One of the problems when scaling a processor is related to its power consumption and sometimes it is necessary to leave parts of it completely inactive. Something that has to do with black silicon and especially when making processors and GPUs more complex in terms of the number of cores.
What is black silicon?
The term Dark Silicon refers to the parts of the processor that need to be turned off at some point for others to run smoothly, this phenomenon is one of the headaches when designing new processors and GPUs. This is at least a “recent” phenomenon, since it did not exist a few years ago and therefore we could have 100% of the surface area of a processor active all the time.
Moore’s Law scaled parallel to Dennard’s scaling, which was based on the fact that as transistors got smaller, the consumption density per unit area has remained constant between node and node. Which in plain language translates as follows: As a transistor gets smaller, its consumption decreases in proportion to the reduction in size.
With the Dennard scaling already broken for fifteen years, the architects find themselves with problems to maintain a consumption density in line with the downsizing of the transistors, one of the solutions being the so-called dark silicon or black silicon. . Which It is based on leaving parts idle in the processor and hence they do not receive external power to function.
How does this affect the design of new processors?
If we look inside any processor, we will see that they are made up of different blocks, so a very common design tip is to divide the area of the processor into different areas where each one receives a pulse of different energy, from that. allows whole parts of the processor to be disconnected when not in use and to distribute power more efficiently.
The problem with black silicon is that as we make processors more complex in terms of the number of cores, we have to add more and more functions to turn off those parts of the processor at some point. Often times, Dark Silicon are not whole units, but for example, there may be instructions that use different parts of the processor than others, and these may be turned on and off depending on the instruction being executed. .
But, which mainly affects the density dedicated to black silicon. This is because as we increase the number of cores, more surface area ends up taking up the black silicon of a processor in total. So this is the main limitation when scaling both CPUs and GPUs in terms of the number of cores.