Nvidia hooks TSMC, ASML, Synopsys on GPU accelerated lithography
What's next – AI designing AI chips? Oh wait... that's exactly what's next
GTC Nvidia's latest gambit? Entrenching itself as a key part of the semiconductor manufacturing supply chain.
At GTC this week, the chipmaker unveiled cuLitho, a software library designed to accelerate computational lithography workloads used by the likes of TSMC, ASML, and Synopsys, using its GPUs.
The idea behind the platform is to offload and parallelize the complex and computationally expensive process of generating photomasks used by lithography machines to etch nanoscale features, like transistors or wires, into silicon wafers.
“Each chip design is made up of about 100 layers and in total contains trillions of polygons or patterns. Each of these 100 layers are encoded separately into a photomask — a stencil for the design if you will — and, using a rather expensive camera, are successively printed onto silicon,” Vivek Singh, VP of Nvidia's advanced technology group, explained during a press conference on Monday.
Originally, photomasks were just a negative of the shape engineers were trying to etch into the silicon, but as transistors have gotten smaller these photomasks became more complex to counteract the effects of optical distortion. If unchecked, this distortion can blur these features beyond recognition. This process is called optical proximity correction (OPC) and more recently has evolved into inverse lithography technology (ILT). In the case of the latter, the photomasks look nothing like the feature they're designed to print.
And the more ornate these photomasks get, the more computational horsepower is required to produce them. However, using GPUs, Nvidia believes it can not only speed up this process, but reduce the power consumption required. The company claims that cuLitho running on its GPUs is roughly 40x faster than existing computational lithography platforms running on general purpose CPUs.
“It’ll help the semiconductor industry continue the pace of innovation that we’ve all come to rely on, and it’ll improve the time to market for all kinds of chips in the future,” Singh claimed.
However, at least in the near term, Nvidia's expectations seem to be a little more grounded. The company expects fabs using cuLitho could produce 3-5x more photomasks a day while using 9 percent less power, which if true, should help to boost foundries' already thin margins
And with the likes of ASML, Synopsys, and TSMC lining up to integrate Nvidia's GPUs and libraries into their software platforms and fabs, we won't have to wait long to see these claims put to the test.
TSMC is already investigating Nvidia's GPUs and cuLitho to accelerate ILT photomasks, while ASML and Synopsys are working to integrate support for GPU acceleration using cuLitho in their computational lithography software platforms.
And while Nvidia execs would love to sell its latest and most expensive GPU architectures to these companies, Singh notes that the library is compatible with GPUs going back to the Volta generation, which made its debut in 2017.
- Semiconductor world in for a rough ride as chip bubble bursts at the high end
- Unless things change, first zettaflop systems will need nuclear power, AMD's Su says
- Nvidia in blast radius as Uncle Sam looks to cut off China's Huawei for good
- Cloud upstart offers free heat if you host its edge servers
While Nvidia is using GPUs to accelerate these workloads, it's worth noting that cuLitho isn't using machine learning or AI to optimize semiconductor design just yet. But it's no secret that Nvidia is also working on that particular problem.
“Much of this has to do with accelerating the underlying primitive operations of computational lithography,” Singh said. “But I will say that AI is very much in the works in cuLitho.”
As our sister site The Next Platform reported last summer, Nvidia has been working on ways to accelerate computational lithography workloads for some time now. In a research paper published in July, engineers at the company used AI to design equivalent circuits 25 percent smaller than those created using traditional EDA platforms.
Nvidia is hardly the only company investigating the use of machine learning to accelerate circuit design. Synopsys and Cadence have both implemented AI technologies into their portfolios, while Google researchers developed a deep-learning model called PRIME to create smaller and faster accelerator designs. And previously, the company used reinforcement learning models to design portions of its tensor processing unit (TPU).
With that said, the addressable market for something like cuLitho isn't that big, and thanks to efforts by the US Commerce Department to stifle China's fledgling semiconductor industry, the number is only getting smaller.
cuLitho will almost certainly be subject to US export controls governing the sale of advanced semiconductor manufacturing equipment and software to countries of concern, which for the moment means China. Pressed on this point, Singh said the library would be "available wherever this end-to-end OPC software is available," but declined to comment further on US trade restrictions. ®