Nvidia's GTC online keynote this week (opens in new tab) introduced a new software tool to accelerate the chip manufacturing process, known as cuLitho, which accelerates the mask design process used in lithography by 40 times.

According to Nvidia, chip manufacturing industry leaders TSMC, ASML, and Synopsys have all signed on to cuLitho. Of course, cuLitho only works with Nvidia GPUs, and those GPUs are manufactured by TSMC.

In other words, Nvidia outsources the manufacturing of its GPUs to TSMC and then sells them outright to TSMC just for use in cuLitho. Well done.

Nvidia states that about 4,000 H100 Hopper server GPUs could replace 40,000 CPU-based servers, doing the same work 40 times faster, yet consuming one-ninth the power.

The task in question is the design of the photomasks used in the lithography process. Simply put, chips are made by shining light onto a silicon wafer through a patterned mask and etching individual components, including wires and transistors, onto the wafer surface.

In practice, the chip's functionality is built layer by layer using a series of masks. Apparently, it takes 89 masks to make Nvidia's own H100 Hopper GPU.

In any case, given the enormous complexity of the latest chips, which contain billions of functions, the mask design requires a great deal of effort. Furthermore, masks that suppress diffraction so that sharp, rather than blurry, images are etched onto the wafer require complex calculations.

As feature sizes become increasingly smaller, the computational task of reducing diffraction only becomes more difficult. Therefore, huge server arrays such as TSMC's are tackling this task.

Nvidia says its new cuLitho can reduce the computational load to create a mask from two weeks to eight hours, speeding up the overall chip design and manufacturing process. Ultimately, Nvidia believes cuLitho will not only speed things up, but also enable higher chip density and better yields.

Whatever it is, selling chips to the company that manufactured them so that they can manufacture them faster is a fun roundabout ploy that only Nvidia could pull off. However, it should help not only Nvidia itself, but all of TSMC's customers, so let's hope for faster and cheaper CPUs and GPUs in the future.