Many times we are hearing in recent times, especially due to the use of ARM architecture by Apple and Intel Foveros, about the big.LITTLE architecture in processors, where “small” cores of low consumption coexist in charge of the lightest tasks and “Big” cores to unleash all the power. At the moment we have only seen this architecture in CPUs but, would it be possible to implement it also in GPUs for graphics cards ?
The concept of this architecture is very simple and at the same time effective: we have a die with two types of cores, some small ones that are always active and that take care of light tasks such as browsing the Internet or writing in Word, and others Larger and more powerful cores, which are normally off and waiting for a task that requires more power to run to wake up and take over. This big.LITTLE architecture is used in processors, where the core count is quite low, especially when compared to a GPU, where the number of cores is in the thousands .
NVIDIA Optimus was already a kind of big.LITTLE on GPU
Many of you will remember the days when almost all laptops with a dedicated GPU had NVIDIA Optimus technology. This technology meant that when we were at the desk or doing tasks that were not 3D, the graphics card integrated in the processor was used to save energy, but when we executed a game or 3D task then the dedicated graphics was used to be able to have all the power available.
This clever way of taking advantage of the fact that the notebook had two GPUs is, in essence, a kind of big.LITTLE but in crude mode (instead of being an entire architecture), but the idea is basically the same: when the GPU is not needed “big” is off and on standby, with the iGPU doing all the work while it can, saving power and reducing the heat generated. When needed, the “big” GPU kicks in and delivers peak performance.
So is it something that can be implemented?
Of course, the idea is very good, but either they integrate two GPUs on the same PCB for this same functionality that we have talked about before or things get quite complicated, and it is not the same to work with the 4-16 Cores a CPU can have than with the several thousand GPUs (to put this in context, a Radeon RX 6800 XT has 3,840 Shader Processors (cores), while an RTX 3090 has a whopping 10,496 CUDA cores).
For a full implementation of this architecture, it would be necessary to redefine the way the GPU works, since currently they already have two types of ALUs: one is used for simple instructions and whose consumption is very low, while others (SFUs) are responsible for perform the more complex operations such as square roots, logarithms, powers, and trigonometric operations. They are not big.LITTLE kernels, but in fact they are already called differently (FP32 ALUs) precisely for this reason.
So you see, actually both in the past with Optimus and today something similar to big.LITTLe is already implemented in the GPU, what happens is that it is obviously different, it works differently and it is not called that. However, since in recent times everything seems to be focused on greater efficiency in terms of consumption, it is something that we cannot rule out, and of course both NVIDIA and AMD have the potential to carry it out. It would, of course, be an ideal situation to save energy and generate less heat, right?