In a Piledriver vs haswell comparison, the numbers work out about as follows from what I have seen:
- 1 thread per hyper-threaded core vs 1 thread per module, haswell is ~80% faster clock for clock.
- 2 threads per hyper-threaded core vs 2 threads per module, haswell is ~30% faster clock for clock.
The above numbers work pretty good when comparing FX based PD chips to the hyper-threaded i3, i7,and E3 chips from the Haswell Family.
- 1 thread per non-hyper-threaded core vs 1 thread per module, the i5-haswell is ~80% faster clock for clock.
- 2 (or 1) threads per non-hyper-threaded core vs 2 threads per module, the Haswell and Piledriver perform the same clock for clock.
The above numbers work pretty good when comparing FX based PD chips to the i5/Pentium chips from the Haswell Family.
When overclocked vs overclocked, PileDriver has ~10-15% higher clock speeds. Performance tuning the CPU-NB and system memory becomes a wash for performance scaling against performance tuning just the system memory on Haswell, so in most OCed vs OCed scenarios where the system is performance tuned all around, we can just refer back to clock speed differences and apply them to the above "rules of thumb" for comparing and contrasting performance under different conditions.
Overclocked vs non-overclocked (often the case when comparing PD vs Haswell in the budget class under $200 CPUs), PD has ~30-70% higher clocks and AND some small gains from CPU-NB/System-memory performance tuning (a few percent usually) that can't be "washed out" on the non-OCing Intel competition.
Many modern workloads scale perfectly proportionally with parallelism and clock speeds, so you can extrapolate theoretical performance easily by adjusting for more or less modules vs more or less cores and higher or lower speeds etc etc.
However, there are also many workloads that do not scale in direct proportion to changes in parallelism (especially real-time workloads), so factor this in when comparing and contrasting.