NVIDIA cuLitho GPU-Accelerates Chip Fab Lithography For A Huge Efficiency Lift
It's kind of ironic, but one of the main drivers of improved computer performance is in fact... improved computer performance. To design ever-faster chips, you need ever-faster chips, and that's true at several layers of the processor design process. At the very beginning, when actually creating these chips, companies like TSMC spend immense computing power on computational lithography to create photomasks for fabrication.
Creating these photomasks is, as Jensen Huang pointed out at GTC today, "fundamentally an imaging problem at the limits of physics." NVIDIA's spent a lot of time and money on computer imaging recently, and it seems it's time to put that expertise to work in the fabrication plant. The company just announced "cuLitho", a library intended to accelerate computational lithography on NVIDIA's GPUs.
NVIDIA makes some big claims about cuLitho in its press release, echoing what Huang said at GTC. Apparently, computational lithography is a highly-parallel process, because moving it from CPUs to GPUs supposedly drops the power consumption from 35 MW to "just" 5 MW—1/7th the power. It also accelerates performance: Huang said that a process that currently takes two weeks on CPUs can be performed in "a single 8 hour work shift" using NVIDIA GPUs and cuLitho.
If his statements are accurate in the general case, then this kind of acceleration could really be transformative for processor manufacturing, and it makes you wonder why this hadn't been done before. NVIDIA says that "longer term, cuLitho will enable better design rules, higher density, [and] higher yields," all of which are obviously very desirable in the manufacturing business.
Team Green has partnered with TSMC, Synopsys, and equipment maker ASML on cuLitho, and says that it expects cuLitho to enter qualification for production in June of this year.