²ÝÁñÉçÇø

World-altering innovation occurs at the intersection of opportunity and complexity.

This according to Rene Haas, CEO of Arm, and Sassine Ghazi, president and CEO of ²ÝÁñÉçÇø, who sat down at the 35^th annual ²ÝÁñÉçÇø User Group (SNUG) conference to share their views on the unprecedented opportunity and complexity of AI.

¡°It¡¯s an incredibly exciting time,¡± Haas said. ¡°I never thought I would have the opportunity to work on the kinds of problems we¡¯re trying to solve with artificial intelligence. And it just seems to go faster and faster and faster.¡±

¡°I strongly believe we are at an inflection point with AI,¡± Ghazi added. ¡°There are certain challenges and problems that humanity has not been able to solve yet.¡±

Like improving health outcomes on a global scale. Haas noted the traditionally methodical pace of medicinal research and elusiveness of breakthroughs.

¡°The time the initial R&D starts to the time you actually have a drug approved is about 15 years,¡± he said. ¡°And the failure rate from early research to end product is 95%. It¡¯s crazy.¡±

With AI accelerating health-related research, these longstanding norms are being upended.

¡°There¡¯s a huge promise, for example, of eliminating some phase one trials and even animal trials by using artificial intelligence,¡± Haas said. ¡°And the idea that we could see a cure for cancer in our lifetime is very, very real.¡±

Taming complexity, improving efficiency

¡°AI is going to touch every industry in terms of applications and productivity,¡± Ghazi said. ¡°It will enable new business models and new paradigms.¡±

But first, complexity must be tamed. Underlying technologies must catch up with the increasing speed of innovation. And ever-growing computing demands must be met in an efficient and sustainable manner.

¡°It's a huge, huge computing challenge,¡± Haas said. ¡°It¡¯s a huge power challenge.¡±

According to the International Energy Agency (IEA), data centers may consume ¡ª as much as the entire country of Japan. And it¡¯s only a near-term projection.

Ghazi and Haas discussed the gigantic data centers ¡ª like those being built for the ¡ª needed to support the insatiable processing demands of AI workloads, which continue to grow exponentially.

¡°The numbers are huge,¡± Haas said. ¡°Multiple gigawatts of energy over a geography that we¡¯ve never done before, thermals that we¡¯ve never dealt with before.¡±

In addition to scaling up and out, processing performance and power efficiency must also shrink.

¡°We focus so much on the infrastructure build-out and the data center, but the moment you start running AI on devices, that's another opportunity,¡± Ghazi noted.

¡°In mobile, the designs are unbelievably complex in terms of power, geometry, cost, and speed. How do I run AI in a wearable?¡± Haas questioned. ¡°How do I run AI in a battery-operated device that still has to run displays, still has to run the operating system, still has to launch your applications?¡±

¡°It¡¯s a system-level problem,¡± said Ghazi.

Holistic design, from silicon to systems

The technologies being used today to drive AI workloads were never designed for those workloads. Silicon devices must become more specialized, Ghazi explained. And they must be purpose-built for the applications and systems they will serve.

¡°Designing a general-purpose chip and using it for many applications ¡ª those days are gone,¡± he claimed. ¡°Especially for AI applications where you need to have an efficient silicon for a specific workload.¡±

¡°People will throw down lots of transistors to address a general-purpose problem,¡± Haas added. ¡°You¡¯re leaving performance on the table, you¡¯re leaving efficiency on the table, and you¡¯re delaying the capability of the software to take advantage of it. There is a gigantic opportunity to do things from a design standpoint that are far, far more power efficient than they are today.¡±

Design optimization won¡¯t be limited to the performance and energy efficiency of silicon. It will encompass entire systems ¡ª from hyperscale data centers to cars, drones, and robots ¡ª and all of the physics they encounter.

¡°You have to consider the mechanics, thermal dynamics, and fluid dynamics in relation to the electrical and electronic components,¡± Ghazi said, noting the need for holistic, silicon-to-systems design. ¡°It¡¯s a level of complexity we¡¯ve never seen before, and AI will help us solve it.¡±