This is the story of a death foretold, and at the same time, one of the greatest challenges the computing world will face in the next decade. The so-called Moore's Law, that almost magical concept that allowed us to double the number of transistors on a chip every 24 months, is running out of air. Intel's tick-tock break and AMD's delays in adopting more compact processes are strong indications of this, but now we have a sort of "official confirmation".
The Official Confirmation
The Semiconductor Industry Association, formed by Intel, AMD and GlobalFoundries, published its so-called "International Technology Roadmap for Semiconductors", whose text suggests something we already considered inevitable: As of 2021, transistors will stop shrinking. Why 2021? Because any attempt at advanced miniaturization will not make economic sense by then. 2021 would also be the year when the first 5-nanometer chips reach the market, unless a different solution appears earlier. In fact, data from Intel proposes abandoning silicon at 7 nanometers, but even with a change of material, we will reach a point where the famous "tunnel effect" will impose its authority.
The latest graphics card from Nvidia, the Titan X, has a total of 12 billion transistors. Its manufacturing is based on the Pascal architecture, which was created on a 16-nanometer node. Intel's Skylake processors, currently available, are at 14 nanometers, and the same situation will repeat with Kaby Lake, which effectively breaks the tick-tock. On average, hardware destined for mobile devices still has one or two generations of advantage, but sooner or later it will reach similar levels. The point is that we cannot keep adding transistors to a chip without making certain compromises. That brings us to the state of Moore's Law.
What Lies Ahead?
Are there alternatives? Several, to be honest. The halt of the miniaturization process does not necessarily mean a pause for the number of transistors on a chip. The manufacturer can, in theory, increase the surface area of its processors, explore three-dimensional designs, or why not return to the concept of dedicated and specialized modules. On the other hand, it would be a mistake to ignore the potential of heterogeneous computing, which many companies are already studying closely. Perhaps the life support for Moore's Law will extend another half decade, however, our wish is that it yields to something better.
