Nvidia Tackles Chipmaking Process, Claims 40X Speed Up with cuLitho
At GTC 2023, Nvidia unveiled a new cuLitho software library that accelerates critical bottlenecks in semiconductor manufacturing workflows. The new library accelerates computational lithography, a technique used to create photomasks for chip manufacturing. With its new approach, Nvidia claims that 500 DGX H100 systems with 4,000 hopper GPUs can do the same amount of work as a 40,000 CPU-based server, but 40 times faster and in 9 minutes. It can run at 1 power. Nvidia claims this cuts the computer lithography workload to create a photomask from weeks to eight hours.
Chip manufacturing leaders TSMC, ASML, and Synopsys have all signed on to the new technology, and Synopys has already integrated it into its software design tools. Nvidia expects this new approach to enable higher chip densities and yields, improved design rules, and AI-powered lithography.
Scientists at Nvidia have created new algorithms that allow increasingly complex computational lithography workflows to run in parallel on GPUs, showing a 40x speedup using Hopper GPUs. New algorithms are integrated into the new cuLitho acceleration library and can be integrated into the mask maker’s software (usually foundries or chip designers). The cuLitho acceleration library is also compatible with Ampere and Volta GPUs, but Hopper is naturally the faster solution. Ultimately, Nvidia claims this will allow 500 DGX H100 systems with 4,000 hopper GPUs combined to do the work of his 40,000 servers using CPUs to process the workload. .
Printing small features on a chip begins with a block of quartz called a photomask. This clear quartz is imprinted with the pattern of the chip design, acting like a stencil. Shining light through the mask, the design is etched into the wafer, creating the billions of 3D transistors and wire structures that make up modern chips. Each chip design requires multiple exposures to build up the chip’s design in layers. Therefore, the number of photomasks used during the chip manufacturing process varies from chip to chip. There can be over 100 masks. For example, Nvidia says he needed 89 masks to make the H100, and Intel lists “50+” masks used for his 14nm chips, for example. .
A new technique has emerged that can etch features smaller than the wavelength of light used to create the features. However, the continual shrinkage of features causes diffraction problems, essentially “blurring” designs printed on silicon. cancels out the effect of However, as features shrink further, this task will become increasingly computationally intensive, allowing billions of transistors per design.
These complex problems include workloads that take up to weeks to process a single photomask (the amount of time depends on the complexity of the chip — Intel says it takes the team 5 They say it will take days).
Nvidia claims that the number of servers required to design modern masks is growing at the same rate as Moore’s Law, making server requirements and the amount of power required to operate them unsustainable. In fact, the staggering computing requirements of new mask technologies such as Inverse Lithography Technology (ILT) using Inverse Curvilinear Masks (ILM)), which has already impeded the adoption of these more advanced technologies. Moreover, high NA EUV and ILT are expected to increase mask data throughput by a factor of 10 in the next few years.
So Nvidia’s cuLitho steps in and cuts the computational lithography workload down to eight hours. The cuLitho library can be integrated into computational lithography software that utilizes ILT (curvilinear geometry) or optical proximity correction (OCP using “Manhattan” geometry) techniques and is already integrated into Synopsys tools. TSMC and ASML have also adopted this technology. Due to the sensitive nature of this type of software, US export controls govern the distribution of software to China and other sanctioned territories.
Intel has long used proprietary proprietary software tools, but is slowly moving toward adopting industry-standard tools, especially as it begins implementing its own external IDM 2.0 foundry operations. So it remains to be seen if other big fabs like Intel and Samsung will adopt the new software into their internal tools.