Profile
Pixel-Perfect at Petascale
Somewhere in a cleanroom right now, a semiconductor manufacturer is writing photomasks that will define transistors smaller than a coronavirus. The shapes on those masks - the templates that transfer circuit patterns onto silicon - are getting curvier. That's not an accident. That's, in large part, Aki Fujimura's doing.
Fujimura founded D2S, Inc. in 2007 with a specific conviction: that electron beam mask writers could be coaxed into producing complex, curvilinear shapes rather than the boxy "Manhattan" geometries that had dominated semiconductor lithography for decades. The industry consensus at the time was that this was impractical. Mask write times would explode. Turnaround would collapse. The economics didn't work.
He disagreed. And he had a tool the previous generation of mask engineers didn't: graphics processing units. GPUs, purpose-built for the kind of massively parallel floating-point arithmetic that rendering video game worlds requires, turned out to be equally suited to simulating a trillion-pixel photomask dose map. D2S built its Computational Design Platform around them - and over forty installations later, the industry has largely caught up with Fujimura's 2007 thesis.
The source of accuracy and precision in semiconductor manufacturing really comes from the mask.
- Aki Fujimura, D2S
To understand why this matters: every advanced chip in production today begins as a design intent that must be translated, faithfully, through a series of physical templates onto a silicon wafer. The photomask is the intermediate step - a glass plate coated with chrome, etched with patterns by an electron beam, then used to expose light onto wafer surfaces in the same way a film negative exposes a photograph. When transistors were large, the shapes could afford to be simple rectangles and right angles. At 2nm, they cannot.
Inverse lithography technology (ILT), the computational process of working backward from a desired wafer outcome to find the ideal mask shape, produces naturally curvilinear patterns. But for years, ILT outputs were "Manhattan-ized" - snapped back to right angles - because eBeam writers couldn't practically write the curves. D2S changed that equation. Its TrueMask ILT platform, combined with its GPU-accelerated CDP, makes entirely curvilinear ILT practical within a production-viable timeframe. Studies comparing curvilinear ILT to conventional Manhattan OPC show an 85% to 100% improvement in process window - meaning the chipmaker has far more tolerance for manufacturing variation.
D2S by the Numbers
100+
US Patents Held
85%
Min Process Window Gain vs. Manhattan OPC
1.8P
FLOPS per Single-Rack CDP (SP)
2007
Year D2S Was Founded
Deep Dive
The GPU Bet Nobody Else Made
What Fujimura saw in 2007 that others didn't was a convergence. NVIDIA was deploying general-purpose GPU computing. Photomask complexity was about to explode with sub-20nm design rules. And eBeam mask writers - the electron beam tools that physically inscribe patterns on mask blanks - were capable of more complexity than the software feeding them could produce. The bottleneck wasn't hardware. It was compute.
D2S built around that insight. Its Computational Design Platform (CDP) combines coarse-grain and fine-grain parallelism to simulate a dose map of one quintillion pixels - that's 10^18 - for a single chip mask. The latest generation uses NVIDIA Ampere A40 GPUs, delivering 1.8 PFLOPS of single-precision compute in a single rack. The result is that simulations which once took weeks can be done overnight. Curvilinear ILT, which was theoretically possible but practically impossible, became both.
D2S's CDP can simulate a 1 quintillion-pixel dose map for a single chip mask - using NVIDIA Ampere A40 GPUs delivering 1.8 PFLOPS in one rack.
The company has also built out partnerships at both ends of the mask writing pipeline. A collaboration with JEOL reduces write times for advanced photomask production. A partnership with NuFlare accelerates CDP deployment. The 2023 acquisition of Gauda expanded D2S's GPU acceleration reach further into the lithography simulation stack.
Fujimura is explicit about where AI fits in. "AI and GPUs are fundamentally changing the way 2nm mask sets will be built in future fabs," he said in a recent interview. That's not hype - it's an observation about physics and compute cost curves that his company has spent 17 years positioning to exploit.
Why It Matters
The Mask Is the Message
There's a reason Fujimura keeps returning to the photomask as the linchpin of semiconductor precision. Advanced chips don't fail at design. They fail at translation - the long chain from circuit intent to silicon reality, where every step compounds error. The mask is where that chain is most brittle.
At 2nm design rules, transistor dimensions are measured in single-digit nanometers. The mask features that define them are larger (roughly 4x, under typical optical projection ratios), but the tolerances are still measured in fractions of a nanometer. A mask that's even slightly wrong - slightly too boxy, slightly misregistered - propagates that error across every wafer it touches. Multiply by millions of dice per wafer run, and the economics of a bad mask are severe.
Curvilinear ILT attacks this problem directly. By computing the mask shape that will produce, after all optical distortions of the lithographic process, the pattern the designer specified, it squeezes every last nanometer of process window out of the physics. The problem has always been compute cost. D2S's CDP reduces that cost to the point where curvilinear ILT becomes production-viable - not a research exercise.
Making a very small contact hole resilient to manufacturing variation is a key. But in wafer and mask manufacturing, you need all of them to look okay. You need the worst-case examples to look okay.
- Aki Fujimura
Fujimura's "Shot Talk" video series - a recurring YouTube format he produces to recap industry conferences - is an unusual outlet for a CEO of a deep-technology company. Most semiconductor executives don't produce direct-to-camera content. That he does, consistently, since 2015, says something about his conviction that the photomask community needs better knowledge infrastructure, not just better tools.
He is a conference institution. At SPIE Photomask Technology, at BACUS, at SEMICON events, Fujimura is a fixture - presenting data, chairing sessions, or anchoring the eBeam Initiative's annual industry survey. The eBeam Initiative, which he co-founded and continues to manage, tracks photomask revenue trends and eBeam adoption in a way that functions as a public good for the whole industry.
