Moore's Law, the observation that the number of transistors on a computer chip doubles every 24 months, has taken a beating as progress miniaturizing circuitry falters. But chip giant Intel has plotted a course to keep the idea alive with a plan to pack 50 times as many transistors onto processors than is possible today.
The progress of Moore's Law, named after Intel co-founder Gordon Moore, has spread chips from expensive mainframes in the 1960s to personal computers in the 1980s and now to smartphones, watches, cars, TVs, washing machines and just about anything with electrical power.
Moore's Law has worked by shrinking transistors, the data-processing elements on a chip. Intel plans to keep shrinking them, but also to increase density by stacking chips into multilayer packages.
"We firmly believe there is a lot more transistor density to come," said Intel Chief Architect Raja Koduri, in a speech Monday for the Hot Chips conference for cutting-edge processor revelations. "The vision will play out over time -- maybe a decade or more -- but it will play out."
Koduri's optimism mirrored the excitement of many other companies at Hot Chips, an engineering conference where researchers detail progress. AMD, Nvidia, Google, Microsoft, IBM and a gaggle of startups showed ways they're advancing both general-purpose chips and those dedicated to tasks such as artificial intelligence, graphics and networking.
How Intel expects to deliver chip progress
Koduri described several steps to cram more transistors into a chip than possible with 10nm chips like its Tiger Lake processor arriving in laptops this fall. First will come the most traditional approach, shrinking transistors and squeezing them closer together. That'll triple the transistor density, Koduri predicted.
Next up are new transistor designs that continue the current transformation of transistors from flat circuitry elements into 3D structures. These steps, called nanowires and stacked nanowires, should quadruple density.
Then come packaging innovations, with chips stacked into a layer cake of processor elements. That should quadruple density again. The total math brings density up by about a factor of 50.
Years of Intel difficulties
Intel's optimism contrasts with difficult times keeping Moore's Law ticking.
Intel, once the unquestioned leader in chip manufacturing, has struggled in recent years. Its move from a manufacturing process with transistor features measuring 14 nanometers to later 10nm took five years instead of two. A nanometer is a billionth of a meter, and with circuitry elements 14nm wide, Intel can pack about 7,000 across the width of a human hair.
Next, Intel delayed its move from 10nm to 7nm manufacturing by six months, and Apple is dumping Intel chips from its Macs. To help adjust, Intel has adopted a more flexible design process that lets it rely more on other chipmakers like its top rival, Taiwan Semiconductor Manufacturing Corp.
Moore's Law, but at what cost?
TSMC, which moved to 7nm manufacturing about two years ago and makes Apple's iPhone chips, last year declared "Moore's Law is well and alive." But unlike in the past, Moore's Law steps now impose new costs for companies that want to employ the most advanced manufacturing processes.
Microsoft's Xbox One in 2013, Xbox One X in 2017, and Xbox Series X coming this year all have chips about the same size, which in the past would have meant that the chips cost about the same price. Now, though, "it's significantly more expensive for the newest one," said Microsoft chip designer Jeff Andrews.
Another challenge besides cost is that new chips often only accelerate specific computing operations. That's useful for tasks like artificial intelligence and graphics, but it makes life harder for software programmers who have to reckon with processors that work in different ways.
Intel is trying to bridge this chip divide with a new software layer it calls oneAPI. It's a notable move: Intel is a hardware specialist, but it's embracing software as an essential step in making its chips useful.
"Increasingly, hardware architecture teams need to be comprised of software experts," Koduri said.
New chip ideas
At Hot Chips, processor makers also detailed a host of innovations. Among the biggest:
- Intel's Tiger Lake processor uses a new incarnation of power-saving technology called DVFS, or dynamic voltage and frequency scaling. Different parts of the chip can run faster for high-priority tasks or slower to save power. Intel now juggles the priorities between its multiple processor cores, the memory system and the communication fabric that connects it all together.
- AMD's competing Ryzen 4000 series chips, code-named Renoir and arriving now in PCs, are the first chips with eight processing cores for super-thin laptops. AMD had initially planned a six-core design but realized a careful design could accommodate eight for better performance on tasks like video and photo editing, said architect Sonu Arora. They use half the power for a given performance level as their predecessors.
- IBM's Power10 processors, which have 18 billion transistors and are due in massive Unix servers arriving next year, can be ganged together into a single powerful server with as many as 240 processing cores. In addition, a "pod" of interlinked servers can share as much as 2 petabytes of memory. That's useful for massive business computing challenges like data mining and managing inventory databases.
- Startup Lightmatter unveiled its Mars chip for accelerating AI work like image recognition. It marries about a billion conventional transistors with tens of thousands of components that use light instead of electricity to transfer data and perform calculations. The idea behind this photonic technology is to cut power usage.