Q2B 2022 SV | Qedma Quantum Computing: Characterization and Error Suppression in Multi-Qubit Devices

0:00 foreign 0:01 [Music] 0:04 good afternoon so I’ll be basically 0:07 describing kadma and the solutions that 0:10 we’ve been building for the last two 0:12 years 0:13 so first uh we who we are 0:15 um I think at cadma we’re really blessed 0:17 to have a really world-class team 0:19 I want to also highlight my co-founders 0:23 oops cifsinai the CEO 0:26 a professor doritaronov was one of the 0:28 first to prove the fault tolerant 0:31 theorem for Quantum competing 0:34 um and we have 34 people out of them 15 0:37 phds and six faculty members 0:40 okay so let me describe also the 0:42 position of kadma in the quantum 0:44 Computing stack 0:45 so if you view kind of uh broad stroke 0:48 the quantum Computing stack at the 0:49 bottom level you have the physics the 0:51 processor the pulse level at the top you 0:54 have the design of quantum algorithms 0:56 and applications 0:57 and in the middle uh what we are 1:00 basically developing at kidma is all of 1:01 the layers that bridge the physical 1:04 level with the algorithmic level 1:06 including characterization gate 1:09 optimization on the pulse level 1:11 compilation meant to reduce and suppress 1:15 errors and later on development of a 1:18 Quantum Quantum development environment 1:22 so if we want to get into more details 1:24 of what we’re addressing at ketma so 1:26 let’s ask you know what does it take 1:29 to actually build a useful quantum 1:31 computer so as we all know there’s two 1:33 major axes in this respect one of them 1:36 is increasing the number of qubits and 1:40 you know I think we’ve been seeing kind 1:42 of steady progress and most Hardware 1:44 providers also have a clear roadmap 1:47 exist at the intermediate time scale of 1:49 how to proceed 1:51 the other axis is of course reducing the 1:54 gate errors 1:55 and these are not orthogonal axes in 1:58 fact I think uh if you view the errors 2:02 in the gates for the larger system 2:04 available 2:05 I think you know this we’re seeing uh 2:09 progress which could be significantly 2:12 accelerated and that’s what we’re 2:14 addressing at ketma basically address 2:16 accelerating the progress in reducing 2:19 gate errors 2:20 so if we want to be more specific we’re 2:22 developing several Solutions the first 2:25 solution that we’ve developed 2:26 is the solution for flexible 2:28 characterization of multi-qv devices uh 2:32 I’ll describe the solution with more 2:33 detail but in Broad Strokes it provides 2:36 very detailed description of the errors 2:38 occurring on when you apply logical 2:40 Gates and it does so with Cutting Edge 2:43 speed to accuracy trade-off 2:46 and then when we have the result of this 2:49 characterization we can fit it in to 2:52 basically two modules that are meant to 2:54 reduce the errors one of them is an 2:56 automatic gain optimization module which 2:59 basically suppresses errors at the 3:01 hardware level 3:03 and and the other one is basically you 3:06 know a compilation module that does 3:08 error mitigation and correction and 3:10 reduces errors on the algorithmic level 3:13 so let me now describe our flexible 3:15 characterization solution 3:18 so as I said this is a very detailed 3:21 characterization of multi-cube devices 3:23 it’s very sensitive to many types of 3:26 Errors like coherent errors and 3:28 crosstalk errors that are very hard to 3:30 extract using conventional methods like 3:32 randomized benchmarking for example in 3:35 this plot we could see 3:36 a description of a whole device of the 3:39 level of errors and single Cube Gates 3:40 and two qubit gates we could zoom in in 3:43 a particular gate say the gate between 3:45 these qubits five and six and we could 3:47 view individual errors of this gate 3:49 including its effects on on the 3:52 neighbors of this pair of qubits in 3:54 particular for example we could choose 3:56 to look at the errors occurring with 3:58 Queen two qubits two and six these are 4:00 crosstalk errors or residual 4:01 interactions between the skill bits 4:03 occurring when you actually apply this C 4:06 naught gate and I’ll go into more 4:08 details in a second 4:10 so uh let in order to kind of describe 4:14 more detail what we’re just what we’re 4:16 giving uh let’s actually discuss how 4:18 we’re basically parameterizing those 4:20 errors so we’re describing the gates as 4:24 process Maps uh and we’re parameterizing 4:27 them as an exponential of linbladian 4:30 which is a timely dependent of course 4:31 the physics is not is strongly time 4:33 dependent but this is just a 4:34 parameterization 4:35 that the Virgin has basically three 4:37 parts it has the ideal gate generator h0 4:40 it has the coherent error part which is 4:43 the basically error hamiltonian it also 4:46 has the dissipation part which is 4:47 basically describing the this or 4:49 dissipative or irreversible errors 4:52 and as an example you can see here we 4:55 can have errors on the active qubits 4:58 either single cubic single qubit errors 5:00 or two qubit errors and also cross talks 5:02 which are basically residual 5:04 interactions uh between Cube the active 5:07 cubes and their surroundings and their 5:09 surrounding qubits and I want to 5:11 emphasize that the parameterization 5:13 we’re using is actually very flexible 5:15 it’s it’s user defined so the user can 5:17 actually Define an error model that he 5:20 thinks described properly uh the physics 5:23 of this device and we can help him 5:24 basically understand which parameters 5:25 are important and which parameters are 5:28 not in order to zoom in on the correct 5:30 physical description of the errors in 5:32 the device 5:34 I also want to emphasize several unique 5:36 features of our characterization uh we 5:40 can characterize non-clifford gates for 5:42 example fractional two cubic Gates we 5:45 have extensive set of tools for model 5:46 verification 5:48 we can characterize full computational 5:50 layers and I’ll show you some examples 5:52 and we can also characterize 5:55 non-markoving errors such as leakage and 5:58 history different errors for example 5:59 this is an experiment in which we’ve 6:01 detected a flux tail which carries a 6:06 kind of memory effects in the system and 6:08 our ability to to basically extract it 6:11 was helped us actually explain uh 6:14 predict kind of experimental data for 6:16 other experiments 6:20 um just in terms of time scales how fast 6:23 our Cartesian protocol runs so if you we 6:26 want to for example just characterize 6:28 just a single two Cubit gate including 6:30 the crosstalk with the neighbors we’ll 6:32 do this with around 10 to 5 shots 6:35 for super super connecting device this 6:37 would take uh around one minute and if 6:40 we want to characterize a full layer 6:42 meaning a situation where we apply 6:45 gating simultaneously uh we’ll do this 6:48 also with the same amount of time 6:49 irrespective of the size of the layers 6:52 how many cubics actually participate in 6:54 that layer with these times we’re able 6:57 to basically give you a 90 statistical 7:01 accuracy meaning the next significant 7:03 digit of each error parameter uh and 7:06 importantly are cost processing 7:08 classical post processing that process 7:10 processing time is a real time which 7:13 means that you can feed indirectly the 7:15 results of this characterization to 7:16 other modules such as modules 7:18 responsible for reducing errors 7:21 okay so now I want to basically give you 7:24 a few examples so the first one would be 7:26 an application of our full device a 7:29 module we call full device 7:31 characterization which basically sweeps 7:33 through the whole system characterizes 7:35 all the gates and I’ll also show you the 7:38 predictive power of the parameters we’re 7:41 extracting 7:43 okay so here’s the uh we we ran this 7:46 basically module on a system called IBM 7:48 Guadalupe this system has 16 qubits uh 7:53 We’ve in this kind of sweep we’ve 7:54 characterized 32 Gates it’s around 2 000 7:57 parameters 7:58 we’ve used around a million shots 8:01 and the wall time for this was around 90 8:03 minutes although most of this wartime 8:06 occurs due to basically controller 8:08 delays which could be fixed and we 8:11 believe the oil be fixed in the very 8:13 near future reducing this time to around 8:15 20 minutes 8:17 so this is an example of uh parameters 8:21 uh error parameters for single cubicate 8:23 this is say qubit number 12. it’s a 8:26 significant acting on this we’re Prime 8:29 we’re basically extracting the errors on 8:31 this qubit and its neighbors you can see 8:34 that there’s a large uh error V error on 8:37 the active Cube these are the magnitude 8:39 of the Eric B’s are positive or 8:41 negatives it’s in a coherent error and 8:44 we also have quite significant crosstalk 8:46 parameters for this single qubit gate in 8:50 this system 8:51 another example is this two cubic gate 8:53 so you see this is the two cubic gate 8:55 between qubits five and eight 8:57 um again there are significant uh 9:00 coherent errors on the active qubits uh 9:03 either single qubit errors or two qubit 9:05 errors those are quite significant Trust 9:07 of errors and I want to focus 9:09 specifically on this crosstalk error 9:12 between Q with 8 and 11. this is a 9:14 residual interaction occurring when this 9:17 gate is being applied 9:20 and what I want to show you here is that 9:22 the parameters this cross the parameter 9:24 that we’re extracting actually has 9:25 predictive value so what we’re doing 9:28 here we’re taking the model we extracted 9:30 and we’re basically using it to predict 9:32 uh the results of very simple 9:34 experiments in which we’re basically 9:36 applying the C note gate end times 9:38 so on this uh plus you see this n is the 9:42 number of times we’ve applied the C 9:43 naught and this is an expectation value 9:45 of some observable 9:47 and uh you can see in the blue plot you 9:50 can see the prediction of our model and 9:53 the uh purple plot you see a model which 9:57 we trained without this crosstalk 9:59 parameter so we basically set this 10:01 crossover parameter to Zero by hand and 10:03 you see that if we do this this 10:04 completely suppose the prediction of the 10:06 model meaning that the value we 10:09 extracted with parameter is actually 10:11 meaningful to explain uh to explain 10:15 predictions or to explain actual 10:17 experiments very simple experiments run 10:19 on this device 10:21 okay so I know you know my team 10:23 complains that I’m very keen on on 10:25 looking at plots 10:27 and I guess you’re also tired missing 10:29 all these parts by now so I prepared 10:31 another slide 10:32 and uh 10:34 basically you can see that we can verify 10:36 all of these parameters with 100 we have 10:39 hundreds of these verification plots 10:41 this is how we make sure that what we’re 10:43 doing is actually meaningful 10:45 here I’ve plotted like a histogram of 10:48 all of the coherent errors in this 10:49 device 10:51 um the yellow or the active qubit errors 10:53 you see this device have actually 10:55 significant coherent errors and on the 10:59 blue you see the crosstalk errors and 11:01 you see that this device also you know 11:03 the the tail of this distribution 11:04 actually also has significant crosstalk 11:06 errors 11:09 so so you know you need to take them 11:10 into account if you really want to do 11:13 things correctly 11:14 okay so I want to move on I want to 11:16 describe our layer characterization 11:18 protocol 11:19 and specifically I want to also show you 11:22 our ability to characterize uh 11:24 non-clifford Gates 11:26 so we’re looking now at a layer where 11:28 I’m applying two non-cliffer gates in 11:30 parallel these are fractional RCC gates 11:33 with an angle of pi over 6. and this is 11:36 some particular layer on IBM called Kata 11:38 you can see the the coherent errors of 11:41 these layers the largest coherent errors 11:43 there’s significant active qubit errors 11:46 here but also significant crosstalk 11:47 errors and in fact uh you know we’ve 11:51 we’ve characterized many layers of this 11:55 type and you can see that the crosstalk 11:58 errors in layers is actually much more 12:01 significant uh than the crosstalkers in 12:04 isolated V8 for example you know there 12:06 are many many crosstalk parameters that 12:09 are Beyond one percent for layers or 12:11 this is not the case for isolated Gates 12:15 okay so now I’ve showed you our layer 12:17 characterization I want to move on to 12:19 the automatic gate Optimizer basically 12:22 using the results of our 12:23 characterization to optimize Quantum 12:26 Gates and increase their fidelity 12:29 so let’s return to the layer I just 12:31 described and again the yellow bars here 12:35 are the errors uh of the kind of raw 12:38 gate this is just a re-skilled C note to 12:41 give us the correct angle of pi over six 12:44 um and you can see that using the 12:46 characterization parameters 12:48 and by adding correction pulses single 12:51 two with correction passes we’ve been 12:53 able to basically 12:54 almost completely eliminate all of the 12:57 errors on the active qubits increasing 13:00 the significantly increasing the 13:01 Fidelity basically almost eliminating 13:03 the coherent contribution to the 13:05 infidelity 13:06 you can see also with that we were not 13:08 able to to correct the crosstalk errors 13:10 with single qubit uh 13:12 um single key with correction pulses on 13:15 the gate level we have other methods 13:17 which I’ll not describe here to also 13:20 take care of these crosstalk errors 13:23 now what I want to emphasize that if you 13:25 have crosstalks in your device and you 13:27 want to do this type of air suppression 13:28 it’s actually important to correctly 13:31 characterize the crosstalk because if 13:32 you don’t do that you’re going to have 13:34 systematic errors on the rest of the 13:36 parameters you’re extracting and for 13:38 example if we do this set of Correction 13:42 passes and air suppression based on the 13:45 characterization which is not sensitive 13:47 to crosstalk we indeed suffer from these 13:50 systematic errors in the purple bars you 13:52 can see what happens when we try to do 13:54 air suppression without uh in the based 13:58 on the characterization that’s not 13:59 sensitive to crosstalk and you see that 14:00 this performs significantly worse 14:03 you can also look at to see what would 14:06 be the result in a kind of a simple 14:08 experiment when you run a circuit when 14:09 you apply this layer n times and again 14:11 the number of times we apply this uh 14:14 layer and you can see that the yellow is 14:17 the performance this is a deviation from 14:20 the ideal value of some observable you 14:22 can see that the observable deviates 14:24 strongly from its ideal value for the 14:26 bare gate uh and it’s the the 14:29 performance is significantly improved by 14:31 doing our air suppression this is the 14:33 blue plot and if you do this air 14:35 suppression with a crosstalk insensitive 14:37 characterization you get a significantly 14:40 worse results this is the purple plot 14:43 finally I want to also show you that you 14:45 know there are very simple experiments 14:46 where this air suppression is visible uh 14:49 so this is just a simple Ramsay 14:51 experiment in in which you do a very 14:52 simple circuit uh and and you see that 14:55 there if you compare the yellow uh 14:57 circuit uh the yellow uh plot to the 15:00 blue one you see that our air 15:02 suppression significantly improves uh 15:04 the performance of this layer and you 15:06 can also see this in the in the 15:07 Improvement in the Fidelity 15:10 so if anything I want to emphasize I 15:12 I’ve described uh several Solutions I’ve 15:14 described that are ability to fully 15:17 characterize errors in multi-cubed 15:19 systems I’ve described our ability to 15:21 use this characterization to 15:22 automatically suppress coherent errors 15:25 uh with our Cutting Edge runtime I did 15:29 not describe our ability to basically 15:32 recompile circuits and suppress errors 15:34 on the algorithmic level using this 15:36 characterization and I’ll be happy we’ll 15:38 be happy to discuss this offline 15:41 and I want to thank uh uh IBM for 15:45 allowing us to share this data and for 15:47 their support uh if you’re a hardware 15:49 company interested uh in testing our 15:52 Solutions we’ll be happy to discuss and 15:55 thank you for listening 15:57 [Applause] 16:01 thank you netanel we do have four 16:03 minutes left for questions so 16:06 let me do the one up front first and 16:08 then 16:09 to what extent is your technology 16:11 applicable to or generalizable across 16:14 all quantum computers versus uh just 16:16 restricted to IBM I know that’s where 16:18 you’ve tested but how generalizable is 16:19 it no yeah so I should have stressed 16:21 that we’ve tested this technology across 16:24 many types of systems 16:27 um and I don’t I don’t say it’s it’s uh 16:29 you know Hardware agnostic I think 16:30 that’s a 16:31 foul but uh 16:34 it it with very kind of uh 16:38 minor level of adaptations it’s 16:40 transferable to any type of uh final 16:42 Computing platforms 16:45 and we have another question from the 16:46 back here 16:48 I think setup was a very nice talk I 16:50 have actually two questions 16:52 um so first one is what kind of error 16:54 mitigation techniques are you using 16:56 and the second one is have you Benchmark 16:59 it against 17:00 um what IBM is already doing for our 17:02 mitigation so I know that they have a 17:04 module for our mitigation so a view yes 17:07 thank you yeah so I did not get into 17:09 details about our error mitigation 17:11 technique uh I don’t also maybe get into 17:14 the question of what exactly uh type of 17:16 error mitigation we’re using I could 17:18 tell you that for the sizes of circuits 17:21 uh that are you know uh manageable today 17:26 with air mitigation we can give 17:28 basically uh an increase in the overhead 17:31 by several orders of magnitude so if you 17:33 compare run time of our error mitigation 17:37 of an error mitigated circuit to the 17:40 best known error mitigation Pub that’s 17:43 no published error mitigation method out 17:44 there our irrigation method will give 17:47 will give you a runtime which is several 17:49 order of magnitude faster achieving the 17:52 same accuracy foreign