In the following years, we will see a growth of built-in accelerators, the reason is that they are cores with their own internal memory that not only take up little space, but also consume very little power during certain tasks compared to other types. processors. All thanks to the fact that they have a few m
What do we mean by movement of data?
A movement of data in computer hardware is the transfer of information from one type of memory to another, regardless of the direction in which said data is going and whether we are talking about DRAM, SRAM, NAND Flash memory, registers. , cache, etc. .
The different processors that are in the hardware of your PC do two main things, on the one hand they are responsible for processing the data they receive and on the other hand they move the data through all the memories in the hardware. This means that the two things a processor does the most are moving data, but that’s what complete hardware does the most because memories only take care of that, moving data to and from. from them.
When designing a new CPU or GPU architecture, engineers not only look for the maximum possible power within the limits of chip size and consumption, but also consider the energy budget of each of the data transactions. to be carried out in architecture. . Unsurprisingly, in the case of smartphone SoCs, they only reach quotas of 67% in terms of energy consumption with the movement of data from one processor to another.
Where data matters
When designing a new processor, the location of the data is important. The reason? Easy, the energy consumption increases the more the distance at which the data is located. Ideally, therefore, the data is processed in the memory closest to the processor, which is impossible, since the large amounts of memory required in a program cannot be placed inside a CPU or a GPU. Sure, chunks of RAM are copied to or from caches, but that also means moving and consuming data.
Until relatively recently, it didn’t matter, if we look at the graph above these lines where a hypothetical 45nm node is compared to a hypothetical 11nm node. This can be seen as if the consumption of executing an instruction with 64-bit floating point precision decreases, as does the access to its registers. The energy cost of communicating the data if it is beyond the records did not, so in the end the situation reversed in terms of design and it is not the energy cost of execution. of a calculation that matters most in the future of processors. , but the cost of moving the necessary data.
It is precisely the cost of data transfer that differentiates a smartphone SoC from a PC SoC, because the interfaces used on the PC to be able to transmit a greater amount of data require a greater amount of energy. The reality is that PCs and smartphones will require the use of new communication methods and structures in order to achieve higher performance rates in the future.
How is data consumption measured?
The way to measure it is the pJ / bit, where the consumption of a data bus is measured for each bit it transmits. To get the power consumption for one second, the amount of peak Joules is multiplied by the bandwidth, in bits per second, not bytes. The reason is that watts, W, are Joules per second.
CPU and GPU companies usually don’t talk about this kind of data because they sell the ability to process the data of their designs, unlike those who are in charge of talking about data bus. Two buses can give the same bandwidth, but be completely different in terms of power consumption due to two different values in terms of pJ / bit.
This is why, for example, GDDR6 and GDDR5 memory, although it has the same bandwidth as HBM memory, has poorer power performance. The less pJ / bit a data interface consumes, the better.