IBM s 127 qubit Eagle quantum computer breakdown and reaction to release

Show video

hello viewers so it seems like every few weeks  there's a big quantum announcement it was only   a couple of weeks ago that we saw the chinese USTC  group release two big results relating to quantum   supremacy and so this week what we're seeing is  a new announcement by IBM of a completely new   quantum computing chip called the eagle  it's a 127 qubit chip and this builds upon   the previous kind of chips that they've had this  is about double the number of qubits that they had   before so what we're going to do today is to look  at a little bit of the video that they used to   announce these results and before that i'm going  to walk through some of the basic specifications   of this chip so let's have a look at this first so  basically this is as i said 127 qubit chip it's in   this kind of configuration it's in this so-called  heavy hexagonal lattice configuration seeing here   is the the colors are the T1 times that's the  decoherence time for relaxation from say qubit   level zero to one and then you see the the gray  or the the colored regions these are of course the   CNOT gates where the colors also represent the the  fidelity so we'll we'll dive into the fidelities a   little bit more but so this qubit configuration  is basically what IBM seems to be going towards   if you remember some of the other super conducting  qubits that people are trying they have this sort   of cross configuration this is a little bit  reminiscent of their earlier geometries but   recently they've been going with this relatively  sparse geometry so there's not really many two   qubit connections actually the density of the  the two qubit connections is actually pretty low   but they do this on purpose because basically  there's all kinds of issues with crosstalk and   spectator qubit errors so at least in their  approach this seems to be apparently the   best way forward for them so these are  some of the T1 and T2 times so this is   basically the T1 times broken down per cubit so  you can see T1 times around 70 microseconds and T2   times this is the dephasing times this is around  about 100 microseconds this is pretty similar   to what they had before and so for example if  you look at their previous generation machine   it's basically comparable T2 times a little bit  improved compared to this one but basically it's   about the same values and if you compare  it to what is happening in the rest of the   superconducting qubit world so this is the  evolution of the various T1 and T2 times for   various types of superconducting qubit  technologies and you can see it's   kind of exponentially increasing so people  are saying this is like a moore's law for   superconducting qubit coherence but you know 70  microseconds is more or less kind of in the middle   of the range here as you can see in this figure  so just in case you're interested the the USTC   machine has a coherence time of 30 microseconds so  it's a little bit less and actually sycamore has   even uh lower coherence times this is the  value that i found for a paper it's about 15   or 19 microseconds this is the the readout errors  actually so you know you have a particular state   and then how well you can actually perform the  measurement so average readout error is about uh   0.09 it's about like 10% error in in the readout  so actually surprisingly high in terms of readout   you know compared to gate fidelities and things  like that these two bottom graphs here i think   what this is talking about is like the probability  of measuring zero when you prepare one so if you   prepare one and you get zero obviously that's an  error right so what's the error of that you can   see it's mostly you know in the sort of ten  percent range although there's a few spikes around even like 80 which i assume it  means that you prepare one and most of   the time it gives zero which is kind of  interesting i don't know actually why the   measurement errors are so high so  the same one thing i wanted to ask   why isn't it symmetric based why is there why are  the errors not symmetric based on what you prepare   yeah i guess because so i'm not really  superconducting qubit expert but you know uh there   is an energy asymmetry with these qubits right so  you know one energy level is actually physically   higher than the other energy level so it's not a  really completely symmetric system so if they were   degenerate yes you would expect it to be symmetric  so these are the actual gate fidelities so   I think these fidelities are so even though in  the web page it's quoted as an error it says one   right so i think that fidelities so basically x  error square root x they're basically perfect so   I think people no longer really even talk about  quoting single qubit fidelities anymore because   they're basically perfect right the CNOT error of  course is a little bit more difficult so they have   larger errors and this is basically going to be  the bottleneck in terms of the actual you know   fidelity of the of the gates in this case so the  CNOT gate has basically around about the error is   0.02 so about two percent error you compare that  with basically what's happening in the rest of the   world so you know it's not about two percent  error so this is fidelity so in this in this case   would be 98 so you can see some other works are  achieving better fidelities you know but they're   probably optimized for that might be smaller  systems too so yeah certainly not not bad for you   okay so that's pretty much all the stats  that they actually give there's not actually   much in terms more information so as Scott  Aaronson has said in his blog basically there's   there's nothing he's been asked to quote you  know yeah like uh so you know what what does he   think of the new developments with this IBM and  basically there's no information right so so i   think what i showed you is pretty much all the  hard data they have they had a so-called state   of the union address of the quantum computing  a couple of days ago so maybe we can extract   a little bit more information from that so  let's let's have a look at the video oh that   looks fancy I want to welcome you all to the 2021  IBM quantum summit computer this is Jay Gambetta expert or what is his expertise he's a quantum  expert yeah so he did his phd with Howard Wiseman   who's a well-known Australian quantum physicist  he's some yes so he's been leading this quantum   computing team at IBM for some time now   we need to feed and provide more  energy okay for our growing populations   well at the same time it's a bit debatable  where the quantum computing will help that today as Dario announced just a few  minutes ago we broke the 100 qubit barrier so yeah so it looks like they are basically  trying to get up to a thousand cubits and then   and then a million i guess by 2024 or  five okay wow using over a thousand qubits   and we still plan to do this  by the end of 2023 being and i kind of like the fact that these guys  seem to be actual researchers so yeah they   came across as like super salesman-ey which  stuff i think that would be actually turned off   committing to scaling our quantum systems each  year and then at last year's quantum summit we   released our hummingbird processor at 65  qubits by the way we haven't chosen these   targets and scaling commute number arbitrarily  in fact the roadmap is fully aligned with the   development of critical enabling technologies  along each step is working there oh no next year   yeah yeah you got a double every year now here's  a visual of what our Eagle processor looks like   it's two chips the josephson junction based  super lithium transplants sit on one chip   and is attached to a separate  interposer chip through bump   i can categorically say that this is the most  advanced quantum computing chip ever built   in fact not only has it been  built but our Eagle has landed. yeah well i do feel a little bit like i'm  watching a sales video because i guess we are   wait are we buying it yeah are we invited to this conference sorry for  this presentation well no you you can just sign   up for this anybody can sign up for yeah i didn't  actually end up signing up but yes but anybody can   here's the device map for IBM washington currently  deployed and it is an exploratory system that the   team is currently busy putting through its paces  with calibrations and benchmarks in fact here's   a circuit that Jay actually just ran this morning  showing entanglement across four qubits but it's   only four cubits okay yeah so i mean i you know  this is showing like if you get uh the corners   this is like making it a cat state i think right  um i was expecting this would be like for a lot   more qubits but okay I guess they're just  testing it and our plan is to have eagle   widely available at the end of the year and most  importantly we are on track with our hardware   now we next plan the scale to 433 qubits  with our IBM Osprey processor in 2022.   and for the team is this is a good thing about  the superconducting qubit technology which is that   once you make one unit i mean just more or  less have to print more of them so making   them is not the hard part it's it's controlling  them effectively and you know i mean D-wave   already made you know 2 000 qubit machines it's  not that nobody's ever made a 2000 qubit quantum   computer it's just that you know these ones  are sort of better performing i'd say as i said   quality is a measure of how good our technology  is of implementing quantum we measured this with   a metric we introduced to the world in 2017 called  quantum volume what is quantum volume oh it's like   a combination of how many qubits you have and  also like how deep your quantum circuits are   and since then we set ourselves the goal of  doubling our quantum volume every year so   far we've doubled it every year and we currently  have a quantum volume of 128. we have succeeded in   improving our T1 times dramatically from about  0.1 milliseconds to 0.3 milliseconds we have  

tested several research test devices and we're now  measuring 0.6 milliseconds closing it on reliably   crossing the one millisecond barrier we have also  had a breakthrough this year with improved data   fidelities here you can see these improvements  color coded by device family our falcon r4 devices   generally achieve gate errors near 0.5 times  to the power of minus 3. so this one is about   0.1 but then the one that they're quoting  actually was more like two percent so i'm   not i'm not quite sure whether i guess this falcon  thing is the the previous generation right so now   in addition to the announcement that we've broken  the 100 qubit barrier with eagle we have a second   major enhancement we're proud to say in recent  weeks we've broken the 0.001 error barrier yes so   i think that's mainly the the technical aspect  that they maybe there's a little bit more here so   this would make our third major announcement today  oh we're officially moving from falcon r5 from our   exploratory system to a core system 2016 when  we put the first quantum computer on the cloud   we created a simple programming model where  circuits could be sent directly to the quantum   computer but this architecture only allows the  QPU to work at five to ten percent efficiency   a GPU works with a CPU and runtime software  to fully utilize this power we see our QPUs   working exactly the same way to get the most  out of quantum processors we need to bring some   classical computing close to the QPU in may we  release the qiskit runtime in beta a program like   VQE which used to take our users 45 days to run  can now be done in nine hours this leap forward   and performance combined with our 127 qubit  Eagle processor means that from this point on   no one really needs to use the simulator anymore i  don't get this third breakthrough i mean they seem   to be saying that way that they can integrate the  you know classical optimizations with the quantum   so what they have is the kind of software library  that abstracts those things for you so you just   like define a problem with your circuit and some  parameters and you know you can both simulate it   or deploy to the quantum device and it all now  integrates those classical you know assisted   things as well okay okay hybrid things  okay but yeah i mean the thing with these   you know variational quantum circuits is that  they're not necessarily sort of faster actually   and probably depends on the problem but  they have they have set up that can automate   this yes i don't know you know i mean  i've just seen like a lot of stuff where   like i saw a paper recently where you know they  were trying to you know optimize some kind of   energy production and you know they ran it on  the quantum computer and they said you know it's   so much faster and so forth but i mean you know  comparing that to like a just a you know genuine   kind of optimizer just running on a classical  super computer you know i doubt that those   whether that would actually still be like that you  know beat classical supercomputing yeah methods so   well i don't know yeah i guess it's still but but  you know to be fair i don't think it would even   be given a chance to reach a full potential if  we would just manually do it every time we want   to try it so it's good that someone is trying to  automate this whole thing at least you can try it   and then you know see if it's faster if it's  not then yeah you don't use it yeah and you know   people will have access to it and do research  so maybe they will find a way to make it better   yes yeah yeah i just worry that when they present  it like this it's like everybody that wants to do   any kind of computation they will expect it to be  faster just by writing it on quantum and it's just   most of the case not not true right it's like  it should it's quite specific and i don't think   you just sort of randomly do it and it's like oh  it's faster great okay i actually didn't know that   how do you know when a problem will be faster  to solve well it's very tricky yeah okay let's   think i mean it's basically like saying how do you  find a quantum algorithm that's actually useful   when i haven't i see it so it's generally quite a  tricky thing and you don't find it randomly okay   you have complexity classes that work better okay  we call this circuit name class of techniques   decomposes a large quantum circuit with more  qubits and larger gate depth into multiple smaller   quantum circuits with fewer qubits and smaller  gate depths then it combines the outcomes together   in classical post-processing this allows us to  simulate much larger systems than ever before   this year we demonstrated circuit knitting by  simulating the ground state of a water molecule   using only five qubits with the specific  technique of entanglement forging water molecule   i mean it's that great that's the fact that  they could do that is also i don't have much   to comment on i mean the ground state of the water  molecule is i mean it hasn't it's been known for   years so free wise nothing really  chemistry-wise it's exciting but i mean   okay you can say you've simulated it but you  know have you solved essentially have you have   you solved the problem exactly or is it the mean  field type thing which is much easier well it's   five qubits so it's probably just you know it's  probably the simplest version of solving for the   ground state of a water molecule although i'm just  curious to see the paper to see yeah what they did   but i i know people like Alan Aspuru-Guzik  have been working on you know the application   of quantum computing to doing these sorts of  electronic structure problems and molecules so   it's interesting to see how much better they  do i'm just curious we've heard of google doing something what has this one  actually done the IBM one yes the speaker   actually this one has done something   what has he done actually oh it hasn't done  anything yes you mean like what circuits is it   right yes well we haven't seen anything like that  yet so this is all we've seen with there's this   video uh and then there's the data which i showed  you right at the beginning yeah that's it yeah but   the birthday by the end of the year you could use  it yourselves oh okay so they didn't test it and well apparently they did a four  qubit calculation this morning so   i think they haven't tested it  very much yeah in terms of um   a lot of circuits or anything like that but to  me that makes more sense than quantum supremacy he actually wants to do something oh i see what  you mean yes after 2023 assuming we continue   at the same rate we are in the land of quantum  advantage we say the land because we don't think   quantum advantage will be a fixed point in time  or a specific event we see it more as a continuum   where applications start off abstract and  esoteric and become slowly more useful over   time so you'd think that this is already in the  quantum computational advantage regime right   yeah it should be it should be right that's  what he said though right they were in the   land of this well he's saying it's continuum  when we don't really uh not a predefined   moment or anything like that so you know it's  a little bit of a different perspective to say   that what the Chinese team or even what Google was  saying like that we've reached quantum supremacy   now yeah we've done it you know maybe they will  compete with in a different task using a task   and if they do better in some of them  then great yeah yeah so perhaps they have   a slightly different philosophy there  i have a question does it run on emacs we should if it doesn't it's crime. The D-wave  computer what is he able to do right now?  

Random number generation no no i mean it's  a quantum annealer so it basically it solves   the Ising model which is equivalent  to MAXCUT and stuff like that oh okay   but so we have a review on that  coming out i'm just advertising   that just throwing that out there but anyway we  have a review on Ising machines which should be   coming out in Nature Reviews Physics sometime soon  and this covers all that and basically compares   all the different approaches of solving such  Ising machine ising models called Ising machines sneak peak is that   the quantum ones don't really perform quite  as well as classical quite yet yeah but you know its still early days so  don't write off quantum yet but yeah   okay so what did you guys think i like it you like it you like that man i like  it the aesthetic's really cool it's so futuristic   have you seen Tron yeah  that's sort of sticking out   it looks like tron yeah maybe  one should look like Dune then any other opinions are you blown  away is your mind blown i don't know   not really not really i'm positively surprised well i would say you know what has it done yet i don't know you want  to see it do something yes yes like if you   saw it factor some you know reasonably  large number it does look yes exactly   use this simulated water molecule they didn't see  how they did the simulation that's the thing yeah   i mean i could simulate the ground state of the  water molecule right now it would take two seconds   i don't know well yeah it's  a five qubit calculations   it doesn't matter what you did with the five  qubits you could of course do that on your   computer or right oh yeah but it's the principle  it's the principle right i guess it shows it   shows that they can apply it to something that's  industry wants that's right which is but there is   one thing you could not do if you do it in two  seconds you could not use the high port quantum i thought you were optimistic well I i believe it's it is there really  something you know noble to chase it so it would   be really cool if you could do regular molecular  electronic structure calculations with quantum   computer i just don't know how they do that yet  because i haven't seen what a quantum version   of like the the really sophisticated electronic  structure methods even looks like yet we're still   developing those for the classical case so it  would be very interesting to see how far they get   then we'll leave it there and so thanks for  watching and we will see you the next video bye

2021-11-24

Show video