This article is more than 1 year old

Nvidia GPUs fly out of the fabs – and right back into them

Let's use AI to make better chips for AI, what could go wrong?

Comment Like salmon returning to their ancestral waters, Nvidia's GPUs make their way back into chip factories.

TSMC, ASML, and Synopsys are all using Nvidia's accelerators to speed up or boost computational lithography. Meanwhile, KLA Group, Applied Materials, and Hitachi are now using deep-learning code running on Nv's parallel-processing silicon for e-beam and optical wafer inspection.

Of course, fabs are packed with electronics so chips of all kinds return to the roost in a way. But it's interesting to see GPUs accelerating this part of the manufacturing process. As we shrink transistor gates, helping us pack more of them into our processors, computational lithography is needed – as is the case for several years – to produce die photomasks capable of etching the gates' ever-smaller features into silicon.

Until recently these workloads have largely run on CPU cores, though it turns out that given the right tweaks, GPUs are pretty good at accelerating such tasks.

For computation lithography, the process is fairly straightforward. Most chips are etched into silicon by projecting specific wavelengths of light — often in the extreme ultraviolet part of the spectrum — through a photomask. To produce smaller and smaller transistors on silicon dies, engineers have had to get creative to prevent, for one thing, distortion from blurring the features. Today, these photomasks are so ornate they're generated on massive compute clusters that can take weeks to complete.

Using GPU acceleration, Nvidia's CEO Jensen Huang claims this process can be sped up by 50x. "Tens of thousands of CPU servers can be replaced by a few hundred DGX systems, reducing power and cost by an order of magnitude," he said this week.

Using AI to build better chips

Anyone who's been paying attention as of late will know that graphics processors are good for more than high-performance computing. They're also the beating heart on which today's large-scale AI depends. And while Nvidia stopped short of claiming artificial intelligence is present in its cuLitho software stack in March, Vivek Singh, VP of Nvidia's advanced technology group, made it clear it's only a matter of time before AI is widely applied to computation lithography.

At the ITF semiconductor conference on Tuesday this week, Nvidia's Huang touted the potential of AI to breathe new life into Moore's Law. "Chip manufacturing is an ideal application for Nvidia accelerated and AI computing," he insisted. Well, he would say that.

While fabs are already highly automated, Nvidia sees an opportunity to apply the work it's doing around robotics, autonomous vehicles, and chatbots to chip manufacturing — and make millions selling GPUs to processor designers, equipment vendors, and foundries.

During his speech, Huang teased VIMA, a multimodal "embodied AI model trained to perform tasks based on visual text prompts, like rearranging objects to match a scene."

Huang didn't explicitly mention applying this to chip design software, but given the conference's emphasis on semiconductor manufacturing, it's easy to see how it could be. "I look forward to physics-AI robotics and Omniverse-based digital twins helping to advance the future of chipmaking," he added.

The time is ripe

Nvidia's sudden interest in selling the semiconductor industry on accelerated computing is hardly surprising.

Following the collapse of the consumer GPU market and the ongoing crypto winter, Nv's datacenter division has done most of the heavy lifting for the corporation. Huang even took a pay cut this past quarter on account of his company's middling performance.

Despite a somewhat recent downturn in chip demand — largely at the high-end — foundry operators, such as Samsung, TSMC, Intel, SK hynix among others, are pushing ahead with new foundry projects. This expansion has been fueled in large part by a series of moves in the US, Europe, and Asia Pacific region, that combined are worth well in excess of $100 billion in tax breaks and subsidies for the semiconductor world. Nvidia is in a prime position to benefit from increased chip development.

AI may have white collar workers worried — be sure to check out this week's Kettle for more on that — but it could address a shortage of skilled workers in the semiconductor space.

Early last year the Center for Security and Emerging Technology warned that the "re-shoring" of US semiconductor manufacturing could be hindered by a lack of skilled workers. By their estimate America would need to train, hire, or import an additional 27,000 workers.

TSMC founder Morris Chang has repeatedly pointed out how much more expensive semiconductor manufacturing is in the US compared to Taiwan.

To fill this void and contain costs, it wouldn't surprise us to see chipmakers going to greater efforts to automate even more of the manufacturing and design process by taking greater advantage of AI/ML. ®

More about

TIP US OFF

Send us news


Other stories you might like