Rainier chips are cooling down

Three Rainier 1st silicon chips are on their way to 10mK. On two of them we are doing  device-level testing, the third has a full 8-qubit unit cell with all the programmable control circuitry bells and whistles… here is the wirebonded chip with the whole Rainier unit cell on it.


For all you many worlds QM types: Kind of cool that this chip may be shared by 2^8 other universes all doing the same experiments… I have to admit, opening doors between parallel universes is a pretty fun job.

26 thoughts on “Rainier chips are cooling down

  1. I, for one, have seen this technology before … in my Macbook.

    Keep it up, Geordie. Perhaps some day you might get a glimpse of Steve Jobs’ behind. Ah well, at least you tried.

    Of course, before all of your other loyal readers flame me, you know I am kidding. Looks impressive, man. Keep it up. When is life bringing you to Southern Ontario? We have deep snowdrifts, perfect for failing knees to snowboard on….

  2. Chester:: I think there is some summer cabin Korn Whiskey thing being half planned by nichol… I will stay away from the winter version for now although me and the quadruped do have handicapped snowboarding imminent.

  3. Dave/Geordie: True, but only because this packaging incorporates some “lessons learned” from the Orion one! (And a couple of idiosyncrasies, but I would not get there in public!🙂 ). Actually wirebonding went surprisingly fast, to the point of most of the people not noticing it happen (but then, again, I was not the one doing it!🙂 ).

    Paul B.

    P.S. David! Good to hear from you! We are planning to go to Costa Rica this winter, any cool pointers?🙂 You have been good with providing those before…

  4. You are going to be running some problems through the one that “has a full 8-qubit unit cell with all the programmable control circuitry bells and whistles”. So with actual performance timing then you would be able get some general performance timing out pretty quick right ? The looking at the clock on the wall to see how much faster correct answers were delivered.

    In BC Business interview:

    D-Wave plans to have a 1,000-qubit system operating by the end of 2009.

    “We can fit roughly 2,000 qubits on our current processor, which is about the limit of where we can go with the current design,” admits Rose. “After that’s achieved, we need to have some other method of going to larger numbers. So the next step in the redesign – or the evolution of the technology – is getting to millions of qubits.”

    Corporate America and the IBMs of the world have said, ‘Everybody will be knocking on your door when you get to 500 qubits.’
    Hope you can meet or exceed that schedule. Look forward to hearing any news on moving to the bigger systems that presumably will put away any doubters and provide the performance so that you can sell the services and the chips.

  5. Brian: in principle you are correct about measuring wall clock time. However our simulations of the chip using AQUA@home show predicted median adiabatic runtimes for 8-variable problems to be on the order of 0.1 nanoseconds. This means that you have proof of global optimality if you run the computation for any length of time longer than this. Our electronics are designed to be “slow” and can’t perform the calculation this fast (the limit is about 1 microsecond). So for 8-variable optimization problems we won’t be able to measure the adiabatic timescale–it’s too fast.

    Testing at this level is more about verifying the functionality of the individual components, those components acting together as a system, and verifying fab yield. Bigger chips are literally just tiling the plane with this circuit. All I/O issues in regards to scaling to low thousands of qubits have been resolved. Going from 8- to 128-qubit circuits, given working unit cells, is mostly a function of fab yield.

  6. So based on the AQUA@home simulations and the current design targets, how many variables and qubits and/or qubit quality would be needed to get to sufficiently world beating performance ?

    You are running 10,000 times slower than possible. Are you going to move to 1,000 times slower, 100 times etc…

    Is that needed ?

    So 1000 qubits and how many variables optimized at one time for a competitive performance ? Probably you would have to be at least 10-100+ times better than conventional in order to clearly be worth migrating challenging problems over to you.

    So you could scale to the two thousand qubit level to fill up your current die now if the fab yield was good enough ? No I/O or other scaling issues are blocking you.

    So a wafer with a few hundred chips. If they were filled up with attempts at 2000 qubits would have how many working chips and could you not turn off some sections of the die that were bad. Say only 1000 qubits out of 2000 if the errors happen in a section where you can just cut back the size to not be effected by them ? Would it not be worth trying to go for full wafer of 2000 qubit chips ? Even you only had a handful that were pristine. Or would that for some reason be a waste of time, because you need to go through more faster iterations with smaller chips now.

  7. Brian: For 128 variable problems of the sort we’re currently studying, the median adiabatic runtimes are about 20 nanoseconds, so even for the largest problems of this type that we can load on the chip we can’t run the machine fast enough to make it fail. If the trends we are seeing continue we should be able to generate problems that we can controllably make fail at the 250-500 qubit level.

    Note that even at 128 qubits the median adiabatic timescale is about 6 orders of magnitude lower than the time it takes CPLEX to solve these problems on a dual quad-core Xeon server. We can’t get all of that speed-up as there are other systems timescales that eat into this gain, but it is likely that the 128-qubit chip even with all included overhead will beat CPLEX on a high-end server in runtime.

  8. Also re. your comment about going directly to much larger chips: this is a good point and a viable strategy. We believe however that there is a market for the 128-qubit chips and are currently focused on potential buyers of the technology interested in this level of integration.

  9. It would seem to be worth a shot sometime in 2009 to go for the full wafer with maxed out dies (2000 qubit). Potential for a PR coup. Plus could confirm a higher end of your Aqua performance simulations. Once you have some happy 128 qubit customers it seems they would pay for you to the roll the dice on some potential 500, 1000 and 2000 qubit chips. the lesser version just coming from sub-standard 2000 qubit attempts wired up to avoid defect. Just as AMD was selling 3-core chips that were failed quad-cores

  10. Hi I’m a programmer and Iv been reading allot lately about quantum computers and have been wondering in how many years an actual contender to the classical computer would be possible and what implications this would have to programing languages inn use to day(“example being creation of new languages and so…”)

  11. I believe that quamtum computers and mathemathical processors should not work in the way of a Turing machine , why the results should be always repeatable? , i will try to explain myself , for thematic parks or simulation of universes that resolve differential equations in a way of integration of initial conditions by the method of an step by step x axis values there is no need of have all the results correctly i mean there is posible to create a quantum computer that calculates cycle by cycle a number of steps of an integration function and a control of errors on each variable so if the error is bigger and the step is not very big can be interpolated , (making a linnear aproximation) the values between steps so several Newton,Navier Stokes,Maxwell or other physics ecuations can be aproximatly resolved in an optimal manner that depends on spatiation between x,y axis and probaiblity of error of the results of the numbers returned by numbers obtained by several qubits, for just simulation cinema effects of fluids or virtual reality there is no need of an exact result its like an mp3 you can discard frequencies that are not audible in the same manner you can interpolate each 2 or 3 values of numbers obtained by qubit numbers with errors.then for errors if detected can be recalculated with a less powerful machine and readjusted interpolated values for future initial conditions.
    Your company is very cool and I see so much future in it also in quantum computers.

  12. David: Yes your comments are quite insightful. A lot of work has gone into figuring out how to solve a variety of differential equations using quantum computers. Some of the work is similar to what you’re suggesting.

  13. Riaan: It’s unlikely that quantum computers will introduce any new programming languages. In the short term these machines will generally be accessed by simple function calls as they will have specifically designed uses. For experts probing the machines at a low level the code will be written in existing languages such as python or lisp. Requiring a new programming language puts too much of a barrier to use. In regards to when people will start using these systems outside D-Wave, we plan to be installing at least one system at a customer’s site by the end of the year.

  14. JP: the first thing we are doing is checking the individual devices used to ensure they are all hitting spec. We have what are called “breakout structures” for the devices that allow each device to be tested in isolation from a circuit to ensure everything is working as it should. For example on the Rainier design, the circuits on the chip that program the qubits and couplers are new. The first thing we’re doing is testing this on-chip programmable control circuitry, as having these circuits functional gates the rest of the work.

  15. do these gates also have to be cooled to 0.4 K to be tested? or maybe a little higher like liquid Nitrogen ?

    also with 128 qubits sound like something you might want to automate when you get to 1K

  16. Geordie, oh okay , the reason behind why I ask , if i had to do comparison of how a computer works i would say it’s 2 dimensional and your quantum computer(if its programmable one day like a classical computer) would be 4 dimensional and existing compilers&languages would then be absolute because they utilize way of writing algorithms developed for 2 dimensional world(1or0 classical computer)… And if this is the case then only way to really get your quantum computer to really do what it should, is to develop new programing logic for it …

    oh and keep awesome work , you guys rock!

Leave a Reply

Please log in using one of these methods to post your comment:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s