If a video card has a thermal throttling circuit, when it overheats it will change its clocks and voltage to lower levels until it reaches normal operating temperatures, and then return to normal. The silicon doesn't automatically run faster based on temperature.
Not true. Dynamically changing voltage and frequency is hard to handle and hard to estimate thermal load. When VRM overheats it simply idles the card for some % of clock cycles bringing average current and thermal output back inline. Far faster, far simpler, and has far less unforseen consequences than trying to dynamically adjust voltage and frequency. When current protection kicks in you will see this as a drop in load % (which is actually load average over one second). If you had milisecond resolution on your graph you would see it as load of 100% then drops to 0% periodically then back to 100%. That averages out to 99%, 97%, 95%, or 90% over one second.
) 
