In addition to the new H100 superchip, Nvidia showed off a brand new technology at the recently concluded GTC conference nothing. This will ensure a fundamental acceleration of the lithographic production process. It focuses on one of the most critical parts of the entire development – the calculation of the photomask, which is needed to transfer the chip design to the silicon wafer itself.
The complexity of this technology is well illustrated by the graph below with the production process on the horizontal axis and the processor time and number of data centers needed on the vertical axes. While eleven years ago about 10 million processor hours per year were enough to calculate the photomask with the 28nm Kepler architecture, with the current Hopper it is up to a thousand times as much. The demands are therefore enormous and unsustainable in the future.
The reason why the calculation of the photomask to transfer the chip design becomes more and more complex is actually simple: diffraction occurring on the optical elements that transfer the image to the wafer. This phenomenon is well known to all photographers who look for the diffraction limit of their lenses: that is, the aperture value at which diffraction begins to affect the sharpness of images. But on the photomask, which contains a chip design with 80 billion transistors, it is not possible to simply change the aperture and eliminate light refraction.
The light source illuminates the photomask containing the chip design. The lens directs the resulting image of this template onto a silicon wafer (source: Nvidia)
On the contrary, the photomask must adapt to diffraction, which contains the design of the chip with such a correction that after passing through all the optical elements it lands on the wafer in the originally intended shape. In addition, each chip consists of tens to hundreds of layers, where a new photomask must be calculated for each of them. For the most complex designs, a single mask can take several days to weeks to calculate. And that despite the fact that, according to Nvidia, up to 40,000 servers take care of it.
On the left is the final design of the chip: the shapes can be imagined as very tiny holes. In order to accurately transfer this design, the photomask must look like the image on the right.
cuLitho can reduce this time many times, and with the currently introduced Hopper architecture and the H100 chip, it should be eight hours to calculate one photomask instead of two weeks. cuLitho is a software library that uses dozens of parallel algorithms that speed up the calculation itself and optimize the use of processor time.
Nvidia has not released many details and, among other things, cuLitho is classified as sensitive software that should not reach China due to trade sanctions. Taiwan’s TSMC, the Netherlands’ ASML and the USA’s Synopsys are currently starting to integrate it into their processes.
