Don't we still have to worry about heat in the case when we add a lot of cores to the chip? In lecture we talked about how we started using multiple cores instead of a single core because we had reached the 'thermal limit' but in this case when we've added so many cores, I'd imagine it would still need some sort of dedicated cooling unit for the chip?
neonachronism
The biggest issue with heat dissipation is surface area. When you increase the speed of a processor by increasing the clock rate, the surface area doesn't change, but the power goes up, so the power per unit area increases. If you add more units to achieve parallelism, the power per unit area doesn't change, so you can cool similarly to a single processor. Additionally, you tend to use less powerful processors on highly parallel processors - GPU cores tend to have far lower single threaded performance than CPUs.
petrichor
@neonachronism I think that the thermal limit was a function of decreasing surface area (ie decreasing silicon transistor feature size) and increasing clock rate, since the two are linked in that to increase the clock rate, you must make the processor smaller due to the limit of the the speed of light and the requirement that across a single core, all the transistors must synchronize to the same clock. Also with the thermal limit, the biggest issue is that it follows a square law, so that if you double your clock rate (and consequently half your feature size) you at least increase your power per unit area by a factor of four.
Don't we still have to worry about heat in the case when we add a lot of cores to the chip? In lecture we talked about how we started using multiple cores instead of a single core because we had reached the 'thermal limit' but in this case when we've added so many cores, I'd imagine it would still need some sort of dedicated cooling unit for the chip?
The biggest issue with heat dissipation is surface area. When you increase the speed of a processor by increasing the clock rate, the surface area doesn't change, but the power goes up, so the power per unit area increases. If you add more units to achieve parallelism, the power per unit area doesn't change, so you can cool similarly to a single processor. Additionally, you tend to use less powerful processors on highly parallel processors - GPU cores tend to have far lower single threaded performance than CPUs.
@neonachronism I think that the thermal limit was a function of decreasing surface area (ie decreasing silicon transistor feature size) and increasing clock rate, since the two are linked in that to increase the clock rate, you must make the processor smaller due to the limit of the the speed of light and the requirement that across a single core, all the transistors must synchronize to the same clock. Also with the thermal limit, the biggest issue is that it follows a square law, so that if you double your clock rate (and consequently half your feature size) you at least increase your power per unit area by a factor of four.