REVERSE LITH | 40X | 2023'

If computer processors make the modern world go round, then Nvidia and TSMC are the flywheels keeping it spinning. When the former claims to have made a breakthrough in semiconductor manufacturing and the latter confirms they are about to implement it, it is important to pay attention.

This week at Nvidia's GTC developer conference, CEO Jensen Huang announced that Nvidia has developed software that speeds up inverse lithography, a chip-making process, by more than 40 times. A process that typically takes weeks can now be completed overnight, and instead of requiring approximately 40,000 CPU servers and 35 megawatts of power, it will require only 500 Nvidia DGX H100 GPU-based systems and 5 megawatts.

"With cuLitho, TSMC can reduce prototype cycle time, increase throughput, and reduce the carbon footprint of their manufacturing, as well as prepare for 2nm and beyond," he said.

On this project, Nvidia collaborated with some of the industry's most well-known names. This summer, TSMC, the world's largest semiconductor foundry, intends to qualify the method for production. Meanwhile, chip designer Synopsis and equipment manufacturer ASML announced in a press release that they will incorporate cuLitho into their respective chip design and lithography software.

What Is Reversible Lithography?

Ultraviolet light is shone through intricate "stencils" to engrave billions of patterns, such as wires and transistors, at near-atomic resolutions onto smooth silicon wafers. This technique, known as photolithography, is used to physically manifest every new processor design, from Nvidia to Apple to Intel, in silicon.

ASML's devices can produce near-impeccable nanoscale works of art on chips at a cost of hundreds of millions of dollars each. The final product, an example of which is presumably humming near your fingertips as you read this, is likely the most complex commodity in human history. (Every six months, TSMC manufactures one quintillion transistors—just for Apple.)

Engineers have had to be inventive in order to produce more powerful processors with more and smaller transistors.

Do you recall the stencil mentioned previously? It is the strangest stencil you've ever observed. Today's transistors are smaller than the light wavelength used to imprint them. Chipmakers must design stencils, or technically, photomasks, that can transform light into interference patterns with features smaller than the wavelength of the light and that precisely match the chip's layout.

In the past, the geometry of photomasks was more straightforward; a rectangle projected a rectangle. However, as time has passed, photomasks have become increasingly complex. Today's most sophisticated masks resemble mandalas rather than basic polygons.

As atomic-scale patterns have decreased, "stencils" or photomasks have become progressively more complicated. Image Source: Nvidia

Engineers invert the process to design these sophisticated photomask patterns.

They begin with the desired design and then run it through a tangled web of equations describing the physics involved in order to create an appropriate pattern. This process is referred to as inverse lithography, and as the disparity between light wavelength and feature size has grown, it has become an increasingly vital component of the entire procedure. However, as the complexity of photomasks increases, so do the computing power, time, and cost necessary for their design.

"Computational lithography is the largest computation workload in chip design and fabrication, consuming tens of billions of CPU hours annually," said Huang. Massive data centers operate nonstop to generate lithography system reticules.

In the broader category of computational lithography, which encompasses the techniques used to design photomasks, inverse lithography is one of the more recent and sophisticated methods. Its advantages include a greater depth of field and resolution, and it should benefit the entire processor; however, due to its intensive computational load, it is currently employed sparingly.

The Parallel Library

Nvidia intends to lessen this burden by making the computation more amenable to graphics processing units, or GPUs. These potent chips are used for tasks that require a large number of basic computations that can be completed in parallel, such as video games and machine learning. Therefore, it is not sufficient to simply execute existing processes on GPUs, which yields only a modest improvement, but these processes must be modified for GPUs.

This is the function of the new software cuLitho. The product is a collection of algorithms for the fundamental operations used in inverse lithography. It was created over the course of four years. By decomposing inverse lithography into smaller, more repetitive computations, the entire procedure can now be partitioned and parallelized on GPUs. According to Nvidia, this expedites everything significantly.

A new library of inverse lithography algorithms can accelerate the process by decomposing it into smaller tasks that are executed in parallel on GPUs. Image Source: Nvidia

"If [inverse lithography] was 40 times faster, would more individuals and companies use full-chip ILT on more layers? Vivek Singh, vice president of Nvidia's Advanced Technology Group, stated as much during a GTC presentation.

With a faster, less computationally intensive process, manufacturers can rapidly iterate on experimental designs, modify existing designs, produce more photomasks per day, and extend the use of inverse lithography to a greater portion of the chip, he explained.

This last point is essential. Wider use of inverse lithography should reduce print errors by sharpening the projected image, allowing chipmakers to produce more working circuits per silicon wafer, and be precise enough to create features smaller than 2 nanometers.

It turns out that hardware is not the only factor in producing superior chips. Software enhancements, such as cuLitho or the increased use of machine learning in design, can have a substantial effect as well.