Codes used within ETMC

• tmLQCD linked with DD\(\alpha\)AMG and QPhiX for gauge generation

• QUDA, DD\(\alpha\)AMG for propagator generation linked to several packages (paramgen, plegma, Nissa, …)

• contraction tools: cvc, plegma and many more

Currently, we have

• GPU based machines (mostly clusters and PizDaint)

• CPU based machines (SuperMUC, JUWELS, …)

in the future

• likely large GPU installation in Jülich

• a new, likely ARM based, European architecture

• complicated vector pipeline and many threads

• peak performance of machines increasing further and further

• however, code efficiency is decreasing if no significant amount of programming and optimisation is invested

• and optmisations need to be done again and again

one would like to have a general enough code to easily adopt to new architectures

If we don't adopt, we risk to loose

• the ability to generate state of the art gauge configs

• our ability for innovations

• our competitiveness

• tmLQCD becomes more and more difficult to adopt to new realities

• it would require a completely new memory management

• and a completely new data layout

• likely not possible with tmLQCD, as it is a C-code

• tmLQCD is the only package supporting twisted 2+1+1 dynamically

• HMC gauge configuration generation with
  • Wilson twisted clover
  • 2+1+1 flavours
  • smearing

• propagator generation

• on the fly contractions

• reasonably optimised kernels for general architectures for all these tasks

• else?

• start a completely new code suite

• hook up with existing project
  • grid
  • Chroma
  • else?

• extend Plegma or tmLQCD

• else?

Any of these requires a dedicated effort and a lot of manpower!

my opinion: need a good compromise between efficiency, flexibility and effort!