AMD: Speculation, Rumors, and Discussion (Archive)

No GPU scales perfectly when it comes to FPS/W, as the parts of the Chip that are needed for it to function but add nothing to the performance (display controllers, video converters) do not scale with the chip size and take up larger parts of the die the smaller the chip gets.

 

Of course, but it's likely that it has a Cxxx model as well. C95x looks close enough in price.

 

Actually, until not that long ago the smallest GPUs would consistently get the performance/watt crowns.

 

Only for AMD though. Even back then 460 and 430 were head-to-head in the metric.

 

Nope, both of Fermi's '04 mid-ranges were consistently ahead of the top-end models:

https://www.techpowerup.com/reviews/NVIDIA/GeForce_GTX_560_Ti/28.html

 

But that is most often because those chips are specified to achieve maximum performance with little thought been given on the power consumption.

 

I do not think this can be taken for granted as we have not seen the charts for voltage scaling,performance vs leakage,etc.
It may be that Samsung node is better suited to slightly lower power and closer to nominal voltages, which are being pushed to the limit with the 480.

The problem for AMD may not necessarily be efficiency, but overall performance with this architecture-node that requires it to be pushed further within the envelope, just not enough info to know for sure.
I think Nvidia managed to get a heck of a lot out of the 16nm, but then their architecture also has different limitations, which could be further compounded with GP106 and GP107.

Cheers

 

Indeed. I was simply offering my take on Dave's comments.

I took them to mean that P10 might (none committal of course) always have intended to use those boost voltages. Thought that doesn't mean AMD meant to use them at the final power/current levels. Being over max connector power levels (regardless of the PCI/six pin juggling) indicates to me that they probably didn't...

 

Fermi had a completely different SM for non-high-end parts to improve perf/watt. GP107 has the same SM as GP104.

 

VULKAN Support is now live in DOOM...let the benchmark wars begin...
https://bethesda.net/#en/events/game/doom-vulkan-support-now-live/2016/07/11/156

AMD Blog Post about it:

https://community.amd.com/community...-to-the-next-level-with-vulkan-implementation

 

Does that change the aforementioned fact?

Finally!

 

Well...f@kk OpenGL and good riddance! Running on a Fury X+ i7 6700K + 32GB RAM with everything set to the MAX (besides the Nightmare textures which I can't select) at 2560x1440... during the same fire fight (disclaimer: Firefight which may not be representative of the full game performance obviously)...never dips below 100fps on Vulkan while it often hits 50fps in OGL...BTW Adaptive Vsync can now be enable on AMD GPU when using the Vulkan path as it was Nvidia only in OGL (but I hate that stuff so I usually play with Vsync ON)

 

Well id software did claim that P11 would be able to play the game at 1080p with solid 60FPS.

 

ASYNC Compute currently isn't supported on Nvidia GPUs but id's working with them to add it..If I had to guess it will probably only work/be enabled on Pascal GPUs like the latest Tomb Raider patch...

DOOM VULKAN FAQ:
https://community.bethesda.net/thread/54585

 

I guess it's time to finally pick this up.

 

Translation: We don't want to slow down the game with async on Nvidia GPUs so we're not enabling it until we at least get equal performance.

 

Nah, it will come for Pascal no doubt. Maxwell and earlier simply can't benefit from it.

 

Reddit users are claiming up to 30% better performance with the Vulkan patch on AMD cards.
AMD claims 27% on the RX480.

Given how their OpenGL performance was unexpectedly competitive, this is putting all GCN 1.1+ cards quite comfortably above their Maxwell counterparts.

 

In fact it benefit too GCN1.0 .... But thats not the only thing, some of the latest shaders extensions as intrinsincs etc are used too with Vullkan ( along that surely more than that ). lets not forget that AMD gpu's was not running on the same OpenGL path that nvidia one too, and was surelly more than optimized for the Nvidia OpenGL extensions..

its clear that AMD wil benefit of some special optimization compatiblle with Vulkan who was not there with the OpenGL version used ( not only of async compute ). I have no idea if Nvidia got them too or similar and yet if they are enabled or if they are working on it, or have different one enabled etc.

It could be interessant to get a real complete analysis, specially on AMD side., outside the traditional Vulkan API.

 

And will the async compute be Pascal only like it is with the recent Tomb Raider patch? Meanwhile it appears that all GCN cards are likely to support it. And will there be as impressive a performance boost with Vulkan on Pascal as there is with GCN based cards?

Surprising that a traditional Nvidia developer has been more successful getting everything to work on AMD cards than Nvidia cards.

Apparently there's an option to disable async compute in Doom on the Vulkan driver. It'll be interesting to see just how much impact that one feature has, or if it's mostly just due to Vulkan being much more efficient than OGL for games.

Can't wait to see sites start to benchmark it.

Regards,
SB