The algorithm is processed in bounded memory, meaning a very large
 number of arrays can be analyzed on also very limited computer systems.
 The only limitation is the amount of memory required to load the final
 chip-effect estimates into memory (as a ExpressionSet).
"justRMA"(csR, flavor=c("oligo", "affyPLM"), ..., verbose=FALSE) "justRMA"(...)AffymetrixCelSet.character string specifying the estimators used in
  the RMA summarization step.doRMA() used internally.Verbose.ExpressionSet.
justRMA
  in affy package quite well.  It does so by still using a
  constant memory profile, i.e. it is possible to use this implementation
  to run RMA on a much large data set than what is possible with
  affy.  At least 20-50 times more samples should be
  doable, if not more.doRMA().