AMD Vega Discussion Thread

[Divide Overflow]

AMD Radeon RX Vega 10 chips differ - physically and quite significantly

http://www.guru3d.com/news_story/am...iffer_physically_and_quite_significantly.html

Here is the official answer from AMD. - "We have multiple manufacturing partners for Vega, one of them uses an epoxy fill to level the SOC surface, otherwise they’re functionally identical."

Conspiracy derailed.

 

[TheoneandonlyMrK]

Many seem to forget Vegas unique strengths and how little they are used , no mining software Yet uses rapid packed math (2-2.5x) and it does well already, same goes with gaming ,the features Amd are leveraging are more and more likely to be used(due to console crossover code) , negating some of this initial pain as will their driver team , Im sure I won't regret buying one.
Ill waterblock it month end.

 

[Divide Overflow]

Ill waterblock it month end.

Please check back and let us know your results!

 

[Cybrnook2002]

Ill waterblock it month end.

Same

 

[notb]

Many seem to forget Vegas unique strengths and how little they are used , no mining software Yet uses rapid packed math (2-2.5x) and it does well already, same goes with gaming ,the features Amd are leveraging are more and more likely to be used(due to console crossover code) , negating some of this initial pain as will their driver team , Im sure I won't regret buying one.
Ill waterblock it month end.

You've really mentioned mining optimizations first - before gaming. What happened to the world?

I'm sure you will regret it. Just give yourself a few weeks.

 

[the54thvoid]

Many seem to forget Vegas unique strengths and how little they are used , no mining software Yet uses rapid packed math (2-2.5x) and it does well already, same goes with gaming ,the features Amd are leveraging are more and more likely to be used(due to console crossover code) , negating some of this initial pain as will their driver team , Im sure I won't regret buying one.
Ill waterblock it month end.

To be fair we have heard all this before when Fiji came out. Not paying dividends yet. The environment exists but the developers don't like the idea of extra work to make AMD's hardware shine.

 

[londiste]

Many seem to forget Vegas unique strengths and how little they are used , no mining software Yet uses rapid packed math (2-2.5x) and it does well already, same goes with gaming ,the features Amd are leveraging are more and more likely to be used(due to console crossover code) , negating some of this initial pain as will their driver team , Im sure I won't regret buying one.
Ill waterblock it month end.

rapid packed math has only minor effect on gaming, if at all. feedback from real devs was that some shaders where rapid-packed math is used on ps4 pro get roughly 30% speed increase. in the large picture though, it is only usable in a small subset of shaders.

i would be careful with that 2-2.5x claim for mining as well. rpm has limitations that do not exist for normal precision calculations and its theoretical maximum is 2x the performance. realistically, it should be (noticeably) smaller than that.

Here is the official answer from AMD. - "We have multiple manufacturing partners for Vega, one of them uses an epoxy fill to level the SOC surface, otherwise they’re functionally identical."

Conspiracy derailed.

conspiracy? if the height of hbm and gpu chips is indeed different as pictures show (2nd picture in guru3d-s article), this has real implications on cooler mounting. especially with a significant amount of enthusiast vega buyers looking at water cooling, this is important.

 

[RejZoR]

I wonder if RPM (Rapid Packed Math) can on any level be done on driver level so it's not developer dependent by replacing existing shaders with FP16 ones, but AMD doing that if user enables such thing. We've seen shader optimizations in the past, I wonder if this is doable or woth investing into if it is.

 

[londiste]

I wonder if RPM (Rapid Packed Math) can on any level be done on driver level so it's not developer dependent by replacing existing shaders with FP16 ones, but AMD doing that if user enables such thing. We've seen shader optimizations in the past, I wonder if this is doable or woth investing into if it is.

that would be a huge scandal. halving shader calculation precision in driver? these kinds of things have been very very bad in the past.

it has been said by devs (for example dice guys regarding bf1) that fp16 is not suitable for everything. quite specific choices will have to be made to use fp16 for optimizations.

edit:
for a specific example - my guess is that using fp16 for anything color related would be bad.

 

[Vya Domus]

To be fair we have heard all this before when Fiji came out. Not paying dividends yet. The environment exists but the developers don't like the idea of extra work to make AMD's hardware shine.

Except Fiji in reality had nothing special , many people got the wrong idea that Fiji has it's special place somewhere in AMD's drivers or in optimizations done by developers. It does not , every optimization that can be done all the way up to Polaris is applicable more or less on Fiji as well , on Vega though not so much. Drivers were built from the ground up for Vega , does that mean there will be a huge uplift in performance in the future ? Who knows , but it does look promising once you go past the memes and hate.

I wonder if RPM (Rapid Packed Math) can on any level be done on driver level so it's not developer dependent by replacing existing shaders with FP16 ones, but AMD doing that if user enables such thing. We've seen shader optimizations in the past, I wonder if this is doable or woth investing into if it is.

Not without losing image/rendering quality.

 

[RejZoR]

I wonder if RPM (Rapid Packed Math) can on any level be done on driver level so it's not developer dependent by replacing existing shaders with FP16 ones, but AMD doing that if user enables such thing. We've seen shader optimizations in the past, I wonder if this is doable or woth investing into if it is.

Can you people stop sperging out? Like, seriously...

 

[londiste]

Except Fiji in reality had nothing special , many people got the wrong idea that Fiji has it's special place somewhere in AMD's drivers or in optimizations done by developers. It does not , every optimization that can be done all the way up to Polaris is applicable more or less on Fiji as well , on Vega though not so much. Drivers were built from the ground up for Vega , does that mean there will be a huge uplift in performance in the future ?

why wouldn't all these optimizations apply to vega? the shaders and structure are pretty much the same. there is no real need to rebuild drivers from ground up. yeah, tile-based rasterization and possibly primitive shaders would have to be supported in drivers, but this is pretty much it.

tile-based rasterization has been topic ever since architectural details ewere revealed and actually a bit earlier than that. amd has had 8 months in public knowledge of the feature to get this working in the drivers. if they have not managed to get huge improvements out of it, what will change now? granted, they definitely have gotten some improvements from this by now, just less than expected.

primitive shaders have been an interesting topic. initally what amd communicated was that these needed to be used by devs. now they are saying those will be enabled in drivers to some degree in some form. we will see.

 

[Vya Domus]

why wouldn't all these optimizations apply to vega? the shaders and structure are pretty much the same. there is no real need to rebuild drivers from ground up. yeah, tile-based rasterization and possibly primitive shaders would have to be supported in drivers, but this is pretty much it.

It does not , Vega has a different instruction pipe-line with new extensions , you can apply the same low level optimizations as to previous GCN iterations but it will either get you nowhere since you'd be ignoring most of what is supposed to speed things up or it would outright not work. This is exactly what happened , no one knows how much time did AMD work on this , my guess not so much. Also no ones know how much more can be done , even if it can be done , RTG can have other things to worry about.

 

[RejZoR]

AMD will seriously have to split their GPU's into two types, gaming and compute. GCN is excellent at compute, but doesn't work too well with games. NVIDIA learned that the hard way with Fermi and even Kepler had some problems. Maxwell is almost entirely focused on gaming and so is Pascal. I'm talking consumer branch. For Navi they'll have to focus on two types, hopefully their modularity will give them this freedom. Otherwise they'll have hard time competing in gaming segment again...

 

[TheoneandonlyMrK]

rapid packed math has only minor effect on gaming, if at all. feedback from real devs was that some shaders where rapid-packed math is used on ps4 pro get roughly 30% speed increase. in the large picture though, it is only usable in a small subset of shaders.

i would be careful with that 2-2.5x claim for mining as well. rpm has limitations that do not exist for normal precision calculations and its theoretical maximum is 2x the performance. realistically, it should be (noticeably) smaller than that.

conspiracy? if the height of hbm and gpu chips is indeed different as pictures show (2nd picture in guru3d-s article), this has real implications on cooler mounting. especially with a significant amount of enthusiast vega buyers looking at water cooling, this is important.

Rpm has No effect on gaming because its not used and until it is your speculating as am i so please sound a bit more like that as you Know what we all Know , nothing yet.
I can play most games at 4k ultra textures etc etc and decent fps with a 580 and its mostly shadows , effects ,AA and post process that slows frame rate production.
Its these and not everything that could be targeted for rpm use Without sacrificing quality.
But im certainly sharing an opinion not facts ,their on the way.

 

[londiste]

fp16 is used in some cases. bf1 (on ps4 pro) is an example. check from slide 82 in the presentation:
https://www.ea.com/frostbite/news/4k-checkerboard-in-battlefield-1-and-mass-effect-andromeda
whether that is already used in pc games running frostbite , i am not sure.
there have been couple other comments about using fp16 but i can't find links right now.

rpm invariably sacrifices quality. with good dev and very selective use, not noticeably.

 

[TheoneandonlyMrK]

fp16 is used in some cases. bf1 (on ps4 pro) is an example. check from slide 82 in the presentation:
https://www.ea.com/frostbite/news/4k-checkerboard-in-battlefield-1-and-mass-effect-andromeda
whether that is already used in pc games running frostbite , i am not sure.
there have been couple other comments about using fp16 but i can't find links right now.

rpm invariably sacrifices quality. with good dev and very selective use, not noticeably.

Well thats a broad selection your getting your facts from then eh , i can see how your an expert now sorry ,your right does anyone want to buy a vega ,this guys opinion appears to be rock solid science style fact. Sigma 6 to the max.

 

[HD64G]

Some great results from Vega56 oc from Gamers Nexus in the video below.

At about the 6 min mark in the video there is a table of results with different clocks and settings

Some interestimg points now:

1) Just by oc the HBM2 up to 950MHz gives 14,7% more!
2) Best stable result is when core is clocked at 1692MHz with memory at 980MHz and it is 35% higher than the stock one (25627 oced vs 18892 stock).

Since drivers and oc programs aren't optimised for Vega yet I can say that this GPU has great potential. If I buy it I will just undervoltage where it is stable and then oc the Vram and gain close to 20% easily with not bad power consuption at all. For the enthusiasts though who don't care about wattage, there is a new game to play with.

 

[Th3pwn3r]

Some great results from Vega56 oc from Gamers Nexus in the video below.

At about the 6 min mark in the video there is a table of results with different clocks and settings



Some interestimg points now:

1) Just by oc the HBM2 up to 950MHz gives 14,7% more!
2) Best stable result is when core is clocked at 1692MHz with memory at 980MHz and it is 35% higher than the stock one (25627 oced vs 18892 stock).

Since drivers and oc programs aren't optimised for Vega yet I can say that this GPU has great potential. If I buy it I will just undervoltage where it is stable and then oc the Vram and gain close to 20% easily with not bad power consuption at all. For the enthusiasts though who don't care about wattage, there is a new game to play with.

That was actually a pretty good video. Definitely makes VEGA FIFTY SIX look good.

 

[TheoneandonlyMrK]

From my own dabbling vega needs a driver update really badly I've tried many times to oc and i end up rebooting due to a weird glitch where it continuously refocuses the desktop like im constantly changing the res between two similar res's.
Only a reboot regains function and overclocking is not working in any useful way ,im eagerly awaiting Amds next driver anyway.

 

[P4-630]

From my own dabbling vega needs a driver update really badly I've tried many times to oc and i end up rebooting due to a weird glitch where it continuously refocuses the desktop like im constantly changing the res between two similar res's.
Only a reboot regains function and overclocking is not working in any useful way ,im eagerly awaiting Amds next driver anyway.

So you were able to buy one of these Vega's?
Which one do you have?

 

[TheoneandonlyMrK]

So you were able to buy one of these Vega's?
Which one do you have?

Reference its Always the one to waterblock , no waterblock yet though.

 

[P4-630]

Reference its Always the one to waterblock , no waterblock yet though.

56?

 

[HD64G]

56?

RX56 isn't for sale till 28th of August me thinks mate.

 

[P4-630]

RX56 isn't for sale till 28th of August me thinks mate.

Ok... Didn't know that.