There is no doubt that the cost of electricity will increase due to the greed of politicians and their measures, easy to say, but they are made with all our money. However, we do not talk about these problems and although we know that the electricity bill has increased artificially, what interests us is to explain why graphics cards consume more and more, like processors, RAM memories. That is, all the components inside our PC.
That new chips increase the bill is inevitable
If we look at any processor architecture, general or specialized, we will realize that more than half of the circuits, if not more than two thirds, do not exist to process data, but to communicate it between the different parts . And we must not forget that today’s chips can have one or more functions at the same time, but they can be summarized in three main functions:
- To process data.
- transmit data.
- Store data.
Well, even though it may seem counterproductive to many, the act of processing information comes at a cost both in the number of transistors inside the processor and in really negligible energy cost. And a big part of designing processors is getting information to execution units so they can process it. The problem is that due to the laws of physics it is more expensive to transport bits than anything else today and one of the consequences is that chips increase the electricity bill, obviously consuming more.
It’s called the Von Neumann wall, because it’s inherent in all architectures, whether it’s a server, a mobile phone, or a video game console, and c t has become the biggest headache for engineers today. And even more with the need to reduce the carbon footprint.
Explain the problem quantitatively
Normally, we usually give chips power consumption in watts (W), which is joules per second. Since bandwidth is in bytes or bits per second, so as not to confuse the two terms, one way to measure a simple data transfer is to see how many joules it consumes. Well, we owe the following graph to Bill Dally, chief scientist at NVIDIA and one of the world’s leading experts in computer architectures:
Well, we have to start from the fact that one nJ or nanoJoules is worth 1000 pJ or picoJoules. That is, if when processing a double-precision or 64-bit floating-point arithmetic operation, depending on where the data is located, the power consumption to perform the same operation will vary:
- If the data is in the records, it will only cost 20 pJ or 0.02 nJ.
- If we have to access the cache to find them, the thing goes up to 50 pJ or 0.05 nJ
- On the other hand, if the data is in RAM, then the consumption rises to 16 nJ or 16000 pJ.
In other words, access to RAM costs 1000 times more energy to perform the same operation as the information contained in the processor. If we add to this the internal intercommunication between the components of a processor and the external, we end up having chips capable of processing large volumes of data, but at the same time they end up needing a large amount of energy to operate.
What solutions will we see in the future?
At the moment, these are only lab solutions, but they have been proven to work and could change the way we understand PCs. Mainly, we have two different solutions to the problem.
Near-memory processing
The first of these is near-memory processing, which involves bringing the memory closer to the processors. The idea is none other than to put memory very close to the processor, unfortunately we cannot place the tens of gigabytes of RAM that we will see in the short term in PCs, but a large level of additional cache which will greatly increase the power capacity to find the data and reduce the consumption.
Interestingly, the strategy of increasing cache sizes is the one that NVIDIA pursued with its Lovelace architecture, increasing L2 by 16 from one generation to the next. It’s not enough. Which brings us to the conclusion that a new level of memory will be needed. Closer to the processor and, therefore, with less power consumption. In other words, in a few years we will talk about fast RAM and slow RAM on our PC. The latter possibly using the CXL interface.
In-memory processing
The second thing is what we call PIM, they are not processors themselves, but memory chips with internal processing capabilities. That is, they are still memories, but some specific algorithms can be executed on them. For example, imagine we need to make several queries to a database that is in RAM to find a customer’s email address. In the classic mechanism, several accesses to a low-power external memory would be required. In this way, instead, the RAM itself looks for the data itself, with very little consumption since it does not need to be consulted from the outside and only has to send data to the processor.
In this way we are able to significantly reduce the amount of transfers between RAM and processor and thus reduce energy consumption. The trade-off is that applications must be designed for this paradigm. Although this is necessary to reduce power consumption and prevent new chips from increasing the electricity bill further.
Table of Contents
