 2026, Siemens Digital Industries Software announced from Plano, Texas, that it had integrated cuLitho with Calibre, its manufacturing software, describing the work as "significant advances in its collaboration with NVIDIA."[10] Siemens made these claims:[10]

  • Up to 12x acceleration for full-chip OPC workloads.
  • "At least 10x acceleration for key lithography simulation workloads."
  • "Verified engineering data supports simulation speed-ups ranging from at least 10x to 17x for key workloads."
  • The integration enables practical full-chip optimization of photomasks for High-NA EUV lithography, including ILT.

Siemens did not name customers or publish the hardware configuration or CPU baseline behind these figures in the press release. Ankur Gupta, executive vice president of the EDA IC portfolio at Siemens, said the combination gives customers "the ability to deploy advanced patterning strategies like curvilinear ILT at scale." Tim Costa, NVIDIA's vice president and general manager of industrial and computational engineering, said it would let engineers "tackle increasingly intricate patterns and explore more extensive optimizations."[10] A companion post by Jim Shiely, chief of staff for Calibre semiconductor R&D at Siemens EDA, said that for photomask makers and chip builders "turnaround times shrink from weeks to days."[11]

NVIDIA's own announcement on the same day, covering CUDA-X adoption across chipmaking and inviting readers to join NVIDIA at SEMICON West, said "Siemens is using cuLitho to speed up Calibre software for chip-pattern refinement and mask preparation."[12] eeNews Europe reported the Siemens figures on October 16, 2026.[13] With the Siemens announcement, cuLitho has been integrated into OPC software from both Synopsys and Siemens, in addition to its use by TSMC and Samsung.[3][6][10]

Reported performance claims

Every figure below was published by NVIDIA or a partner (the imec paper is co-authored by NVIDIA researchers). The figures measure different things (individual workloads, full OPC flows, system counts, cost or cycle time) against different CPU baselines, so they are not directly comparable.

ClaimantDateClaimSource
NVIDIAMarch 2023Up to 40x over current lithography; 500 DGX H100 systems do the work of 40,000 CPU systems[1]
Jensen Huang (keynote)March 2023Per-mask time from about two weeks to eight hours; TSMC power from 35 MW to 5 MW[14]
NVIDIAMarch 202440-60x; 350 H100 systems replace 40,000 CPU systems; 45x curvilinear and nearly 60x Manhattan flows in TSMC tests[3]
NVIDIAMarch 2024Generative AI adds a further 2x to OPC[3]
SynopsysMarch 202515x OPC speedup with Proteus on H100 and cuLitho; up to 20x projected on Blackwell[5]
Samsung / NVIDIAOctober 202520x gain in computational lithography performance[6][7]
NVIDIA and imec researchers202657x end-to-end on 25 imec EUV designs; 35% better process window in silicon[8]
NVIDIAMay 202620-50% improvement in cost effectiveness or cycle time at TSMC[9]
SiemensOctober 2026Up to 12x for full-chip OPC; at least 10x to 17x for key simulation workloads[10]

Expanded article table

See also

References

  1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8NVIDIA. "NVIDIA, ASML, TSMC and Synopsys Set Foundation for Next-Generation Chip Manufacturing." NVIDIA Newsroom, March 21, 2023. nvidianews.nvidia.com/...ration-chip-manufacturing
  2. ^1 ^2NVIDIA. "NVIDIA cuLitho: Accelerate computational lithography." NVIDIA Developer. developer.nvidia.com/culitho
  3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16NVIDIA. "TSMC and Synopsys Bring Breakthrough NVIDIA Computational Lithography Platform to Production." NVIDIA Newsroom, March 18, 2024. nvidianews.nvidia.com/...c-synopsys-nvidia-culitho
  4. ^1 ^2 ^3 ^4 ^5Singh, Vivek. "TSMC and NVIDIA Transform Semiconductor Manufacturing With Accelerated Computing." NVIDIA Blog, October 8, 2024. blogs.nvidia.com/...itho-computational-lithography
  5. ^1 ^2 ^3 ^4 ^5Synopsys. "Synopsys Accelerates Chip Design with NVIDIA Grace Blackwell and AI to Speed Electronic Design Automation." Synopsys News, March 18, 2025. news.synopsys.com/...-Electronic-Design-Automation
  6. ^1 ^2 ^3 ^4 ^5 ^6NVIDIA. "NVIDIA and Samsung Build AI Factory to Transform Global Intelligent Manufacturing." NVIDIA Newsroom, October 30, 2025. nvidianews.nvidia.com/...samsung-ai-factory
  7. ^1 ^2 ^3 ^4Samsung. "Samsung Teams with NVIDIA to Lead the Transformation of Global Intelligent Manufacturing through New AI Megafactory." Samsung Newsroom Singapore. news.samsung.com/...ing-through-new-ai-megafactory
  8. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Mukhopadhyay, Saumyadip, et al. "Transforming Computational Lithography with AC and AI -- Faster, More Accurate, and Energy-efficient." arXiv:2602.15036, 2026. arxiv.org/...2602.15036
  9. ^1 ^2 ^3NVIDIA. "NVIDIA and TSMC Bring AI Into Fabs to Advance Semiconductor Design and Manufacturing." NVIDIA Newsroom, May 31, 2026. nvidianews.nvidia.com/...-design-and-manufacturing
  10. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Siemens. "Siemens transforms computational lithography for next-generation AI semiconductors with NVIDIA cuLitho." Siemens Newsroom, October 14, 2026. news.siemens.com/...siemens-calibre-nvidia-culitho
  11. ^1 ^2 ^3 ^4 ^5Shiely, Jim. "The virtuous cycle of silicon: How Siemens is accelerating the era of Physical AI with NVIDIA cuLitho." Siemens Calibre IC Design & Manufacturing blog, October 14, 2026. blogs.sw.siemens.com/...cal-ai-with-nvidia-culitho
  12. ^1 ^2 ^3 ^4NVIDIA. "NVIDIA and Semiconductor Leaders Advance AI Across Chip Design and Manufacturing for AI-Defined Fabs." NVIDIA Newsroom, October 14, 2026. nvidianews.nvidia.com/...uring-for-ai-defined-labs
  13. ^Adhimi, Asma. "Siemens and NVIDIA speed up AI chip lithography by 12x." eeNews Europe, October 16, 2026. eenewseurope.com/...-up-ai-chip-lithography-by-12x
  14. ^1 ^2 ^3 ^4"計算リソグラフィを40倍高速化するAIライブラリ：マスク作成時間は2週間から8時間に" (AI library that speeds up computational lithography 40x: mask creation time cut from two weeks to eight hours). EE Times Japan, April 4, 2023. eetimes.itmedia.co.jp/...news079_2

Improve this article

Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.

2 revisions · v2 · 3,004 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Reviewer note: Independent verification 2026-10-17 (xg19 V5): Siemens, NVIDIA, Synopsys, imec and EE Times sources checked; 3 minor fixes applied to TSMC timeline and speedup attributions.

Cite this page: AI Wiki. "NVIDIA cuLitho." aiwiki.ai, updated 16 Oct 2026, fact-checked 16 Oct 2026. CC BY 4.0. https://aiwiki.ai/wiki/nvidia_culitho

Suggest edit

Where to go next

Topics:AI ComputeAI HardwareNVIDIA