Nvidia Tackles Chipmaking Course of, Claims 40X Pace Up with cuLitho
At GTC 2023, Nvidia introduced its new cuLitho software program library for dashing up a crucial bottleneck within the semiconductor manufacturing workflow. The brand new library speeds computational lithography, a way used to create photomasks for chip manufacturing. Nvidia claims its new method permits 500 DGX H100 programs wielding 4,000 Hopper GPUs to do the identical quantity of labor as 40,000 CPU-based servers, however achieve this 40X sooner and with 9X much less energy. Nvidia claims this reduces the computational lithography workload to supply a photomask from a number of weeks all the way down to eight hours.
Chipmaking leaders TSMC, ASML, and Synopsys have all signed on for the brand new tech, with Synopys already integrating it into its software program design instruments. Over time, Nvidia expects the brand new method to allow greater chip density and yield, higher design guidelines, and AI-powered lithography.
Nvidia scientists created new algorithms that permit increasingly-complex computational lithography workflows to execute on GPUs in parallel, exhibiting a 40X speedup utilizing Hopper GPUs. The brand new algorithms are built-in into a brand new cuLitho acceleration library that may be built-in into masks makers’ software program (usually a foundry or a chip designer). The cuLitho acceleration library can be appropriate with Ampere and Volta GPUs, although Hopper is the quickest resolution.
Printing the small options on a chip begins with a piece of quartz referred to as a photomask. This clear quartz has an imprinted sample of a chip design and works very like a stencil — shining a lightweight via the masks etches the design onto the wafer, thus creating the billions of 3D transistors and wire constructions that comprise a contemporary chip. Every chip design requires a number of exposures to construct up the chip’s design in layers. As such, the variety of photomasks used through the chipmaking course of varies primarily based on the chip; it could even exceed 100 masks. As an illustration, Nvidia says it takes 89 masks to create the H100, and Intel cites ’50+’ masks used for its 14nm chips.
New methods have emerged that now permit etching options smaller than the wavelength of the sunshine used to create them. Nonetheless, the continued shrinkage of the options has led to points with diffraction, which primarily ‘blurs’ the design that is being printed onto the silicon. The sector of computational lithography counteracts the influence of diffraction via advanced mathematical operations that optimize the masks structure. Nonetheless, this activity is changing into more and more compute-intensive as options shrink even additional, thus enabling billions extra transistors per design.
These advanced issues require giant clusters of computer systems, usually numbering tens of 1000’s of servers (Nvidia cites 40,000), that crunch via the numbers in parallel on CPUs in a workload that may take as much as weeks to course of a single photomask (the period of time varies primarily based on chip complexity — Intel says it takes its workforce 5 days to create a single masks).
Nvidia contends that the variety of servers required to design a contemporary masks is rising on the similar fee as Moore’s Legislation, thus pushing the server necessities and the quantity of energy wanted to function them into unsustainable territory. In reality, the unbelievable compute necessities for brand new masks tech, like Inverse Lithography Technology (ILT) which uses Inverse Curvilinear Masks (ILM), has already hampered the adoption of those extra superior methods. Moreover, Excessive-NA EUV and ILT are anticipated to extend the quantity of knowledge processing for masks by 10X within the coming years.
That is the place Nvidia’s cuLitho steps in, lowering the computational lithography workload to eight hours. The cuLitho library will be built-in into computational lithography software program that leverages ILT (curvilinear shapes) or Optical Proximity Correction (OCP, which makes use of ‘Manhattan’ shapes) methods, and is already built-in into Synopsys’ instruments. TSMC and ASML are additionally adopting the tech. Given the sensitivity of those kinds of software program, US export controls will govern any distribution of the software program to China and different areas topic to sanctions.
Intel has lengthy used its personal proprietary software program instruments however is slowly shifting to adopting industry-standard instruments, notably because it begins implementing its personal exterior IDM 2.0 foundry operations. As such, it’s but to be seen if different massive fabs, like Intel and Samsung, will undertake the brand new software program for their very own inside instruments. Regardless, the assist from Synopsys, ASML, and TSMC assures broad uptake of the cuLitho library and Nvidia’s GPU-based options with main semiconductor producers over the approaching years.