NVIDIA has announced a breakthrough in advanced chip design as AI power grows exponentially.
As applications for artificial intelligence are developing rapidly, the need for ever higher computational power to train massive datasets for AI puts a premium on chipmaking technology. Conventional photolithography is increasingly difficult as gate density increases on modern semiconductors, and single digit nanometer devices are more time-consuming and difficult to fabricate.
Access all episodes of This Week in Engineering on engineering.com TV along with all of our other series.
* * *
Episode Transcript:
Artificial intelligence is on everyone’s mind today. With lightning-fast advances in the capability of AI platforms like ChatGPT, the number of potential use cases for the technology is also growing exponentially, including in engineering.
Morgan Stanley analyst Joseph Moore, commenting on recent developments at AI chip maker Nvidia, described artificial intelligence as “one of the most significant developments in technology since the development of mobile Internet”.
At the recent Nvidia developer conference, the company made several announcements directly related to engineering, notably in the automotive space and in semiconductor manufacturing. Announcements included BMWs selection of Nvidia to create the automaker’s first entirely virtual factory. Omniverse is a platform-as-a-product offering that’s notable for the replacement of conventional file formats with Universal Scene Description, or USD, an open standard for composing and simulating in 3D.
The development is significant because USD was not developed for digital twin applications but may offer photorealism and interactivity during the design phase for complex production processes or factory layouts. Invented by Pixar Animation Studio, USD is widely believed to be the future global standard for development in the metaverse.
On the hardware side, the company announced a new development in computational photolithography, using Nvidia software integrated by chip maker TSMC and design automation firm Synopsys. But significantly, equipment maker ASML will use the lithography systems and also work with Nvidia on graphics processor units, allowing higher component density and faster production.
Nvidia predicts 40 times better performance compared to current lithography, enabling 500 NVIDIA DGX H100 systems to achieve the work of 40,000 CPU systems, running computational lithography processes in parallel. Gate dimensions on the single digit nanometer scale were once thought to be impossible; a 7 nm technology is now a reality.
According to Nvidia, the combination of XML, Synopsys and TSMC will allow two nanometer scale, again pushing the limits of physics. And that task is complex. A separate photomask for each of up to 100 layers must be generated and the masts themselves are not simple negatives of the act, but must incorporate optical correction to address distortion, making the masts themselves a serious engineering challenge beyond a simple negative of the layer pattern.
Speed in lithography is essential, as larger die sizes translate to fewer devices per crystal wafer, resulting in lower overall component yields. Chiplet architecture implemented by several of the majors combines multiple smaller dies onto one package, but the need to train AI systems is exploding just as the task of manufacturing the high-performance processors needed, is becoming more difficult.
The open question is whether the computational intensity of AI workloads will favor developers with access to quantum computers, or if a new generation of conventional integrated circuits with higher densities will allow faster, cheaper AI applications in advanced high-value sectors such as scientific research, medicine and critically, the replacement of human beings in customer facing roles such as sales, customer service and human resources.