$350 H100 replaces 40,000 CPUs! TSMC applies NVIDIA platform to accelerate computing and trigger a semiconductor manufacturing revolution

During the "NVIDIA AI Summit" held in Washington on Tuesday, October 8th, NVIDIA emphasized the achievements of its collaboration with Taiwan Semiconductor in the accelerated computing field: NVIDIA's cuLitho computing lithography platform is now in production at Taiwan Semiconductor, significantly increasing the speed of the chip manufacturing cornerstone step of computing lithography and reducing energy consumption.

NVIDIA stated that cuLitho is introducing accelerated computing into the field of computing lithography. The production deployment of cuLitho enables Taiwan Semiconductor to accelerate the development of next-generation chip technologies, as the current production processes are approaching the physical limits. Applying cuLitho in production at Taiwan Semiconductor can enhance the speed of manufacturing next-generation advanced semiconductor chips and break through physical limitations.

Computing lithography is a key step in computer chip manufacturing. ASML defines it as using computer modeling, simulation, data analysis, etc., to predict, correct, optimize, and validate the imaging performance of lithography processes under a series of patterns, processes, and system conditions. It involves complex calculations such as electromagnetic physics, photochemistry, computational geometry, iterative optimization, and distributed computing, making it the most computationally intensive workload in the entire semiconductor design and manufacturing process. Wafer foundries typically set up large data centers specifically for these calculations. With the size of advanced chips getting smaller, down to 3nm and below, the need for precise computing lithography is increasing, and the time required for computing lithography is also growing. Without more powerful computing lithography, it is difficult to achieve complex mask designs. Computing lithography has always been a bottleneck in bringing new technology nodes and computer architectures to the market.

At last year's NVIDIA GTC developer conference, NVIDIA introduced the cuLitho computing lithography technology software library built on GPUs, which is considered a game-changing software in the field of computing lithography. The core of cuLitho is a set of parallel algorithms invented by NVIDIA scientists, where all parts of the lithography process can run in parallel. What used to require 40,000 CPU systems to complete can now be done with just 500 NVIDIA DGX H100 systems. Wafer fabs using cuLitho can increase their daily mask production by 3-5 times, while reducing power consumption by 9 times compared to current configurations.

This Tuesday, NVIDIA reiterated the achievements disclosed at this year's GTC conference. Advanced foundries spend hundreds of billions of hours of CPU computing time on computing lithography each year, with a single chip mask set potentially requiring 30 million hours or more of CPU computing time, necessitating large data centers for foundries. By accelerating computing, 350 sets of NVIDIA H100 Tensor Core GPU systems can now replace 40,000 CPU systems, speeding up production, while reducing costs, space requirements, and power consumption.

TSMC CEO C.C. Wei stated at this year's NVIDIA GTC conference that by integrating GPU accelerated computing into TSMC's workflow in collaboration with NVIDIA, TSMC has significantly improved performance, increased throughput, shortened cycle times, and reduced power consumption NVIDIA revealed at this year's GTC conference that since its launch last year, cuLitho has brought new opportunities for Taiwan Semiconductor's innovative patterning technology. Tests conducted on the shared workflow with cuLitho showed that NVIDIA and Taiwan Semiconductor have respectively increased the speed of curvilinear process and traditional Manhattan process by 45 times and nearly 60 times. The difference between these two processes lies in the fact that the light mask shape in the curvilinear process is curved, while the light mask shape in the Manhattan process is limited to horizontal or vertical.

On Tuesday this week, NVIDIA also mentioned its developed generative AI application algorithm. It has been proven that this algorithm enhances the value of the cuLitho platform. Building on the accelerated process speed of cuLitho, the new generative AI workflow has further doubled the speed. Applying generative AI can create nearly perfect inverse light masks or reverse solutions, solving the diffraction problem of light in computational lithography, and then deriving the final light mask through traditional rigorous physical methods, thereby speeding up the entire Optical Proximity Correction (OPC) process by two times.

OPC has been applied in semiconductor lithography for thirty years. NVIDIA stated that in the past thirty years, few technologies like accelerated computing and AI have brought such rapid transformation to OPC. These technologies make physical simulations more accurate and achieve mathematical techniques that were once resource-intensive.

Many changes in wafer fabrication processes currently require modifications to OPC, increasing computational workload and creating bottlenecks in the development cycle of wafer fabs. The significant speed increase in computational lithography accelerates the creation of each mask in wafer fabs, thereby shortening the total cycle of developing new technology nodes. More importantly, it enables new calculations that were previously impossible.

NVIDIA cited an example that twenty years ago, reverse lithography technology was proposed in scientific literature, but due to long computation times, accurate calculations on a full-chip scale were largely unachievable. With cuLitho, this is no longer the case. Leading foundries will use it to enhance reverse and curvilinear solutions, which will help create the next generation of powerful semiconductors