Can a CPUs speed be increased by (physically) upgrading the CPU clock?

I don't know much about CPU operation, but I'm learning. I read somewhere that making a CPU work faster is as simple as increasing the clock speed (one method among many), and that the real limiting factor being electron travel speed in the circuits, which wont be met until extremely high THz rates, or well over PHz.

My question is what would prevent me from speeding up a CPU if I replace the clock with one that ticks at 1+ THz instead of the typical few GHz for modern processors? Assuming I manage to completely prevent heat build up, would such a simple change to the processor boost the speed like this? If not, what other factors have to be considered?

• This is a question about the behaviour of electronic components, rather than about computer science per se. I think Electrical Engineering would be a better place. – David Richerby Nov 16 '14 at 17:14
• At high clock speeds, it becomes impossible to synchronize across the whole chip. Current chips are about 45mm across so it takes light about 0.15ns to cross the chip; a terahertz clock ticks 150 times in that interval. Very high clock speeds bring many problems. – David Richerby Nov 16 '14 at 17:58
• Even with liquid nitrogen cooling and a substantial voltage increase AMD's FX 8370 was only able to be overclocked to just over 8.7 GHz (source) from a base 4.0 GHz (with all 8 cores working). The Wikipedia article on overclocking presents some of the considerations constraining overclocking. From CNN on an earlier record: "We destroy motherboards, processors, and graphics cards at an alarming rate" – Paul A. Clayton Nov 16 '14 at 20:13
• [sarcasm] Wow, you've hit upon a great idea. Nobody ever thought of that before! Go tell Intel all they have to do is increase the clock frequency to make faster chips and you'll be rich [/sarcasm]. – Olin Lathrop Nov 16 '14 at 23:28
• Speed is a key marketing point for CPU chips. Do you think the manufacturer would deliberately undersell one of their most important performance metrics? No, they aggressively push themselves to sell the fastest clock speed that can work reliably for a given design, over temperature and process variation. Overclocking degrades that reliability. – MarkU Nov 16 '14 at 23:33

I read somewhere [...] that the real limiting factor being electron travel speed in the circuits, which wont be met until extremely high THz rates, or well over PHz.

Pure fiction. Electron travel speed itself is relatively low. Electromagnetic wave travel speed - that is the interesting one - is in the order of the speed of light. At 1 THz - or in 1 ps (picosecond, 1e-12s) - your signal would travel just 0.3mm.

what would prevent me from speeding up a CPU if I replace the clock

The critical path would prevent you from going above a certain frequency that is usually not much higher than specified. In a nutshell this is the signal path that takes the longest time, but must be finished in one clock cycle. Once you rise the clock speed above that limit, the CPU will no longer operate correctly.

• I suspect the OP might be referring to record transistor switching speed when mentioning THz (among other things, not realizing that the processor frequency is not equal to the fastest transistor switching speed). Further explaining the critical path limit (and basic pipeline constraints) and perhaps some basic aspects of circuit speed (switching speed and transmission delay [beyond just giving speed of light travel time for a ps]) might be helpful to the OP, but the tradeoffs of answer length and accessibility (and your time!) are difficult constraints. – Paul A. Clayton Nov 17 '14 at 10:23

Really!? It never ocurred to you that if all you had to do was increase the clock frequency to make a processor run faster, that someone would have done it by now. You actually think that Intel and others with 1000s of engineers looking at this problem didn't realize that all they had to do was increase clock frequency instead of spending a few 100 G$on a new fab? Duh! No, it doesn't work that way, obviously. First, electrons actually travel remarkably slowly in electrical conductors. However, that doesn't matter. It's the propagation speed that matters. Second, signal propagation speed is not the limiting factor in modern processors. Stop and actually think about it. What's the speed of light? Even figuring signal propagation speed is half that due to the impedance of the transmission line, how long would it take a signal to cross the die? No, actually go figure it out. The real limits in today's leading edge digital logic come from having to charge and discharge the inevitable capacitance on any conductor you are trying to switch the voltage of, and the reaction time of the semiconductors. Both those cause delay from starting to drive the input of a gate until the output reaches the threshold where downstream circuitry will reliably interpret it as the intended high or low level. Each individual gate may be quite quick, delaying less than a ns. However, to do meaningful things in a processor can a number of successive stages of gates. Much of high performance processor design is reducing the gate delays in the worst case path. Often shorter delay can be traded off with complexity. For example, look at a basic adder. Each stage takes in the two bits to be added, the carry from the previous stage, and produces the output bit and the carry that is passed onto the next stage. In a basic adder, the gate delay therefore grows with the number of bits. Higher bits can't be added until the carry from lower bits is available. This basic adder is also called a ripple carry adder. There are other types of adders that have lookahead carry. These take more gates, but can add two wide numbers faster. More gates of course means more cost and more power consumption, which means more cooling required, etc. Nothing is free. This same principle of more gates to make things faster applies in many places. Then there are other things, like memory, which sometimes work on a different principle than gates, and also have their inherent delay times. Back to the original question, the point is that these gates all function at some minimum delay from when the inputs are stable until the outputs are guaranteed to be correct. One way this is guaranteed is by latching the inputs on one clock edge, then not using the outputs until a subsequent clock edge. The people that design the processor very carefully decide how fast it can be clocked over what temperature and voltage range so that all the gates have the right answers by the time those answers are used. Another limiting factor is being able to get rid of the heat. All those little parasitic capacitances being charged and discharged causes current proportional to the charging and discharging frequency. This means faster clocks cause higher current, which causes more heat, which has to be gotten rid of safely so that the chip still functions. Silicon stops being a semiconductor at around 150°C, and of course you need some margin below that. If the gates are packed too closely, clocking them too fast would cause them to get too hot to function within your capability to remove the heat. This is why in some cases you can overclock some processors by cooling them more than intended. Note that this only addresses one of the limiting factors, so you can't keep overclocking a processor no matter how much you are able to cool it. Anyway, the maximum clock rate of any processor is a complex subject. No you can't just clock it faster to make it run faster and still have it work, and no, the limiting factor has little to do with how fast electrons travel. • Your answer could have done without the arrogantly rude first paragraph. You need to remember that things that seem so basic to you are not the case to everyone. You also were not born magically knowing this information. – krb686 Nov 17 '14 at 2:28 • Furthermore, his question never implied it was possible to do just that and why hadn't Intel figured that out, his question was centered on what else sets them back, so the sarcasm is a little out of place. – krb686 Nov 17 '14 at 2:41 • Also, it's terribly cute to think that a fab costs "a few 100 G$." Intel's latest fab was budgeted at 7,000,000 G\$, which I think is slightly higher than "a few" – Cort Ammon - Reinstate Monica Jul 17 '17 at 3:35

Besides what others said, electron propagation DOES matter but not on whole circuits but on the transistors itself. As soon as the electromagnetic wave reaches a transistor it MUST move electrons (Thats the whole point of a transistor) and those might be slow. Thats why, for example, NMOS tech was faster than PMOS and CMOS of the era (circa 1970/80), because as electrons are faster than holes (carriers) the p-type transistor (used in cmos) slowed down signal propagation. The advantages of CMOS where so big that they got used anyway, and the solution came with the miniaturization of the transistors itself. There is a relationship between path delay costs and transistor size and thats usually why two processors if done on different node technologies will show different clock speeds. Smaller transistors are shorter paths for electrons to flow. When this flow reaches the metallic interconnections speeds will increase accordingly, so that the limiting factor is still the transistors that make up the processor circuit.So yes, the electron flow speed is the limiting factor for processor clocks.

Besides that, CMOS transistors uses a metal oxide between the gate and the channel, this effectively makes up for a capacitor (and makes the input impedance quite high). Intrinsic capacitance between the gate and the channel delays the speed at wich the transistor will switch, and this too add to the path delays that will set the max clock speed of a processor or digital device.