Gian-Carlo Pascutto (gcp.delete@this.sjeng.org) on June 19, 2017 2:28 am wrote:
> Adrian (a.delete@this.acm.org) on June 18, 2017 10:45 am wrote:
> > The results are not surprising, of course Ryzen with 2 memory channels is slower
> > than a Broadwell with 4 memory channels and for a single thread it is slower
> > than a Kaby Lake with higher IPC and with higher clock frequency.
> >
> > Nevertheless, the results are slightly better than I expected, i.e.
> > Broadwell is only around 25% faster for SPECrate (both int & fp).
>
>
> How do the SPECint/SPECfp numbers look single-core, normalized for clockspeed? i.e. what's
> the real IPC difference? (Preferably with MSVC, don't care for Intel C numbers)
>
> Can you estimate the relative HT gains? (Should be SPEC/GHz vs
> SPECrate/GHz). I guess multi core turbo makes this complex.
>
> --
> GCP


With MSVC 2017, single thread:

SPECint
Ryzen: 35.4 / 4 GHz = 8.85
Broadwell: 38.8 / 3.7 GHz = 10.49
Kaby Lake: 47.2 / 4.5 GHz = 10.49


For single-thread SPECfp, Ryzen has a 11% ... 13% advantage, but I have no idea which might be the clock frequencies that were sustained during the benchmark.

You must keep in mind that these results can give an indication about what you can achieve with MSVC on these processors, but you cannot infer much about the performance limits of hand-optimized programs.

Using the Intel compiler with Ryzen, gives you a SPECint of 47.4, which is higher than the 47.2 obtained with the Microsoft compiler on a 4.5 GHz Kaby Lake !!


Since the compiler can influence the results more than the processors, this benchmark can be useful for estimating the performance of legacy programs, but gives little information about what can be achieved with these microarchitectures.