“We are stuck with silicon for the manufacture of circuits for at least another 7-10 years”

In the past, it was normal for a new manufacturing technology with doubled transistor density and lower costs, higher clock frequencies and lower voltage to be released every two years. This is a development that has gradually slowed down for a little more than a decade, which makes it increasingly difficult to develop better and more high-performance circuits.

During a conference on high-performance data centers (High Performance Computing), Forrest Norrod, AMD’s head of data centers and embedded solutions, talks about the challenges facing the semiconductor industry. The first has to do with performance, where he notes that frequency scaling has long since stopped.

On the consumer side, Intel has started to deliver individual models with frequencies closer and up to 5.0 GHz, but then only with a few cores with rampant power consumption as a result. In servers and data centers where many cores are commonplace, Norrod says that the frequencies have reached a peak around 3-4 HHz and that in the future they may take smaller steps backwards.

The same trend is also beginning to be felt in transistor shrinkage, but so far there has been a steady stream of new and more transistor-tight manufacturing techniques. Here, instead, the problem is that manufacturing techniques are becoming increasingly expensive and that some parts cannot be shrunk as much as others.

When Norrod predicts the future and the next step, he believes that it will take at least 7-10 years before other materials, such as graphene, replace silicon. In the short term, he sees that silicon has more to give two more generations, where the step after 7 nanometers is 5 nanometers and the second generation he is talking about is 3 nanometers.

I think there’s a clear path on using traditional silicon to about 3 nanometer geometry. So another couple of process nodes. Which we’ll get to in about, probably about 5 to 6 years. I think after that it gets a little bit more, it gets a little fuzzier. Obviously we got the dice rolling on quantum off to the side. I am sure that quantum will mature in 10 to 100 years. It’s not 10, it’s probably more like 20 or 30. So I think we’re stuck on silicon here for about, at least the next 7 to 10 years. And I don’t believe anyone who tried to predict any longer than that sum.

After 3 nanometers were launched in about 5–6 years, Norrod believes that it will be more unclear for the years to come. The closest at hand is a successor to silicon, while quantum computers are still in their infancy. However, he sees it only as a matter of time before the technology matures, but that it is 20-30 years away in time.

Instead of relying entirely on new manufacturing techniques to produce better and more high-performance products, Norrod sees other solutions as the way forward. Not entirely unexpectedly, he cites AMD’s upcoming Epyc processors as an example, where eight computing circuits of 7 nanometers are connected through a large circuit with memory controllers and other input / output (I / O) of 14 nanometers.

The step after this will be so-called 3D circuits, where circuits made on different techniques are stacked on top of each other. This is something Intel intends to release later in the year with Lakefield, where the computing part itself (CPU and GPU) is manufactured at 10 nanometers.

An energy-optimized variant of 22 nanometers houses memory controls, I / O and cache memory that do not always scale well and can even work better on a larger node. On top of this, Intel will stack DRAM circuits.

Read more about future manufacturing techniques: