In 7.0.3's testsuite, allT in perf/should_run says: test('T3738', [stats_num_field('peak_megabytes_allocated', 1, 1), # expected value: 1 (amd64/Linux) # expected value: 12800 (x86/OS X): if_wordsize(32, stats_num_field('bytes allocated', 12700, 14000)), if_wordsize(64, stats_num_field('bytes allocated', 60000, 80000)), # expected value: 72608 (amd64/Linux) only_ways(['normal']) ], compile_and_run, ['-O']) which looks quite amazing. On my linux box I consequently get bytes allocated 60608 is more than maximum allowed 14000 *** unexpected failure for T3738(normal) which looks quite horrible. But considering that 6.12.3 allocated nearly 500,000 bytes for that, it's pretty fine. In HEAD's testsuite, the figures have been adjusted to if_wordsize(32, stats_num_field('bytes allocated', 40000, 50000)), and HEAD does that, allocating 46008 bytes. A trivial 'main = return ()' produces ./nuff +RTS -s 47,496 bytes allocated in the heap 1,376 bytes copied during GC 34,036 bytes maximum residency (1 sample(s)) 19,212 bytes maximum slop here Apparently the allocation figures drastically vary by arch and OS, it would probably be necessary to test on several such and be more generous with the limits. Cheers, Daniel