Taiwan Semiconductor Manufacturing Co. (TSMC) is overhauling a core part of its chipmaking process. For its next-generation 2nm chips, the foundry giant will adopt curvilinear masks, abandoning the straight-lined “Manhattan” geometry used for decades.
This change allows for more precise patterns to be printed on silicon, boosting chip performance.
This leap is enabled by new multi-beam mask writers and advanced software like Nvidia’s GPU-powered CuLitho platform.
While costly, the investment is fueled by the booming AI market, where high-margin chips from customers like Nvidia justify pioneering these complex new manufacturing techniques.
From Manhattan Grids to Curvy Lines: A New Geometry for Chipmaking
For the first time in over a decade, the fundamental geometry of chip design is being redrawn.
TSMC’s 2nm (N2) process node will be its first to use curvilinear masks, a significant departure from the industry’s reliance on rectilinear, or straight-lined, patterns often called “Manhattan” geometry.
Curvilinear masks in semiconductor manufacturing are photomasks whose patterns include curves or free-form shapes rather than being constrained to conventional rectilinear (straight-edged) shapes aligned only vertically or horizontally.
Specifically, curvilinear masks can contain edges that are neither Manhattan (strictly 90-degree or 45-degree angles) nor straight lines, but instead can include smooth arcs, circles, ovals, splines, or piecewise linear polygons with angled edges beyond standard orthogonal geometries.
These masks are designed using advanced optical proximity correction (OPC) and inverse lithography techniques (ILT) that optimize the photomask shapes as curves instead of approximating them with many small rectangles or Manhattan polygons.
The use of curvilinear shapes allows better fidelity in printing complex and small features on silicon wafers, yielding larger lithographic process windows, improved depth of focus, and reduced process variation.
This shift accompanies the move to Gate-All-Around (GAA) transistors, marking one of the most substantial technological transitions in semiconductor fabrication in nearly 15 years.
From Academic Photolithography Concept to Production
Photolithography, the process of printing chip designs onto silicon wafers, is governed by the physics of light.
Light naturally diffracts and distorts, and it does not favor sharp, 90-degree angles. Curvilinear designs, which use smooth curves, more accurately model how light behaves, resulting in a more faithful transfer of the intended pattern onto the wafer.
This widens the overall process window, making manufacturing more resilient to minor deviations and ultimately improving chip yield and performance.
For years, engineers have known that curved designs are theoretically superior. Using a technique called Inverse Lithography Technology (ILT), they could work backward from the desired pattern on the wafer to calculate the optimal, often psychedelic-looking, mask design.
However, this ideal remained an academic concept because the tools to create these masks didn’t exist.
Traditional mask-writers, known as Variable Shaped Beam (VSB) systems, can only produce rectangles and squares. To create a curve, they had to approximate it with thousands of tiny, overlapping rectangles in a process called “Manhattanization.”
This conversion was not only imprecise, creating fuzzy edges, but also incredibly slow.
A VSB machine writes each rectangle with a single “shot” of its electron beam. The sheer number of shots required for complex, Manhattanized patterns created a severe throughput bottleneck, with mask write times stretching from hours into days.
The Tech Behind the Curves: Multi-Beam Writers and GPU-Powered Physics
Powering this leap in manufacturing precision is a convergence of breakthroughs in hardware and software. The first critical enabler is the rise of multi-beam mask writers, developed by firms like IMS Nanofabrication and NuFlare.
Instead of a single electron beam, these machines split the beam into hundreds of thousands of tiny, individually controlled “beam-lets.”
By moving the mask stage and turning these beam-lets on and off like pixels on a screen, the machine can effectively “paint” complex, curved patterns with high fidelity.
Bringing this technology to market was a monumental engineering challenge. Developers had to solve tricky problems like identifying and capturing defects in complex curved patterns and streaming the massive amounts of design data to the machine at high speed.
The development costs were immense; KLA-Tencor, for instance, spent over $226 million on a multi-beam project before abandoning it in 2014. Success required a decade of persistence and deep investment to overcome these hurdles.
The second piece of the puzzle is a massive increase in computational power, driven by the GPU revolution.
Calculating an ILT mask design for a modern chip with billions of transistors is an immense task, sometimes requiring up to 30 million CPU hours. A data center with tens of thousands of CPUs could take over a week to complete the job.
Nvidia’s cuLitho, a software library of parallel algorithms, dramatically changes this equation. According to Nvidia, 500 of its H100 GPUs can now perform the computational work of 40,000 CPUs for these tasks.
This accelerates workflows by up to 60 times, turning a two-week calculation into an overnight process.
Recognizing this potential, TSMC, Nvidia, and the design software firm Synopsys announced in early 2024 they were moving the CuLitho platform into production, paving the way for the N2 node’s adoption of curvy masks.
Why Now? The AI Boom Pays for a Manufacturing Revolution
Driving the massive investment required for this transition is the insatiable and high-margin demand from the artificial intelligence market.
Chips designed for AI accelerators, like those from Nvidia and AMD, must deliver the absolute highest levels of performance. Dr. Lisa Su, Chair and CEO of AMD, has previously highlighted the company’s deep collaboration with TSMC, which “has enabled AMD to consistently deliver leadership products that push the limits of high-performance computing.”
For these key customers, the benefits of the 2nm node and its curvilinear patterns are direct and substantial. F
or Nvidia, it means more powerful and energy-efficient GPUs to dominate the data center. For a customer like Apple, it translates to longer battery life and faster processing for future generations of iPhone and Mac silicon.
Unlike the mature and price-sensitive mobile phone market, the AI sector has the financial margins to absorb the high costs of pioneering these advanced manufacturing technologies.
This dynamic justifies the multi-billion dollar investments in new mask writers and the extensive R&D needed to bring curvilinear lithography to high-volume production.
The focus on next-generation technology is central to TSMC’s strategy of maintaining its leadership position.
The company has consistently denied rumors of operational mergers with Intel or others, with CEO C.C. Wei stating firmly, “TSMC is not engaged in any discussion with other companies regarding any joint venture, technology licensing or technology.”
Instead, the foundry is pushing forward on multiple fronts, including the development of advanced panel-level packaging to meet future AI demands.
The adoption of curvilinear masks is more than an incremental update; it’s a foundational shift in manufacturing, paid for by the AI boom, that will redefine the limits of chip design for the next decade.