The Kasli SoC is much stronger than the Kasli in several metrics, as explained in:

Lam et al., Combining processing throughput, low latency and timing accuracy in experiment control (2021)

Practical example:

  • An important metric is how long an empty RPC call takes, see for example the code and the histograms in artiq/issues/2601.
  • In December 2024, we repeated this measurement with a Kasli SoC v1.1.1 and a Kasli v2.0.1. We flashed them separately with the latest firmware on artiq release-8 and measured histograms with 10,000 empty RPC calls.
  • The Kasli SoC's histogram was peaked at 160 microseconds and fully contained between 100 and 250 microseconds per RPC call, whereas
  • the Kasli's histogram was peaked and mostly contained between 1300 and 1400 microseconds per RPC call.
  • Clearly, the Kasli SoC is better by almost a factor 10 and, clearly, one should upgrade their master device from a Kasli to a Kasli SoC.

Questions:

  • Will the above speed-up also translate to master-satellite communication via the SFP ports?
  • If we run all of our kernels on the master, does it still make sense to upgrade the satellite devices from Kasli to Kasli SoC? Why yes? Why no?

Subkernels and DDMA

So if we use neither subkernels nor DDMA, we might as well use the "old" Kaslis as satellites?