AMD Vega Discussion Thread

[xkm1948]

Friendly notice, this is not your daily hype train. VEGA WILL NOT run on rainbow power and output 16K@60fps.


Source:
https://www.chiphell.com/forum.php?mod=viewthread&tid=1716524&page=1&authorid=42930

Using google translate, here is what I can get out of it.

1. Consumer version VEGA will have 8GB HBM2. HBM2 will be from Micron. 2048 bit at 700MHz.

2. The Professional level MI25, which is based on VEGA architecture is suppose to have 4096bit. It seems RTG is intentionally hiding their top performer this time. (MI25, or big VEGA maybe 16GB HBM2)

3. VEGA base core clock currently at 1200MHz. The leaker said the boost clock of VEGA should be pretty high. He suggested somewhere around 1500MHz boost for VEGA, at least that is for engineering samples currently available.

4. VEGA will be air cooled, unlike FuryX

5. Possible price is targeted between 1080 and 1080Ti

6. MI25 may be a larger scale VEGA, the big VEGA versus small VEGA we have been seeing in rumors.


Well that's all folks. Remember, this is NOT you hype train.

 

[uuuaaaaaa]

These are some spicy news!

 

[the54thvoid]

choo choo!

 

[xkm1948]

choo choo!

Nope nope nope nope. Derail your train right off the track.

 

[r9]

I want to know how many ncu's it has.

 

[xkm1948]

I want to know how many ncu's it has.

4096 it seems. Looks like a 14nm Fiji core with both frequency and efficiency optimization.

 

[r9]

4096 it seems. Looks like a 14nm Fiji core with both frequency and efficiency optimization.

I saw that number I thought was related to the memory bus.

 

[the54thvoid]

I'm saying 50-75% faster than Fury X.

(Even Ryzen is only 50+% faster than BD in IPC)

 

[crazyeyesreaper]

Meh its Vega its a 64 CU design so unless its RADICALLY different from the past designs its basically 2x 470s

Fury X was 2x Tonga GPUs essentially but single die. Looking at specs of M125 its likely 2x 470s in a single die package. scale it up call it a day.

M125 = 12.5 Tflops
RX 470 = 4.9 Tflops

9.8 Tflops at clocks of 1206. so 10 Tflops

Boost clocks by 20% and you end up around 12.5 Tflops.

1200 / 1500 = 80 so 20% boost would get it to 12 Tflops at least add in higher bandwidth of HBM and you likely end up with a GPU die comprised of just 2x 470s. Make 1 design scale it up and scale it down. Same as with Ryzen.

 

[xkm1948]

I have been seeing this 2x470 BS a lot recently? Who started it? This is stupidity beyond stupid.

Applying this stupid analogy to everything else:

Meh, 1080 is just 2x1060, double the stream processors right? Like who cares
Meh, 6.4L engine is just 2x3.2L engine right? Like who cares
Meh, 1 man is just 2xballs right? Like who cares

Jesus fu*cking Christ this is stupid.

 

[R-T-B]

Kill the hype train.

 

[Vya Domus]

Didn't AMD said Vega is not entirely Polaris based but rather a different architecture ? I think they even released some detailed info on that. So it being like 2x470 is kinda dumb to say.

 

[crazyeyesreaper]

I have been seeing this 2x470 BS a lot recently? Who started it? This is stupidity beyond stupid.

Applying this stupid analogy to everything else:

Meh, 1080 is just 2x1060, double the stream processors right? Like who cares
Meh, 6.4L engine is just 2x3.2L engine right? Like who cares
Meh, 1 man is just 2xballs right? Like who cares

Jesus fu*cking Christ this is stupid.

Really its stupid look at the specs Fury X vs Tonga based 380x

Antigua rebranded Tonga
2048 Cores 128 TMUs 32 Rops

FURY X
4096 Cores 256 TMUs 64 Rops

In a game where Xfire worked the 380x Crossfire performed the same as the Fury X because.

AMD makes 1 fucking design then scales it up and scales it down.

RX 470 is 32 CUs now considering GCN hasnt changed all that much (better tessellation unit few odd changes but nothing mind blowing or performance inhancing.

M125 is 64 CUs 4096 shaders cut that in half you have a 2x 470s

AMDs current claim to fame with Vega is HBC or High Bandwidth Cache, Now if a game taxes advantage of that it likely will perform extremely well. However like most things it likely needs to be coded for meaning in the majority of todays games looking at AMD's design and how they operate the Vega GPU is likely similar performance wise with 2x 470s with a higher fucking clock.

CU count / TFLOP performance. All point to that. Maybe you cant see the writing on the wall but its pretty damn obvious.

Same for the 7000 series. 7750 was half of a 7850 and 7770 was half a 7870. This is a common design trend for AMD because it keeps R&D low allowing them to make products that hit the majority of the market segments for less outstanding cost.

 

[Vya Domus]

AMD makes 1 fucking design then scales it up and scales it down.

Take a look at these pairs : 2 generations of Fermi , 2 generations of Kepler , Maxwell and Pascal , all of these were the same design scaled up and down at at different clocks and minor differences. How is Nvidia different may I ask ? That's the way this industry works.

 

[eidairaman1]

Lets not turn this into a flamewar/clickbait.

 

[crazyeyesreaper]

Nvidia does something similar yes but its more similar to AMD now than in the past.

Nvidia usually designs a die around 2 products. so for example Titan XP / 1080Ti 1080 / 1070 1050 etc.

In the past Nvidias designs between product stacks were more varied. AMD however has become far more homgenous.

128 bit 256bit 512bit
1024 cores 2048 cores 4096 cores
2gb 4gb 8gb
16 rops 32 rops 64 rops

AMD's performance is easy to predict between generations. Nvidia might use a die for two gens. But fundamentally speaking AMD hasnt really changed much at all from the 7000 series up till now 7000 200 300 400 series 4 generations with the same base arch.

whre as Fermi got a refresh however 480 to 580 Nvidia stripped out compute performance in favor of better gaming performance. This process continued. Then Kepler was newer design rebadged for 700 series with Maxwell being used for upper and some entry level products. Then we get to pascal.

usually AMD will design a mid range GPU say 470 cut it in half to get entry level then double it to get the high end.

Where as AMD. GCN GCN GCN GCN. fundamentaly the same but new features added as they become available. Nothing wrong with that either. It just makes AMD entirely predictable.

Nvidia side 1080 is very similar to 2x 1060s but Nvidia didnt directly double the design. if it was double it would have 96 rops not 64 as the 1060 is 48 rops. it would also have been 12GB GPU with 384 bit bus not a 256.

Nvidia can be harder to pin down. AMD not so much.

 

[Vya Domus]

Nvidia does something similar yes but its more similar to AMD now than in the past.

Nvidia usually designs a die around 2 products. so for example Titan XP / 1080Ti 1080 / 1070 1050 etc.

In the past Nvidias designs between product stacks were more varied. AMD however has become far more homgenous.

128 bit 256bit 512bit
1024 cores 2048 cores 4096 cores
2gb 4gb 8gb
16 rops 32 rops 64 rops

AMD's performance is easy to predict between generations. Nvidia might use a die for two gens. But fundamentally speaking AMD hasnt really changed much at all from the 7000 series up till now 7000 200 300 400 series 4 generations with the same base arch.

whre as Fermi got a refresh however 480 to 580 Nvidia stripped out compute performance in favor of better gaming performance. This process continued. Then Kepler was newer design rebadged for 700 series with Maxwell being used for upper and some entry level products. Then we get to pascal.

Where as AMD. GCN GCN GCN GCN. fundamentaly the same but new features added as they become available. Nothing wrong with that either. It just makes AMD entirely predictable.

And why is this something relevant at all? You do realize the future of silicon chips is multiple identical dies on an interposer , if anything AMD is ahead by getting ready for this when the time arrives.

 

[crazyeyesreaper]

Your looking for a needle in a hay stack never said it was bad design choice. people keep expecting AMD Vega to be awe inspiring god GPU its not its design will be rooted in 2x 470s. Thats what the argument is about. Then again I have all day to keep arguing so... keep it coming gives me something to do while benchmarks run.

That said tho looking at past design Vega is likely 2x 470s with some new features taking up die space. Likely performance will be somewhere around a 1080 reference while using more power and having a larger die thus more expensive to produce.

People can feel free to disagree with me thats life.

Vega is just something to fill out a product stack. The next big thing GPU wise for AMD will likely be Navi and will come about when a newer better smaller node is available.

 

[Vya Domus]

Your looking for a needle in a hay stack never said it was bad design choice. people keep expecting AMD Vega to be awe inspiring god GPU its not its design will be rooted in 2x 470s. Thats what the argument is about. Then again I have all day to keep arguing so... keep it coming gives me something to do while benchmarks run.

Well , you could have said that from the beginning , I don't think any one with enough knowledge expect VEGA to be smash every chart. You are still wrong about it being 2x470 , I have already said it in another comment make a quick search on Google , Vega is not Polaris , there is nothing to argue , you claim something that simply makes no sense in this context.

 

[crazyeyesreaper]

Well , you could have said that from the beginning , I don't think any one with enough knowledge expect VEGA to be smash every chart. You are still wrong about it being 2x470 , I have already said it in another comment make a quick search on Google , Vega is not Polaris , there is nothing to argue , you claim something that simply makes no sense in this context.

Believe what you want when the GPU comes out and its performance is the same as 2x 470s when crossfire scales properly ill just laugh.

 

[eidairaman1]

Vliw was before GCN, so amd has phased out their drivers for vliw and are fully focused on gcn arcs.

To me this is what I speculate the true Fury is supposed to be

 

[r9]

Personally I think 2x470 gives good base line. Hopefully it will give us gtx 1080 for much less I'm pretty sure it not gonna beat 1080ti.

 

[crazyeyesreaper]

Personally I think 2x470 gives good base line. Hopefully it will give us gtx 1080 for much less I'm pretty sure it not gonna beat 1080ti.

AMD could scale up to beat a 1080Ti but die size would be huge and expensive.

Fury X was 596mm2

470 is 232 mm2 double that take into account HBM and possible new feature somewhere around 460-500 mm2 for Vega

so somewhere between 450-500 MM2 for AMD to compete with Nvidia 314mm2 1080. 1080Ti is 471 mm2 So AMD would likely need a nearly 600mm2 die like Fury X to compete and thats just not a good path to go down. as cost per die will be high and yields will likely be pretty bad pushing to something that large.

 

[Frag_Maniac]

Intended as a hype train or not, I can't get into anything AMD regarding GPUs lately, because their drivers seem to be getting worse and worse. At first I thought it was the old (7970) model I'm using. Then I saw on an AMD forum thread, that many have been reporting Crimson scaling not working, Radeon settings not installing, ReLive not working properly, etc, etc, even on much more current models. And the idiot Ray moderating the forum labeled the thread as solved, even though many are still reporting the problems.

Software issues have long been a problem with ATI/AMD, and now it appears to even be affecting their new CPU line. AMD seems to be blaming the W10 Scheduler for some of Ryzen's threads being missed, due to thread and cache identifying errors. Seems to me though, W10 is the OS AMD was testing this chip on for a LONG time before launch, and they're just now finding this out? Something's fishy with that. Another article I read said the W10 Scheduler was found to not be the problem.

No my friends, it's never, save for Bulldozer, the hardware end I worry about with AMD's new products, it's the software.

 

[jboydgolfer]

I'll say one thing for AMD , it takes a hell of a lot of Will and coordination to not only roll out competitive CPUs but also competitive discrete GPUs. Against companies that don't have the same burden on them in respect to the CPU/gpu market.it's a situation where simply staying afloat requires positive movement, never mind getting ahead in the market....you HAVE to tip your proverbial hat to them in respects to that fact.