measuring QPU access time
I want to understand the different timing components during an anneal by looking at the sampler's timing info, see below,
1. What are the units in which those times are quoted? Is it in micro-seconds? I am not able to find the answer from that page.
2. If I just want to benchmark the time it takes the QPU to solve a particular problem as compared to a CPU, disregarding all overhead related to connecting to the device and such, what would be the best time to look at?
Thanks.
Comments
Hello,
1. The units are microseconds. You can see that here:
https://docs.dwavesys.com/docs/latest/c_qpu_timing.html
2. The page you mentioned is probably the best source of timing information.
The "Breakdown of QPU Access Time" shows the time used on the QPU based on how many samples are taken.
All of the fields in the breakdown should be present in the readout you provided above.
It is a difficult comparison between QPU and CPU, as they do not function the same way, so benchmarking isn't reallly possible.
I hope this helps clarify some things.
Just as a follow-up:
A CPU solves a problem by executing a series of machine instructions that take about a nanosecond each -- the program says what instructions to execute.
The QPU solves the same problem by executing a single machine instruction that takes a several microseconds, plus milliseconds for the programming/readout (I/O) parts.
As we can see, it doesn't make sense to compare instruction to instruction: we have to compare the time for the single QPU instruction to the time needed to run the program.
However there are lots of ways to write programs, and lots of inputs to test on, and it's hard to say anything "definitive" about how much time it takes to solve a given problem.
Also, there are many definitions of "solve" in the problem domains we're interested in.
There are lots of research papers comparing QPU to CPU in terms of time to solve the total problem, but none are definitive, so we are likely to see many more papers in the future.
Please sign in to leave a comment.