Manufacturing of chips utilizing modern course of applied sciences requires extra compute energy than ever. To deal with necessities of 2nm nodes and past, NVIDIA is rolling out its cuLitho software program library that makes use of the corporate’s DGX H100 methods primarily based on H100 GPUs and guarantees to extend efficiency accessible to masks retailers inside an inexpensive quantity of consumed energy by 40 occasions.
Trendy course of applied sciences push wafer fab tools to its limits and sometimes require finer decision than is bodily attainable, which is the place computational lithography comes into play. The first function of computational lithography is to reinforce the achievable decision in photolithography processes with out modifying the instruments. To take action, CL employs algorithms that simulate the manufacturing course of, incorporating essential knowledge from ASML’s tools and shuttle (take a look at) wafers. These simulations help in refining the pellicle (photomask) by intentionally altering the patterns to counteract the bodily and chemical influences that come up all through the lithography and patterning steps.
There are a number of computational lithography strategies, together with Decision Enhancement Expertise (RET), Inverse Lithography Expertise (ILT, a way to scale back manufacturing variations by using non-rectangular shapes on the photomask), Optical Proximity Correction (OPC, a method for bettering photolithography by correcting picture inaccuracies ensuing from diffraction or process-related impacts), and Supply Masks Optimization (SMO). All of them are extensively used at in the present day’s fabs.
In the meantime, compute-expensive strategies like inverse lithography expertise and supply masks optimization are particular to a given design. They should be applied individually for every chip to make sure applicable decision and keep away from yield-limiting hotspots. Synthesis of pellicles that use RET, ILT, OPC, and SMO entails the utilization of computational lithography. As nodes get thinner, the complexity of computations will increase, and compute horsepower turns into a bottleneck for masks retailers as every trendy chip makes use of dozens of pellicles. For instance, NVIDIA’s H100 makes use of 89 of them.
NVIDIA says that computational lithography at present consumes tens of billions of CPU hours yearly and, subsequently, huge energy. In the meantime, extremely parallel GPUs like NVIDIA’s H100 promise larger efficiency at decrease value and energy consumption. Specifically, NVIDIA says that 500 of its DGX H100 methods packing 4000 of its H100 GPUs (that devour 5 MW of energy) and utilizing computational lithography software program that makes use of cuLitho can supply the efficiency of 40,000 CPU-based methods which devour 35 MW that TSMC makes use of in the present day. The corporate additionally goes on to say that masks makers can produce 3 – 5 occasions extra pellicles per day utilizing 9 occasions much less energy than they use in the present day as soon as they begin counting on GPU-accelerated computational lithography, one other declare that requires verification by precise masks retailers, however which provides a fundamental understanding the place the corporate desires to go.
“With lithography on the limits of physics, NVIDIA’s introduction of cuLitho and collaboration with our companions TSMC, ASML, and Synopsys permits fabs to extend throughput, scale back their carbon footprint and set the inspiration for 2nm and past.”
Whereas efficiency targets set by NVIDIA are spectacular, it must be famous that the cuLitho software program library for computational lithography have to be integrated in software program supplied by ASML, Synopsys, and TSMC effectively utilized by their companions, amongst masks retailers. For current-generation lithography (assume 7 nm, 5 nm, and three nm-class nodes), masks retailers already use CPU-based computational lithography options and can proceed to take action for at the least some time. That is maybe why NVIDIA is discussing its computational lithography efforts in context with next-generation 2 nm-class nodes and past. In the meantime, it is sensible to anticipate foundries and masks retailers to at the least strive deploying NVIDIA’s cuLitho for a few of their upcoming 3 nm-class nodes to extend yields and efficiency. TSMC, for instance, will begin to qualify cuLitho in mid-2023, so anticipate the platform to be accessible to the corporate’s prospects starting in 2024.
“Computational lithography, particularly optical proximity correction, or OPC, is pushing the boundaries of compute workloads for essentially the most superior chips,” mentioned Aart de Geus, chief govt of Synopsys. “By collaborating with our associate NVIDIA to run Synopsys OPC software program on the cuLitho platform, we massively accelerated the efficiency from weeks to days! The team-up of our two main corporations continues to power wonderful advances within the business.“
An official assertion by NVIDIA states that “A fab course of change typically requires an OPC revision, creating bottlenecks.” “cuLitho not solely helps take away these bottlenecks, but it surely additionally makes attainable novel options and modern strategies like curvilinear masks, excessive NA EUV lithography, and sub-atomic photoresist modeling wanted for brand spanking new expertise nodes.“
Further compute horsepower accessible for computational lithography purposes will are available in significantly helpful for the subsequent technology of manufacturing nodes that may use Excessive-NA lithography scanners and can mandate the utilization of ILT, OPC, and SMO to contemplate bodily peculiarities of lithography scanners and resists to make sure respectable yields, low variation (i.e., foreseeable efficiency and energy consumption), and predictable prices. In the meantime, computational prices for RET, ILT, OPC, and SMO will inevitably enhance at 2 nm and past, so it seems like NVIDIA will introduce its cuLitho platform at a superb time.