Debate Over Health Of Moore’s Law Continues

Semicon West 2019 was kicked off by the ‘AI Design Forum’ and featured a panel of CEOs that debated if Moore’s Law was still making power, performance and area optimization possible in the same way as it had been.

Synopsys chairman and co-CEO Aart de Geus asserted that Moore’s Law is completely alive. “The discussion of Moore’s Laws invariably goes back to the ‘65 document, and then people exactly track where are the economics, where’s the technology; that’s not the way to think about it. The way to think about it is Moore’s Law is the behavior of an exponential that has techonomic feedback on the exponential that drove a revolution of what mankind can do. The reason I’m saying it’s completely alive is because right now we’re facing another decade or two of amazing opportunities that themselves economically will drive the push for technology without a stop. Maybe it’s not exactly the same curve that Moore actually drew, it doesn’t matter. The impact is what matters of the exponential.”

Victor Peng, CEO of Xilinx agreed with de Geus in the sense that while there will still be quite significant improvements in computing, those improvements will take a very different form. “In the general sense, if you take any processor, and every two years you move it to the next node, even if you didn’t change the architecture, the performance, the power, and the area would get better. Today, you definitely get one; you get to pick which one you want. You could work really hard or depending upon the application you get two, [but] it’s really hard to get all three. In that essence, in a more traditional sense of Moore’s Law, that’s not really working. On top of that, it’s been said for a lot time, this happens every two years, and clearly the clues for when the next node comes is taking longer than for economic as much as technical reasons.”

Since Lisa Su, CEO of AMD reminded that she began her career as a process technologist, she can’t bring herself to say Moore’s Law is dead. “Moore’s Law is slowing, but Moore’s Law is something that we all strive to in terms of how do we get more performance for our products. There are lots of things that you can do in addition to process technology to keep that performance trend going. That’s what we’re all trying to do, whether it’s in hardware or software or systems. The key is how we keep that performance trend for our industry, that’s driving all of the applications, going? There’s a tremendous amount of effort to augment Moore’s Law.”

When it comes to memory and storage, Sanjay Mehrotra, CEO of Micron Technology said for almost a decade, Moore’s Law has slowed down significantly and actually has been less relevant. “A lot of innovations have been driven in memory and technology, for example, going from 2D NAND to 3D NAND that has the capability to bring down costs as well as gain more Gigabits per wafer. In DRAM, certainly there’s a lot of innovation happening with high bandwidth memory stacking of multiple chips and DSP technologies to gain performance, to get the benefit of higher capacities. But there is no question that Moore’s Law is significantly challenged in memory and storage. Looking at 10 years ago versus today in NAND as well as in DRAM, the year/year bit growth that you could get from one technology transition to the next technology transition, that bit growth, that cost decline capability has more than halved now, and certainly there are challenges.”

He remains optimistic: “Engineers are always finding new, innovative ways to address memory and technology challenges, whether from a process point of view or a design point of view, or an architecture point of view.”

Gary Dickerson, CEO of Applied Materials referred to classic Moore’s Law. “You think about shrinking in two dimensions. Today, Victor’s correct: it doesn’t deliver simultaneous improvements in power, performance, area and cost. Over the last few years, that classic approach has really run out of gas but we’ve talked about five different drivers in a new playbook. One is architecture innovation and especially with AI and big data with workload specific types of applications, in the edge, cloud, training, inference architecture innovation. This is a huge focus for many different companies. The innovation around new structures, Sanjay talked about 2D to 3D NAND, where the playbook in NAND for decades was to strengthen two dimensions and then now we go vertical so it’s a different driver. You hit a physical limit and you hit a cost limit so now it’s about new structures, about new materials, also how you connect the chips together, in packages and heterogeneous integration. That’s another driver. The classic 2D scaling, no question, that has run out of gas, but there are all of these new opportunities in structure innovation, materials innovation, architectures, packages, in addition to 2D.”

One outstanding question for the industry is that, if costs aren’t falling, will semiconductors continue to have the same kind of dramatic impact on the world as they have had?

“Everybody, correctly, knows what the technical challenges are on the individual steps and the cost per transistor is a challenge,” de Geus said. “But what has changed so fundamentally is where the cost matters. It matters in the ultimate application domain, i.e. the verticals and there’s no question whatsoever that we’re facing a multi decade change of every economic vertical being impacted by the computation that will be possible. That compensation is predicated of course of faster, more memory, more data, you name it, but the economic impact is actually not limited to the semiconductor industry. It is limited to every vertical and if the impact is as big as we think it is, which is machine learning, AI plus plus plus, clearly would indicate that is, then the economics will continue to be sufficient to try and to go vertical to put more things into the same package, to continue to invent and push on new materials.”

As such, every door that’s semi-open will be pushed rapidly, he said, because the economics are there.

“We may have been surprised that the semiconductor industry has done as well as it has with its ups and downs. But overall the impact is not in question whatsoever. I would propose that therefore our own thinking should be not limited. Certainly it should be supportive of the technical challenges but it should be driven by the notion of the impact of the exponential,” de Geus added.

Related Stories
The Impact Of Moore’s Law Ending
Chips will cost more to design and manufacture even without pushing to the latest node, but that’s not the whole story.
Unsticking Moore’s Law
Applied Materials’ VP looks at what’s next for semiconductor manufacturing and the impact of variation, new materials and different architectures.
Sidestepping Moore’s Law
Why multi-die solutions are getting so much attention these days.
Moore’s Law Knowledge Center