 Hi, I'm Nirdis in the race and today I want to share some of the things my colleagues and I have learned over the last few years of working to optimize the stability and performance of cloud based quantum systems at IBM. In this talk, I will touch upon changes we've adopted in the design and screening process. What we've learned about automated calibration and optimization of gate parameters. How we try to maintain the consistency and the performance of the machines day to day. I'll introduce some modified cross resonance. That we're widely using in these systems. And hopefully by the end of the talk, I'll have given you a good overview of how we realize is the performance improvements in these systems. In different metrics, for example, reduction in two qubit a error. All the IBM quantum systems available today share some of the same basic design choices. We're using fixed frequency transmon qubits and the echoed cross resonance gate as a two qubit entangling gate operation. As a reminder, the standard echoed cross resonance gate includes two pulses on opposite phase labeled CR to the control qubit at the target qubits frequency. Separated by a pie pulse on control combined with single qubit gates we can construct a CNOT out of this. So one of the most obvious changes between earlier 20 qubit devices such as Tokyo and a more recent one is the reduced connectivity. In the latter design, each qubit has at most three nearest neighbors whereas in the earlier Tokyo, a qubit had as many as six. So this change came about to the challenge due to the challenges that we were facing on managing qubit frequency distribution across large devices and the impact that qubit spectators have on gate fidelity. The much simpler layout enabled significantly higher single and two qubit gate fidelities as well as helped us increase the yield of the number of devices that successfully passed our screening process. All right, some of these sort of collisions that we're specifically looking for when we're talking about, you know, frequency distributions and why they're unassailable. Let's just look at two examples. So the first case is perhaps the most obvious one where you have two couple of qubits, and they have nearly the same transition frequency. So, as it sort of approaches the limit we're limited in the speed at which we can drive the gates and ultimately this will limit our gate fidelity. The second transition is when the zero one transition of one qubit is close to the one to transition of a couple qubit. And as we approach this resonance condition, the ZZ static ZZ diverges. And in the worst case we cannot drive good to, or even one qubit gates. So we're trying to avoid these situations. And so this plot at the bottom has a representative distribution of the static ZZs across several IBM Q back ends. Just to demonstrate that with this device design change and screening we're able to down select devices with lower ZZs predominantly less than 100 kilomers. So to calibrate a device from the ground up, and to bring it up to a stable state. And working on a suite of automated calibrations. In addition to the obvious convenience of having it automated, this ensures that every device is treated with the same systematic optimization. And that's it also allows us to roll out new features or updates as we continue to improve our calibration and optimization routine. Here I want to talk specifically about the cross resonance gate length calibration. Given the distribution of qubit frequencies that I just discussed with you, each coupled qubit pair has a different optimal cross resonance date length for maximizing date fidelity. This is influenced by many factors, including coherence time, J and ZZ coupling, as well as qubit spectators. Therefore, for every two cubic gate we have, we have swept the cross resonance gate length tracking error per gate and leakage. And so the length that produced the best result is then automatically updated into our latest gate calibration set. On the right, I've just included two examples of such gate length sweeps where you can see for these two pairs. The optimal date fidelity which is the error per gate which is tracked in the dash dotted line and leakage which is represented in this sort of shaded green occurs at different gate lengths. I've included as the solid green line just as a guide to I sort of estimate coherence limit. As we go into the gate length, we also test driving the gate in both directions. As in some circumstances, driving the gate in the reverse direction where the target qubit is higher in frequency than the control is actually preferable. So this is an example of three qubits from one of our actual hardware back ends, where you can see here that the spectator plays an important role in determining the gate direction of the other pair. If you're trying to tune up the one to, or for example, an entangling gate between qubit to two and three, we find that it is what it is much better to drive the gate with qubit one as a control rather than qubit to because of that, because qubit one and qubit three are so close together in frequency, so that spectator plays a very important role in determining which direction gives is better to drive in. So once we have completed that thorough two qubit gate calibration, we've tuned up single qubit gates, read out, reset. The next task is to maintain the quality of the calibrations through the lifetime of the deployed devices. So for our devices and our electronics, we found that once daily find calibration of control pulse parameters is sufficient. It is much more daily to monitor the device to performance, but also to try to correlate and learn from trends and gate fidelity coherence and other metrics. In some circumstances, this might tell us we need to recalibrate our cross residence gate lengths, or to swap the direction of the gate to make it more robust. Of course, if we only test once per day, we can't guarantee that something catastrophic doesn't happen between those checkups can also be constantly calibrating so we found a more agreeable middle ground where very frequently about every half an hour or so we're testing the stability of our calibrations. In our systems we found that amplitude drift is probably the type of error we're most sensitive to. And so we use a simple area amplification sequence to quickly measure the amplitude stability or at the validity of our calibrations. And we use this as a sort of early warning sign of large drifts for large drifts and calibration and also monitor the health of our electronics. And when these sort of hit surpass the threshold that we've set, they automatically pause the queue and alert us that something has gone wrong. And so this sort of automated constant monitoring which sort of interleaves with all other usage of the devices was really allowed us to increase the number of back ends available as each device is essentially responsible for its own health monitoring condition. But also, it really helps us learn what things we need to improve on in order to make these devices more stable going forward to make them more resilient to these types of drifting calibrations. As I mentioned that we're using a modified cross resonance. There have been many variations over the years to the standard CR sequence designed to address various issues such as classical cross stock for example. In scaling to larger and larger systems, we have begun to introduce a target drive in phase with the control in phase with and simultaneous with the control pulse. To address some of the coherent error that's present. And so this is an additional amplitude and phase calibration, however, it does not increase the gate time and in practice has been very easy to systematically implement. This additional drive has really contributed to improve to cuba gate fidelities on top of all the other changes that I mentioned thus far. And we are currently working on a manuscript to summarize this and hope to be on the archive soon. So, in addition to just to cuba gate fidelities we've been tracking various metrics for the currently deployed back ends and we've been seeing a steady improvement in all those over the last few years. Notably, all the improvements I've highlighted in this talk have significantly contributed to the yearly doubling in quantum volume, including the recent 32 measurement on the 28 cuba rally device. So my colleague here measure measured the heavy output probability up to quantum volume 32, and this play in this plot. We're highlighting that in addition to all those changes, the additional target drive really boosted the heavy up probability for cuba 32. So, in summary and this virtual APS talk, I've highlighted some of the significant changes that have led to overall improved device performance. One of the key design decisions to reduce connectivity to improve gate fidelity and yield. And even after this trade off connectivity and gate fidelity we're seeing that net improvement. So much that in the next generation of devices, including the Falcons and the recent 53 cuba hummingbird, we move to an even further reduction connectivity using the heavy hexagon layout. We've developed and we continue to develop a suite of automated device calibrations and optimizations, which has been crucial for scaling the size of the device as well as the number that we can support. And also it means that every time we meet an edge case that doesn't calibrate well we use this opportunity to learn why it's an edge case, manually figure out how to optimize that and then generalize it to make more robust and better optimization procedures. So throughout the deployed life of advice we're constantly learning how to better maintain, operate and optimize these performance and we continue to do so with every new backend we deploy. I'd like to thank all my colleagues at IBM who've contributed to the results I presented today and to all of you for checking out my virtual talk. Thank you.