Sixty-two people in San Jose write the software that decides where every transistor on your phone ends up. They have been doing it, quietly, since 2007.
It is a Tuesday on Moorpark Avenue. Inside a low office building you would never photograph, a few dozen engineers are arguing about pixels. Not the kind you scroll past - the kind that decide whether a 3-nanometer transistor turns on. D2S writes the software that converts a chip designer's intent into the curlicued, wavy, faintly alien patterns etched onto photomasks. Those masks are then beamed onto silicon by lithography machines the size of school buses. If any part of that chain stutters, your phone gets slower, your laptop gets hotter, and your AI bill goes up.
D2S is what the semiconductor industry calls infrastructure software. Which is a polite way of saying: nobody outside the industry has heard of them, and everybody inside it depends on them.
Photolithography is a stubborn kind of magic. To print a feature smaller than the wavelength of the light printing it, you cannot just draw the shape you want. You have to draw the shape that, after diffraction, after scattering, after the light bends around its own intent, will land as the shape you want. The drawn shape stops looking like a chip. It looks like a hairball.
For years, the industry tried to keep mask shapes rectangular - "Manhattan" geometry, all right angles, all polite. It worked until it didn't. Below 20 nanometers the right angles stopped doing their job. Engineers needed curves. Curves print better. Curves are also a computational nightmare. A modern chip has billions of features. Each curve has to be simulated, corrected, written, and inspected.
Two things were true at once: curvilinear masks were necessary, and curvilinear masks were impossible to compute at scale. Which is the kind of contradiction that, if you are an engineer, you find irresistible.
Fujimura, a Cadence and Tangent Systems veteran, made a bet in 2007 that looked premature at the time and obvious in retrospect: graphics processors, then known mostly for shooting aliens, would eat the world of physical simulation. He had also, years earlier, coined the phrase "Inverse Lithography Technology" - the idea that you should compute the mask backwards from the wafer you want, not forwards from the design. ILT is now the standard term. In 2007 it was a curiosity.
The bet was simple to state and brutal to execute. Use GPUs to make ILT and model-based mask correction fast enough to be practical. Then build everything around it: the e-beam simulation, the mask data preparation, the deep-learning models, the appliance you sell to mask shops so they don't have to assemble their own. D2S would be the GPU shop for an industry that, at the time, still ran most of its serious math on CPU farms.
1. GPUs were not toys. 2. Curvilinear masks were inevitable. 3. The industry would collaborate if someone organized the room. The first two were technical bets. The third turned out to be the hardest.
The D2S product line reads like a startup that finally got around to its naming meeting. The TrueMask family is, fundamentally, one idea expressed seven ways: take the physics of e-beam writing, simulate it on a GPU, and use the result to fix the mask before it goes near a real mask writer. Then, when the mask writer is done, simulate the wafer too.
Model-based mask data preparation for Manhattan, curvilinear, and ILT shapes - at write times that don't bankrupt the fab.
Full-chip inverse lithography. Designs curvilinear mask features that print sharper at the wafer.
Mask process correction and dose simulation for the e-beam mask writer itself.
Deep-learning mask process correction. Faster than the old model-based loops, surprisingly close in accuracy.
Wafer plane analysis - the loop that ties mask geometry back to wafer reality.
The GPU-accelerated physical model the whole TrueMask family quietly imports.
A turnkey GPU+CPU computational design platform. Mask shops plug it in. 50+ have.
You can tell a lot about a quiet company by who refuses to ignore it. D2S's investors include Cadence and Advantest. Its partners include NuFlare and JEOL, the two dominant mask-writer manufacturers. Its consortium - the eBeam Initiative - has, at various times, included companies that compete with each other for the same fab business. They keep showing up.
NuFlare. JEOL. Advantest. Cadence. Plus the dozens of mask shops, foundries, and IDMs that quietly run TrueMask in production. D2S also runs the eBeam Initiative out of its own offices - 40+ members and counting.
If you cornered Aki Fujimura at a SPIE conference and asked him what D2S is for, you would probably get something like this: use GPUs, simulation, and deep learning to make advanced mask manufacturing practical. Practical is the load-bearing word. Photomasks have always been technically possible. The trick has always been making them in time, on budget, and with enough accuracy to print billions of features correctly.
This is why D2S spends time on a deep-learning consortium - CDLe - and on a vendor-agnostic industry group - the eBeam Initiative - that on paper it has no reason to run. The mission isn't just to sell software. It's to keep the road open for the entire industry to move to curvilinear, inverse-lithography masks. D2S happens to make excellent tolls for that road.
Every node shrink makes the mask weirder. Multi-beam writers - JEOL and NuFlare's latest generation - finally make full-chip curvilinear writing affordable. But weird masks need weird software. Inverse lithography is moving from luxury to default. Deep learning is moving from research demo to production loop. Both trends compound. Both run on GPUs. Both, conveniently, are what D2S has been doing since 2007.
If AI workloads keep eating fabs - and there is no current evidence that they will stop - the photomask is one of a small number of physical objects whose precision determines how fast the model in your pocket can think. D2S sits closer to that precision than anyone wearing a hoodie at a foundry.
The pixels are still being argued about. The mask writer down the supply chain still hums on a swing shift in Tokyo, in Hsinchu, in Albany. The chip that will end up in next year's phone is, somewhere in this process, becoming the wavy alien hairball that will, in turn, become the polite straight transistor you will never see. The 62 people in San Jose go home. They come back the next morning. They open another file. There are more pixels.
D2S is not famous. That is fine. Photomasks are not famous either, and they are how every modern chip gets made.