One of the great improvement in the more demanding applications of consumer PCs has been in splitting general computation (done by the CPU) from graphical computation (done by the GPU) with physics sometimes being done by a GPGPU.
As well, turbo boost allows a multicore CPU to have one of its core run above its usual frequency.
Here’s some speculation I’d like to have feedback on:
The processor that acts as the CPU needs to have as high a potential frequency as wattage and cooling systems will allow since it affects overall performance and acts as a catch-all processor for any tasks not assigned to other processors.
On the other end of the scale, graphics is highly parallelizable. A GTX Titan Z has 5760 processors and an (overly enthusiastic) user can use 4 of them at the same time for a total of 23 040 processors. Presuming they’re using a 3840 X 2160 resolution, that means 1 processor for every 360 pixels*. Within 1 or 2 decades, each pixel may have its own assigned processor running below 100MHz.
In-between those two ends of the spectrum, there may be different levels of amenability to parallelization. For example, physics may benefit from having a frequency and a number of processors that sits in-between (on a log-scale) the CPU and GPU.
What other processes would benefit from parallelization?
Might we end up with all-in-one systems with 1 level0 central core running at 5GHz, 1000 level1 cores running at 500MHz and 1 000 000 level2 graphics cores running at 50MHz? Or even more tiers to leverage different levels of parallelizability and ultimately reduce required wattage and heat?
Will liquid cooling, graphene or extensive use of copper** become common at the consumer level? If we compare 90s CPU cooling to 2010s CPU cooling, it seems that way. Something like a graphene layer right next to the processors, thick copper leading out of the main case to wafers of aluminum in water with fans as the last step?
Related questions:
Am I correct that performance improvement for a given technology level is chiefly limited by power requirement and heat generation? IOW, they’re the bottleneck in boosting performance?
For a given technology level, what’s the approximate relationship between processor frequency and required wattage & generated heat? Is it close to a square or cubic function? Some other function? Does the same hold for the cost of producing the processor?
- Yes, I know graphics isn’t only pixels but this gives up an idea of the parallelization involved.
** As opposed to a timid use of a plate and a few thin tubes that goes through aluminum plates.