New Computer Breakthrough is Defying the Laws of Physics
for years we relied on Moore's Law the idea that the number of transistors on a chip doubles approximately every 2 years and it has led to exponential growth in computing power powering the digital revolution up until now however this rate is slowing down and we are bumping into some hard physical limits in this video I will break down a new computer chip which will escape the lab already this year and might break through all the limits there are many billions of personal computers and approximately 100 million servers operating worldwide and soon we could have a cloud for every person with artificial intelligence embedded everywhere and this consume tremendous amount of energy but why physicists have been asking this question for a long time but it wasn't until 1961 when Rolf Landauer at IBM discovered a surprising answer information has a cost each bit of information costs us energy and that is tied to the laws of physics particularly thermodynamics and since then it became known as a Landauer Limit which states that for every operation you perform you need to expand a minimum amount of energy in fact flipping a bit from zero to one costs us that much energy at room temperature and that's an incredibly small number typically we should not care but when you're computing billions or trillions of bits like the latest NVIDIA GPU for example it all adds up consuming lots of energy but what do we get for this energy we get information as the output the result of the operation what's interesting almost 100% of the energy in the modern computer chips are dissipated as heat so essentially wasted now just think for a moment what if we could store this energy in the system and recycle it to reuse it for future computations could we build a chip that runs at no energy at all never heating up and breaking the Landauer Limit in the times of Landauer it was neither possible nor practical however a new chip which is coming out this year might break this limit why do we care about breaking this limit in fact for the past 60 years semiconductor industry was able to exponentially reduce the amount of energy spent per bit on each operation because the transistors we're getting smaller operating voltage going down capacitor you need to charge is getting smaller so you spend less and less energy per bit on each operation and that's beautiful what's happening right now is that even though the Landauer Limit is very small we are getting very close to it right now we are just a few orders of magnitude away from it when you talk to chip designers one of the things that you will hear is that oh we are so far away from Landauer's Limit right and this is true and false at the same time because yes we're fairly far away but with CMOS we're not it's CMOS the problem in CMOS we're actually basically we're basically at the end of the road right and so for us is that sure Landauer's Limit is is super far away but the point is seem it's already finished and so what do we do we're going in this gap that is between 1 Landauer and where we are today right and where we can grow reversibly and no one else can interestingly Landauer himself found that actually it's reversibility that places a lower limit on energy consumption later on my all-time favorite Richard Feynman explored this problem in depth questioning whether it might be possible to surpass the Landauer Limit and actually he concluded that there was no theoretical minimum of energy required for computation if in the process we do not lose the bits of information essentially if we don't erase the data and that's very interesting because for the entire history of computing computer chips were built in a way that we intentionally erase the information when it's no longer needed this is the best understood by an example if we take a simple logic gate which all computer chips are constructed from an end gate for example it's a very simple structure that combines two input bits into one output bit ensuring that multiple conditions are true before allowing an action so if both inputs are true both inputs are one it gives true so one and the output if one of the inputs is zero it gives an output of zero the problem is that all modern computers implement so-called irreversible logic which means once computed you cannot run it simply backwards you cannot reproduce the inputs looking back at the gate if I know that I have zero at the output and I try to run the operation in the reverse I can't reproduce it even when I know that the output is zero I'm still not clear what the input was there are many possible combinations of inputs which gives me zero at the output this kind of logic is irreversible so once it's computed it's gone and you can't get it back and the most important the energy is wasted but what if we can build a computer where we don't waste any information can we then recover this energy Landauer himself proved it was possible he found that the loss of energy was due to the destruction of information not from the execution of operation itself so if we were able to build a reversible computer theoretically we could compute using no energy at all it turns out that when you use reversible сomputing you can really kind of decouple the generation of heat from the processing of information and it's difficult to translate that into kind of terms where we could construct logic but when you figure that out it then allows you to in principle significantly reduce the energy dissipation there by the energy consumption of computing at first it's really hard to comprehend just because most of the things we experience in our lives are irreversible take aging for example or when you pour hot coffee in the mug it spreads heat to the mug right and into the air and the coffee is cooling down and it's irreversible without external intervention and we know that from the Second Law of Thermodynamics which says that the total entropy of the system either stays the same or increases never decreases what is even more interesting that the most fundamental laws of physics are reversible means if you know the state of a closed system at some point you can always run it in reverse and determine its state at any previous moment if we take a game of billiards for example if you film it it would look normal played forward or backward because the physics of collisions is the same so you could easily predict their past or their future positions and the same reversibility holds for Quantum Computing take D-Wave Quantum Computer for example Quantum gates are always reversible couldn't we just apply the same principles to Conventional processors 60 years ago Landauer considered this idea impractical back then if we wanted to make computation reversible we would need to store every input and every intermediate result and doing so would quickly fill up the memory but later on his colleague Charles Bennett found another way he thought what if instead of just storing intermediate results in memory once the result is no longer needed one could just reverse the computation this has a huge advantage because in this case we only need to store the result of the operation and one of the inputs and then we can reverse it back let's get back to our example if we take the end gate and add an extra output copying one of the original inputs then knowing this two values we can reverse the operation if I know that the output is zero and one of the inputs was one I'm immediately certain that the second input was zero with that I can reverse the operation or decompute it getting back the energy that was used in computation of course at the computer level it's way more complicated but in principle we understand now that it is possible to recover the operation to reverse the operation but how does it help us to get back the energy let's have a look Vaire is a startup that is building the first commercial reversible computer this computer chip implements reversible logic which we've just discussed so they took the standard logic gates and adjusted them to work in reversible fashion and typically it's quite a hustle but in this case it's good that this is compatible with our traditional manufacturing process process and of course these gates are still take energy to switch from zero to one but the trick is as long as this energy is not dissipated as heat but stored in the transistor they able to recover it at the decomputation step and they're getting close to it the first prototype is coming out already this year it's kind of interesting so normally reversible computing requires both a forward and a backward step this isn't quite necessarily true so it kind of depends on the paradigm of reversible computing you're considering it is the easiest approach to do reversible computing where you compute one step forward and then one step backwards and this is because you can embed any of traditional algorithm that we've already developed over the last 100 or so years of computing you can embed that in a reversible computer by kind of computing it forwards saving the output and then decomputing kind of all of the intermediate and temporary data that you generated and so kind of this allows us to just automatically reversal anything that we already have and save almost all of the energy now let me guess what you thinking how does computing twice help us to recover the energy and it's a good one here they're actually connecting the logic gates to a resonator through the power rails and this is a very interesting trick yeah we do that by essentially embedding the reversible logic circuits in a resonant oscillator currently it's a lc oscillator consisting of an inductor and a capacitor which is a capacitance of the logic as the energy slashes back and forth so to speak between the logic and the resonator or between different parts of the logic through the inductor and the resonator you know only a small fraction of it is dissipated and so you're essentially recovering and recycling most of the charging energy rather than dissipating it as heat so that's kind of the basic principle well a resonator is a kind of a pendulum that swings it oscillates naturally back and forth and it's just like a swinging pendulum that stars energy as it moves in case of this computer chip instead of a swinging pendulum we have a resonator where at the resonance frequency energy is bouncing back and forth between an inductor and a capacitor and if there were no friction it could bounce there forever on the up swing the computation step is performed and the energy goes to the logic charging the capacitor and on the down swing decomputation step is performed and the energy goes back to the resonator it's a very clever way it took me quite some time to understand this but essentially it comes down to basic electronics if your logic gates seem as a capacitor and you connect an inductor to it and you make it to resonate the energy bounce back and force and if you still haven't gotten it just yet yet don't worry just swing through this part once again of course in reality there will be always some energy lost due to the parasitics the resonator quality device characteristics so we will need some energy to make it to keep going what's interesting there is one more trick they're using together with reversibility they're using so-called Adiabatic Technique for bringing the energy to the logic gates for charging those capacitors this means that instead of charging them abruptly they are slowly ramping up the voltage with controlled current and this allows to further reduce the energy consumption it turns out that if you are following strictly the rules required for fully Adiabatic switching it requires the logic to be reversible you can't you can't erase information under that constraint and so when you compute some information and then you want to get rid of it you have to then decompute it by essentially doing the opposite of the transformation that computed that information in the first place and so that's the connection to a reversible logic and with this approach they can theoretically achieve energy efficiency gains of more than three orders of magnitude compared to conventional chips eliminate heat dissipation and potentially break the Landauer Limit and this approach is very promising for applications like machine learning AI inference and in general for building low power systems and the most exciting part that their first prototype is coming out in early 2025 and I'm looking forward to see the measurement results of how much energy they're actually able to recover in practice and then recycle in the system and at the same time they are working on the next chip which is designed to perform multiply-accumulate operations for AI influence applications right now everyone is recycling zero right we just need to prove that it's non-zero right because the moment that we prove that is non-zero and that we don't use more energy which basically simulation suggest that's the case at that point we know we open up the entire tech tree for rest of computing and then becomes like an engineering problem right of course when you're pioneering a new approach to computing there are many challenges to solve the first big trade-off of this approach is area reversible logic takes more space on the silicon and the resonator itself is quite big it's in micrometer range with so much effort going into shrinking transistors why would we be willing to sacrifice area in this case well in fact if you take the most powerful NVIDIA GPU or an AMD GPU what's happening already now that the big chunk of logic cannot be computing all at once due to the power and thermal constraints in conventional chips the performance per area is not actually limited by the area taken by individual logic gates it's limited by power dissipation constraints right so you know at least by a factor of 10 or 100 you know you could compute much much faster right if it wasn't for the power dissipation of the gates and there's a lot of dark silicon on today's chips because of this you can't actually actively run you know you can't tile the chip with gates that are actively switching on every cycle it would just overheat and so because of this you know we've actually got some breathing room to play with you know you could afford to spend a little more area on your gates if you make them more energy efficient and you could actually get greater aggregate throughput per die area and the ongoing trends building larger chips vertical integration stacking chips on top of each other making the situation even worse and apart of this of course many challenges remain the first one has to do with the technology and manufacturing building reversible logic gates and building a high quality resonator and then integrating them all together into a single chip and on top of that they have to build the whole software stack for this new hardware and we know that's a lot of work what's making it even more complicated is that typically reversible computers require reversible algorithms and reversible programming but luckily in this case they manage to hide so to say all the reversibility at the circuits level as we get to more and more advanced reversible computers you can generate more reversible algorithms which kind of do not necessarily require these backward steps but for now we can already get a lot of gains just by adding in this extra backward step without really sacrificing performance as I mentioned in my previous videos I always admire people who are pioneering new approaches to computing because it's a lot of hard work which sometimes take a decade but just think about it if we could build a reversible computer which operates adiabatically we could solve the biggest problem the heat dissipation problem and then we could truly go vertical stacking more chiplets on top of each other building truly 3D chips and this could open the whole new era in computing interestingly there are alternative implementations of reversible computers ongoing and one of the prominent approaches is building reversible computers with light and in general computing with light is a very promising approach and I'm working on one big episode about it so make sure to subscribe not to miss it as the researchers are working on a super conducting implementation of reversible computers and also on the molecular implementation using DNA but don't ask me how DNA Computing works to me the new year is a perfect time to develop new skills and take your career to the next level this video is brought to you by Boot.Dev the go-to platform for learning how to code with Boot.Dev you can master back-end development using Python and go in a self-paced game like manner their platform guides you through writing real code combining essential theory with hands-on projects after all the best way to learn programming is by building real world applications as a computer engineering major I spent years coding in C# and C++ but back then I didn't have an opportunity to learn Python it wasn't until I started working as a chip designer I realized that Python is fundamental essential for scripting and beyond and today knowing Python is a must have skill for anyone in tech so if you're looking to take your career to the next level Boot.Dev is a perfect place to start use my code ANASTASI to get 25% off the entire first year if you choose an annual plan check them out through the link below and thank you Boot.Dev for sponsoring this episode
2025-01-19 16:59
