Nvidia hooks TMSC, ASML, and Synopsys on GPU accelerated lithography

GTC Nvidia’s newest gambit? Entrenching itself as a key a part of the semiconductor manufacturing provide chain.

At GTC this week, the chipmaker unveiled cuLitho, a software program library designed to speed up computational lithography workloads utilized by the likes of TSMC, ASML, and Synopsys, utilizing its GPUs.

The thought behind the platform is to dump and parallelize the complicated and computationally costly technique of producing photomasks utilized by lithography machines to etch nanoscale options, like transistors or wires, into silicon wafers.

“Every chip design is made up of about 100 layers and in complete accommodates trillions of polygons or patterns. Every of those 100 layers are encoded individually right into a photomask — a stencil for the design if you’ll — and, utilizing a moderately costly digital camera, are successively printed onto silicon,” Vivek Singh, VP of Nvidia’s superior know-how group, defined throughout a press convention on Monday.

Initially, photomasks have been only a damaging of the form engineers have been making an attempt to etch into the silicon, however as transistors have reduced in size these photomasks grew to become extra complicated to counteract the results of optical distortion. If unchecked, this distortion can blur these options past recognition. This course of is known as optical proximity correction (OPC) and extra just lately has advanced into inverse lithography know-how (ILT). Within the case of the latter, the photomasks look nothing just like the characteristic they’re designed to print.

And the extra ornate these photomasks get, the extra computational horsepower is required to provide them. Nonetheless, utilizing GPUs, Nvidia believes it cannot solely velocity up this course of, however scale back the ability consumption required. The corporate claims that cuLitho operating on its GPUs is roughly 40x sooner than present computational lithography platforms operating on basic objective CPUs.

“It’ll assist the semiconductor business proceed the tempo of innovation that we’ve all come to depend on, and it’ll enhance the time to marketplace for every kind of chips sooner or later,” Singh claimed.

Nonetheless, at the least within the close to time period, Nvidia’s expectations appear to be slightly extra grounded. The corporate expects fabs utilizing cuLitho might produce 3-5x extra photomasks a day whereas utilizing 9 % much less energy, which if true, ought to assist to spice up foundries’ already skinny margins

And with the likes of ASML, Synopsys, and TSMC lining as much as combine Nvidia’s GPUs and libraries into their software program platforms and fabs, we can’t have to attend lengthy to see these claims put to the check.

TSMC is already investigating Nvidia’s GPUs and cuLitho to speed up ILT photomasks, whereas ASML and Synopsys are working to combine assist for GPU acceleration utilizing cuLitho of their computational lithography software program platforms.

And whereas Nvidia execs would like to promote its newest and costliest GPU architectures to those corporations, Singh notes that the library is appropriate with GPUs going again to the Volta technology, which made its debut in 2017.

Whereas Nvidia is utilizing GPUs to speed up these workloads, it is price noting that cuLitho is not utilizing machine studying or AI to optimize semiconductor design simply but. Nevertheless it’s no secret that Nvidia can be engaged on that exact downside.

“A lot of this has to do with accelerating the underlying primitive operations of computational lithography,” Singh mentioned. “However I’ll say that AI may be very a lot within the works in cuLitho.”

As our sister website The Subsequent Platform reported final summer time, Nvidia has been engaged on methods to speed up computational lithography workloads for a while now. In a analysis paper revealed in July, engineers on the firm used AI to design equal circuits 25 % smaller than these created utilizing conventional EDA platforms.

Nvidia is hardly the one firm investigating using machine studying to speed up circuit design. Synopsys and Cadence have each applied AI applied sciences into their portfolios, whereas Google researchers developed a deep-learning mannequin known as PRIME to create smaller and sooner accelerator designs. And beforehand, the corporate used reinforcement studying fashions to design parts of its tensor processing unit (TPU).

With that mentioned, the addressable marketplace for one thing like cuLitho is not that massive, and because of efforts by the US Commerce Division to stifle China’s fledgling semiconductor business, the quantity is barely getting smaller.

cuLitho will nearly actually be topic to US export controls governing the sale of superior semiconductor manufacturing tools and software program to international locations of concern, which for the second means China. Pressed on this level, Singh mentioned the library can be “accessible wherever this end-to-end OPC software program is on the market,” however declined to remark additional on US commerce restrictions. ®