Yield per watt is one of the elements that marketers commonly use to promote their products as a performance metric. But is this metric really valid or is it a smoke and mirrors trick? We are going to explain the reason why you do not have to rely on this metric exclusively to know if one product is better than another.
One of the issues that most “concern” people is energy consumption, especially in times when climate change is an issue that has passed from the world of environmentalism to the political sphere. Which has made it a mass issue and has opened the opportunity for the sale and promotion of products that are more aware of climate change.
Performance per watt as misleading marketing
The problem with performance per watt is that it only matters when it is a metric that companies win at . If we think about it coldly, nobody buys hardware for its performance / consumption but for the ability it has to do the job for which it was designed.
When we talk about performance / watt, contrary to other rates , which are based on a fixed denominator value such as time, whether they are frames per second, floating point operations per second, etc. It is that the watt concept is variable by the manufacturers and they place the value in the denominator that interests them most.
For you to understand, the performance per watt varies according to the number of watts that we put in the denominator when making the calculation, so the same processor can have several different performance / watts.
The trick is easy to demonstrate
The amount of watts of consumption depends on three values , of which we have two of them that are not fixed and are fully configurable: voltage and clock speed, but especially voltage .
That is why the undervolt exists, which consists of lowering energy consumption by only lowering the voltage, but without lowering the clock speed . A processor or graphics card that has received an undervolt has a higher performance / watt than one that has not.
Performance per watt is not constant
When we compare rates where the denominator is time, this is treated as a constant that does not vary, one second is one second and it lasts as long as it lasts. On the other hand, within the same processor the performance per watt depends on the number of watts in the denominator.
So companies have what is called in English a “sweetspot” which is the amount of watts in which their architecture has the peak performance / watt. This leads to misleading comparisons where processors from different brands are compared at a specific consumption level.
Deceptive tricks to falsify the amount of watts
There is a trick to measuring performance by disconnecting some I / O interfaces during performance tests. I / O interfaces that are essential are seen completely disconnected in benchmarks to measure performance / watt.
Although it may seem surprising to you, the fact of moving the data consumes much more energy than executing an instruction, which is why what some processors do is have different energy domains and disconnect the domain of the I / O peripherals in most tests performance.
Another trick that is used is the fact of marking the cache as if it were RAM, in this way if the benchmark is small enough it treats the last level cache as if it were RAM, this avoids having to use the external memory interface of the processor, which is the one that consumes the most.