By: Contrarian (Whatever.delete@this.hotmail.com), October 4, 2015 10:53 am
Room: Moderated Discussions
Anders (anders.delete@this.no.reply.co) on October 3, 2015 11:04 pm wrote:
> itsmydamnation (no.delete@this.way.com) on October 3, 2015 3:47 pm wrote:
> > Heikki Kultala (hkultala.delete@this.iki.fi) on October 3, 2015 8:01 am wrote:
> >
> > > Though the 128-bit FPU's are one thing which Dresdenboy quessed without
> > > any better knowledge/any source. They may also be 256-bit.
> >
> > i saw somewhere that the patch Dresdenboy got the other specs
> > from it stated that 256bit ops are fast path doubles.
> >
> The patch Dresdenboy referenced in his blog post stated all `znver1_avx256_*` (add, iadd, mov, log, etc.) entries
> are `znver1-double`, instead of `znver1-direct` is attributed on all the `znver1_sseavx_*` entries.
>
> Link: https://patchwork.ozlabs.org/patch/524324/
Alright, we know AVX2 is easy to implement, but is AVX2 a waste of die space?
Do games use AVX2 and for those games that do use AVX2 do they get a significant speed up?
Is the gain largely wiped out by the thermal throttling that kicks in when you use AVX2, etc.
Is AMD better off with a 128b FPU on each CPU rather than sharing a AVX2 unit.
With hyper threading sharing a FPU would no longer make sense.
The big thing I see is that AVX2 needs 256b reads and writes, which impacts thermals on the integer side, or not.
AMD has to pick it's battles, is trying to compete with AVX2 on the die a losing position, verses using 128b and getting more cores per die and 'better' thermals. (Intel can afford to spend more optimizing thermals and does, this is more about giving AMD a better chance to compete at all code but AVX2 optimized tasks, which are few.)
> itsmydamnation (no.delete@this.way.com) on October 3, 2015 3:47 pm wrote:
> > Heikki Kultala (hkultala.delete@this.iki.fi) on October 3, 2015 8:01 am wrote:
> >
> > > Though the 128-bit FPU's are one thing which Dresdenboy quessed without
> > > any better knowledge/any source. They may also be 256-bit.
> >
> > i saw somewhere that the patch Dresdenboy got the other specs
> > from it stated that 256bit ops are fast path doubles.
> >
> The patch Dresdenboy referenced in his blog post stated all `znver1_avx256_*` (add, iadd, mov, log, etc.) entries
> are `znver1-double`, instead of `znver1-direct` is attributed on all the `znver1_sseavx_*` entries.
>
> Link: https://patchwork.ozlabs.org/patch/524324/
Alright, we know AVX2 is easy to implement, but is AVX2 a waste of die space?
Do games use AVX2 and for those games that do use AVX2 do they get a significant speed up?
Is the gain largely wiped out by the thermal throttling that kicks in when you use AVX2, etc.
Is AMD better off with a 128b FPU on each CPU rather than sharing a AVX2 unit.
With hyper threading sharing a FPU would no longer make sense.
The big thing I see is that AVX2 needs 256b reads and writes, which impacts thermals on the integer side, or not.
AMD has to pick it's battles, is trying to compete with AVX2 on the die a losing position, verses using 128b and getting more cores per die and 'better' thermals. (Intel can afford to spend more optimizing thermals and does, this is more about giving AMD a better chance to compete at all code but AVX2 optimized tasks, which are few.)