jtr1962
Storage? I am Storage!
I'm sure everyone here is intimately familiar with Moore's Law, which basically states that computing power will double every 18 months or so. Of course, it was never a law in the strict sense, but merely a means of predicting the evolution of computers which thus far has been fairly reliable.
I've recently come to the conclusion that, short of a breakthrough that makes it out of the lab with record speed, Moore's Law will end by the year 2005, at least with regards to CPUs. Variously advances in storage have ensured that Moore's Law should continue in that area for at least another decade before bumping into inherent limits of their own.
My conclusion is not based on the usual "at some point we will reach the physical limits of silicon" but rather on the issue of power dissipation. What led me to start thinking about this was rather interesting. For several years experimenting with thermoelectric modules has been a hobby of mine. Since large(40mm x 40 mm) modules can put out in excess of 100 watts of heat, one of the initial problems I faced was finding suitable heat sinks that would give a low(<10° C) temperature rise above ambient while dissipating a hundred or so watts. Generally, I ended up purchasing fairly large aluminum extrusions of perhaps 5" x 12" x 2 " high and using a 120mm, 120 CFM fan, or making my own copper liquid heatsinks. I had never given much thought to using smaller heatsinks designed for microprocessors until recently, and I have Moore's Law to thank for it. When I first began this hobby in 1995 or so, microprocessors such as the Pentium were dissipating ~10 or 20 watts, and got by with fairly small heat sinks that were totally unsuitable for large TE modules. It was only recently that microprocessors like the Athlon and P4 began approaching the power levels of TE modules, and therefore needed similar heatsinks, but in a much smaller package, of course. I recently purchased a P4 heatsink to test for such purposes, and was frankly impressed by its performance, which was suitable for anything up to a 50W TE module. Better yet, the price($11.67) was half what a large extrusion and fan would cost me, and the size was about one third that of an extrusion with similar thermal performance. Since processors were only going to get more powerful(and thus need even better heatsinks), it seems that my TE cooling requirements for the foreseeable future are satisfied, and at a very reasonable price.
This insight combined with something I read in Electronics Design about high performance ICs approaching 200W by 2005, and with this article:http://www.electronics-cooling.com/html/2000_jan_a2.html. It then dawned on me that we are probably about two or three years from a big road block, and no solution is in sight. For years, transistors have been scaled down in size and core voltages have dropped. Both of these tend to decrease power dissipation. However, the trends towards more transistors and ever higher clock rates have conspired to produce a net increase in power consumption despite the lower voltages and smaller features. This will lead to several problems:
1)The ever increasing power is being removed from an ever smaller die area, resulting in an increased power density. Diamond dies and carbon nanotubes can overcome this problem, most likely, for a while, but sooner or later there will come a time that the power density(W/mm²) is so high that the temperature rise will be unacceptable, even with an infinite heat sink. For an analogy, imagine 100 W trying to leave an area the size of a few square mm. No material can remove that much power from so small an area while keeping it's temperature below, say, 65° C, except possibly with huge amounts of active cooling.
2)Besides the power density increasing, the absolute power being dissipated is also increasing. It is clear to me that for a variety of reasons, 150W or so is the upper limit to what a home PC processor will be allowed to dissipate.
3)The power supply must put out this power, and combined with other peripherals and the M/B, we are likely in the 300W to 400W range, which will affect supply reliability.
4)The effect of millions of new PCs that draw 200W more than their predecessors will be such that perhaps the equivalent of 20 or so 1 GW power plants will need to be added in the US alone at a time when we are under increasing pressure to cut back atmospheric emissions, and cannot build new nuclear plants due to NIMBYism. Clock throttling can only accomplish so much, and it will be difficult to justify replacing a machine that is basically used for web surfing and word processing(at least for the masses) with another consuming twice as much power.
5)Even assuming 1) thru 4) aren't show stoppers, the fact that the power must be removed from the PC will be. Given that PCs are unlikely to be any larger(probably the opposite), or any louder(again, probably the opposite), there are inherent physical limits as to how much heat can be removed with a heat sink of a given volume. First of all, the heat sink is already at a disadvantage since it is using the hotter air in the case, and with the general increase in power this air will be hotter still unless we have more noisy airflow to increase the air exchange rate inside the case(basic law of physics). Assuming we can keep the air in the case at 35°C, we still need to keep the microprocessor under 50° C for long term reliability and stability. For a processor dissipating 200 W, this implies a heat sink with a thermal resistance of (50°C - 35°C)/200W or 0.075° C, probably in a package not much larger than the P4 heatsink that I purchased(which incidentally had a manufacturer's rating of 0.26°C/W). Given that the heat sink I purchased incorporated nearly every trick in the book(thin, closely spaced fins, thick baseplate, powerful fan), I am at a loss to figure out exactly how the performance will be improved by a factor of nearly four. A tip magnetic fan will get you another 15%, perhaps optimizing the fin spacing another 15%, making the thing out of copper ~30%. As Scotty said to Captain Kirk: "I can't change the laws of physics", and that is the problem I see here.
6)Given the cost constraints and mass production problems, liquid cooling is not an option, and even if it were, it would only buy another year or two before the heat problem would rear it's ugly head again. You can mount a fairly large highly efficient liquid to air heatsink on top of the PC case and use ambient air to cool it, and then use the cooled liquid to cool the CPU with a copper liquid heat sink. The total performance of such a setup will be at best around 0.05° C/W, and since you will be using ambient(25°C) air, rather than heated case air, your CPU power dissipation can be (50°C - 25° C)/0.05°C/W or about 250W to keep the CPU temperature under 50° C. And that's it! So even with exotic cooling solutions, and putting aside all the other considerations, 250W is the physical limit that any passive cooling system will handle, and we'll be there within 3 years.
7)Active cooling is definitely not an option. Putting aside the usual cost and mass production concerns, the usual device used for active cooling(the Peltier or thermoelectric module) is grossly inefficient, and compressors are too heavy, expensive, and noisy to even consider. A Peltier module is about 10% to 25% as efficient as a compressor, depending upon the temperature differential, but is nevertheless used in many niche aplllications where simplicity of design and size outweigh increased power consumption. In general, you're lucky to get a COP(coefficient of performance) of even 1 with TE modules, meaning that if you need to remove 100 W of heat, you need to power the modules with 100 W of power. So for your 200W microprocessor you need to supply another 200W to power the TE modules, and you need to somehow remove all that heat from the computer case. Now we're starting to approach the power of a space heater.
8)It is not foreseeable that silicon ICs will be able to operate reliably at higher temperatures than 50°C. If it were, then the cooling requirements mentioned above would be immaterial. We could just use a regular P4 heatsink and standard case cooling, turn it on, watch the internal case temperature rise to ~60°C and the CPU die rise to ~115°C. Unfortunately, as transistors shrink they become more suspectible to thermal diffusion at higher temperatures, so if anything newer CPUs will be even more sensitive to temperature than their predecessors, not less. For example, I can happily operate a MOSFET at 125° C but wouldn't dare try it with a CPU.
By now I'm sure you're either beginning to see that the picture is looking rather gloomy 3 years down the road, or you know something I don't. The problem is simple-using current and foreseeable technologies, there is no way to increase computing power without also increasing heat production. Even multiprocessors are not a solution. Each generation of CPU is more efficient than its predecessor in terms of computing power per watt, so using multiple cooler running CPUs will consume more power, and occupy far more space. Thinking about this, I would need, say 4 PII-450s to equal the performance of a P4 2 GHz machine, and those 4 PII CPUs would consume in total about 140W versus 70W for the P4, and occupy more space to boot.
So that's it, folks. Sort of a major breakthough, which will likely not get out of the lab in time, I think our machines will reach a plateau in a few years, perhaps at around 5 GHz. Oh, did I mention there are also a myriad of other problems designing circuits operating at those near microwave frequencies?
I've recently come to the conclusion that, short of a breakthrough that makes it out of the lab with record speed, Moore's Law will end by the year 2005, at least with regards to CPUs. Variously advances in storage have ensured that Moore's Law should continue in that area for at least another decade before bumping into inherent limits of their own.
My conclusion is not based on the usual "at some point we will reach the physical limits of silicon" but rather on the issue of power dissipation. What led me to start thinking about this was rather interesting. For several years experimenting with thermoelectric modules has been a hobby of mine. Since large(40mm x 40 mm) modules can put out in excess of 100 watts of heat, one of the initial problems I faced was finding suitable heat sinks that would give a low(<10° C) temperature rise above ambient while dissipating a hundred or so watts. Generally, I ended up purchasing fairly large aluminum extrusions of perhaps 5" x 12" x 2 " high and using a 120mm, 120 CFM fan, or making my own copper liquid heatsinks. I had never given much thought to using smaller heatsinks designed for microprocessors until recently, and I have Moore's Law to thank for it. When I first began this hobby in 1995 or so, microprocessors such as the Pentium were dissipating ~10 or 20 watts, and got by with fairly small heat sinks that were totally unsuitable for large TE modules. It was only recently that microprocessors like the Athlon and P4 began approaching the power levels of TE modules, and therefore needed similar heatsinks, but in a much smaller package, of course. I recently purchased a P4 heatsink to test for such purposes, and was frankly impressed by its performance, which was suitable for anything up to a 50W TE module. Better yet, the price($11.67) was half what a large extrusion and fan would cost me, and the size was about one third that of an extrusion with similar thermal performance. Since processors were only going to get more powerful(and thus need even better heatsinks), it seems that my TE cooling requirements for the foreseeable future are satisfied, and at a very reasonable price.
This insight combined with something I read in Electronics Design about high performance ICs approaching 200W by 2005, and with this article:http://www.electronics-cooling.com/html/2000_jan_a2.html. It then dawned on me that we are probably about two or three years from a big road block, and no solution is in sight. For years, transistors have been scaled down in size and core voltages have dropped. Both of these tend to decrease power dissipation. However, the trends towards more transistors and ever higher clock rates have conspired to produce a net increase in power consumption despite the lower voltages and smaller features. This will lead to several problems:
1)The ever increasing power is being removed from an ever smaller die area, resulting in an increased power density. Diamond dies and carbon nanotubes can overcome this problem, most likely, for a while, but sooner or later there will come a time that the power density(W/mm²) is so high that the temperature rise will be unacceptable, even with an infinite heat sink. For an analogy, imagine 100 W trying to leave an area the size of a few square mm. No material can remove that much power from so small an area while keeping it's temperature below, say, 65° C, except possibly with huge amounts of active cooling.
2)Besides the power density increasing, the absolute power being dissipated is also increasing. It is clear to me that for a variety of reasons, 150W or so is the upper limit to what a home PC processor will be allowed to dissipate.
3)The power supply must put out this power, and combined with other peripherals and the M/B, we are likely in the 300W to 400W range, which will affect supply reliability.
4)The effect of millions of new PCs that draw 200W more than their predecessors will be such that perhaps the equivalent of 20 or so 1 GW power plants will need to be added in the US alone at a time when we are under increasing pressure to cut back atmospheric emissions, and cannot build new nuclear plants due to NIMBYism. Clock throttling can only accomplish so much, and it will be difficult to justify replacing a machine that is basically used for web surfing and word processing(at least for the masses) with another consuming twice as much power.
5)Even assuming 1) thru 4) aren't show stoppers, the fact that the power must be removed from the PC will be. Given that PCs are unlikely to be any larger(probably the opposite), or any louder(again, probably the opposite), there are inherent physical limits as to how much heat can be removed with a heat sink of a given volume. First of all, the heat sink is already at a disadvantage since it is using the hotter air in the case, and with the general increase in power this air will be hotter still unless we have more noisy airflow to increase the air exchange rate inside the case(basic law of physics). Assuming we can keep the air in the case at 35°C, we still need to keep the microprocessor under 50° C for long term reliability and stability. For a processor dissipating 200 W, this implies a heat sink with a thermal resistance of (50°C - 35°C)/200W or 0.075° C, probably in a package not much larger than the P4 heatsink that I purchased(which incidentally had a manufacturer's rating of 0.26°C/W). Given that the heat sink I purchased incorporated nearly every trick in the book(thin, closely spaced fins, thick baseplate, powerful fan), I am at a loss to figure out exactly how the performance will be improved by a factor of nearly four. A tip magnetic fan will get you another 15%, perhaps optimizing the fin spacing another 15%, making the thing out of copper ~30%. As Scotty said to Captain Kirk: "I can't change the laws of physics", and that is the problem I see here.
6)Given the cost constraints and mass production problems, liquid cooling is not an option, and even if it were, it would only buy another year or two before the heat problem would rear it's ugly head again. You can mount a fairly large highly efficient liquid to air heatsink on top of the PC case and use ambient air to cool it, and then use the cooled liquid to cool the CPU with a copper liquid heat sink. The total performance of such a setup will be at best around 0.05° C/W, and since you will be using ambient(25°C) air, rather than heated case air, your CPU power dissipation can be (50°C - 25° C)/0.05°C/W or about 250W to keep the CPU temperature under 50° C. And that's it! So even with exotic cooling solutions, and putting aside all the other considerations, 250W is the physical limit that any passive cooling system will handle, and we'll be there within 3 years.
7)Active cooling is definitely not an option. Putting aside the usual cost and mass production concerns, the usual device used for active cooling(the Peltier or thermoelectric module) is grossly inefficient, and compressors are too heavy, expensive, and noisy to even consider. A Peltier module is about 10% to 25% as efficient as a compressor, depending upon the temperature differential, but is nevertheless used in many niche aplllications where simplicity of design and size outweigh increased power consumption. In general, you're lucky to get a COP(coefficient of performance) of even 1 with TE modules, meaning that if you need to remove 100 W of heat, you need to power the modules with 100 W of power. So for your 200W microprocessor you need to supply another 200W to power the TE modules, and you need to somehow remove all that heat from the computer case. Now we're starting to approach the power of a space heater.
8)It is not foreseeable that silicon ICs will be able to operate reliably at higher temperatures than 50°C. If it were, then the cooling requirements mentioned above would be immaterial. We could just use a regular P4 heatsink and standard case cooling, turn it on, watch the internal case temperature rise to ~60°C and the CPU die rise to ~115°C. Unfortunately, as transistors shrink they become more suspectible to thermal diffusion at higher temperatures, so if anything newer CPUs will be even more sensitive to temperature than their predecessors, not less. For example, I can happily operate a MOSFET at 125° C but wouldn't dare try it with a CPU.
By now I'm sure you're either beginning to see that the picture is looking rather gloomy 3 years down the road, or you know something I don't. The problem is simple-using current and foreseeable technologies, there is no way to increase computing power without also increasing heat production. Even multiprocessors are not a solution. Each generation of CPU is more efficient than its predecessor in terms of computing power per watt, so using multiple cooler running CPUs will consume more power, and occupy far more space. Thinking about this, I would need, say 4 PII-450s to equal the performance of a P4 2 GHz machine, and those 4 PII CPUs would consume in total about 140W versus 70W for the P4, and occupy more space to boot.
So that's it, folks. Sort of a major breakthough, which will likely not get out of the lab in time, I think our machines will reach a plateau in a few years, perhaps at around 5 GHz. Oh, did I mention there are also a myriad of other problems designing circuits operating at those near microwave frequencies?
Does anyone remember all those endless jokes about computerised toasters? Maybe it's not so silly as it sounds!
