What is Dark Silicon and what is it currently used for?


You can also be interested in these:


A term that surely seems curious to you is Dark Silicon, we reckon it seems taken from a science fiction movie. However, it is a phenomenon that has been occurring recently in all processing units such as TPUs, CPUs and GPU. It is necessary that you know everything about this phenomenon and above all, if your processor can be a victim of it.



What is the phenomenon known as Dark Silicon?

At first, processors were manufactured with the firm intention of performing any type of task without resting, with the idea that all resources would be used for a long time. However, the microarchitecture of today’s processors forces some transistors to stay in a constant standby state.

Not many decades ago, only a few transistors were put to sleep as part of the normal functionality. But today, there can be even thousands that remain idle in the different types of processors. This would not normally be a problem if the transistors did not have a consumption, or at least, the consumption that they had was decreasing with the advances in technology.

What is the problem that this trigger? A terrible one: Processors that have as much consumption as the necessary to keep a light bulb of more than 100W on, at least in the case of a traditional CPU. If we go to the field of GPUs, these can have such a demanding consumption as to light a home with several ordinary light bulbs.

Due to this phenomenon, it is necessary for microprocessor manufacturing companies to turn off certain sections of transistors, both to avoid excessive consumption with frequent use of the computer, and to avoid temperatures that can represent a challenge for current cooling systems.

These areas of idle transistors are known as Dark Silicon. Now, is Dark Silicon an ally or an enemy?

A large-scale solution: Keep multiple sections off

Turning off the transistors when there is no resource requirement saves power and prevents overheating. This is a long-standing solution to microchip manufacturing, leaving large sections of transistors unused. This keeps the temperature from being excessive and at the same time allows the other sections of the processor to work at full capacity.

A performance issue

In the original theory it was better to have all the transistors working for better performance, which cannot be done today due to problems with overheating and excessive power consumption. This generates the problem that the performance will not be the maximum that could be obtained. Despite this, the “extra” transistors continue to be manufactured “just in case” they have to be used at some point when the system requires. This represents a larger surface area of ​​the microchip, a greater number of components necessary to manufacture them and a higher price.

What is Dennard Scaling?

Dennard Scaling’s law is simply the process on which transistors are reduced in size as technology makes it possible. Formerly there were micrometers, that represented a technological advance. Today there are nanometers and thus progress is being made with miniaturization further on.

Dennard Scaling’s Law works in conjunction with Moore’s Law and Pollacks Law, which make possible the scalable miniaturization process that the computing industry has been using in recent decades. But it is important to know that as transistors are miniaturized, there will also be more transistors that are not used, because there cannot be thousands of transistors working at the same time for the reasons we have already explained.

The law itself was invented in part by Robert Dennard, who first applied it to the well-known Mosfet-type transistors, a theory that was later applied to FinFET transistors and so on until today.

In itself, the theory proposes that the current and voltage can decrease as the size of the transistors also decreases, but this does not mean that the clock frequency does, as it can increase without causing consumption to skyrocket a significant way. This at least a few generations of processors ago, since today consumption has been increasing significantly.

Since 2016, a solution to this problem has been investigated, using MCM packaging and chiplets as a testing ground, including the new 3D packaging technology.

Is there a solution for Dark Silicon?

If you like the news regarding the world of technology, surely you have seen in some reviews that a certain processor is using a technology free of Dark Silicon or at least considerably reducing its effects. This happens mostly on ARM-type processors, which found a way to avoid “dark silicon”. This is how they do it:

1. Implement the use of coprocessors

By implementing the use of dedicated coprocessors, you can increase the efficiency of the main processors. This happens because the data is processed better if it is floating point type and not as part of an integer pipeline. This can cause FPU SIMD type processors, and their instruction set to be debugged.

2. Heterogeneous cores

A group of powerful cores and another group of less powerful cores but energy savers. This is the technology called big.LITTLE that allows you to use all the transistors but save processing power when you don’t need to do large amounts of calculations. Windows 11, for example, identifies tasks by total processing weight and assigns them to the corresponding group of cores.

3. Accelerators

Sometimes the manufacturer includes software that identifies the tasks so that they can be attended to with the appropriate priority. Currently there are SoCs that are using it in TPUs and DSPs and thus optimize the tasks that are performed.

All these proposals have their advantages and disadvantages. For example, as the main disadvantage we have, in all three cases: A highly complex microchip would be needed, which would make it more expensive and even with more transistors that can perform each type of task, however, this has demonstrated not be the case.

The solutions presented have shown that they can make all the transistors work without increasing the consumption and temperature of the processor.

Conclusions about Dark Silicon in current processors

We have seen how the processing capacity of PC components has increased over the years. Also, that the arrival of 3D Packaging and the MCM is close. This means that the miniaturization process continues to advance as well as the number of transistors in the CPU continues to increase. The good news is that there are already several solutions that are being put into practice to avoid Dark Silicon and take advantage of the full capacity of a processor in an efficient and economical way.

We believe that dark silicon will no longer be a problem for the next decade as manufacturers are taking measures so that in the next 5 years the processors are more efficient but at the same time with greater savings and less overheating. However, synchronizing the microarchitecture planning process, manufacturing and meeting future expectations is something very difficult to achieve despite the great advances in this field of microchip manufacturing.


More stories like this