AMD Radeon RX VEGA to Launch at SIGGRAPH 2017

[LAwLz]

20 minutes ago, tom_w141 said:

ITS NOT BEING RELEASED ANY TIME SOON. Sorry you keep saying "volta tomorrow" and "volta next month", I thought raising my voice might get past the ear wax  

 

IF it was next month then yes you'd be fair to say so. But if you don't compete with something released 8-9 months after then that's hardly an issue and is clearly what your next product competes against. How many different people have to say this before you get it? You are semi right, in the sense that if volta followed near immediately but it won't. Or it won't unless NVidia have to launch it early as damage control  because vega clearly beats what they have (which it seems like it won't).

 

tldr: on sale at the same time is not equal to competing against. RX 480 and GTX 1080 were on sale at the same time. They did not compete.

Sorry. I remember reading that Volta would be released in Q3 2017 so that's what I based my original comment on. After reading replies and googling a bit it seems like it will be out Q1 2018. I don't think that invalidates my point though.

Navi won't be out until 2019 (according to rumors), so that means that if AMD gets Vega into stores next month then we will have august to Q1 2017 with competition between AMD and Nvidia (so something like 6-7 months), and then it will be Vega vs Volta for the next ~12+ months.

 

AMD will not be in a good spot if Volta is better than Vega, because those will be the cards people will decide between for the majority of time Vega will be on sale.

Vega needs to be a compelling alternative compared to Volta.

 

 

24 minutes ago, tom_w141 said:

tldr: on sale at the same time is not equal to competing against. RX 480 and GTX 1080 were on sale at the same time. They did not compete.

Because AMD completely pulled out of the high end category, and Vega is suppose to be them re-entering it again.

People don't want another mid-range AMD card. They want something that can compete with Nvidia. If AMD can only compete with Nvidia on the high end for 6 months and then get crushed for the next 12 months again then I don't think they succeeded with their goal.

[Jito463]

8 minutes ago, LAwLz said:

and then it will be Vega vs Volta for the next ~12+ months.

That assumes AMD can't eke more performance out of Vega in the meanwhile, like they did with Polaris (400 to 500).

[mr moose]

2 hours ago, tom_w141 said:

If Nvidia bring volta forward to Q3 2017 then I absolutely agree, but if we see Vega at the end of July with supply in august and no Volta till 2018 as planned then they definitely don't compete.

 

 

I don't think release time has much to do with whether two products are competing or not. Otherwise we could be arguing that the 1060 (almost a year out now) is not competing against the RX580.   Even though they are similar in price and performance.  Which means the only logical conclusion we can draw is that unless AMD come out with something bigger and better than Volta when it drops (regardless of when it drops) then you are right, becasue Nvidia being a year in front again will make it hard for anything to compete. .

[Shreyas1]

Now that all this negative press starts, I bet AMD will rush out their cards, making things worse...

[Valentyn]

2 hours ago, leadeater said:

It's actually quite funny how well Fury X/Fiji does in certain workloads yet translates so poorly to gaming performance. It also doesn't take much critical thinking to realize that since Vega is still based on GCN architecture family that a 50%-60% increase in clock speed with no other optimizations would make a Fury X/Fiji card slot right in between the 1080 and 1080Ti.

 

I mean that alone should be enough to know Vega FE is not actually representing the proper gaming performance it should have, not even including the architecture improvements unless people are getting bold enough to say AMD have actually produced an architecture with lower performance than the last. If that was actually the case you would just cut your loses and node shrink Fiji and increase the clock speed.

Exactly, scaling Fiji to 1600Mhz makes it a very strong GPU; and it wouldn't have taken them over 2 years to simply manage a Die Shrink of Fiji.

The performance for specs make no sense, even given Fiji's horribly bottlenecked front end.

 

I get the doom and gloom from many, the first viewing of Retail Vega is not good in the slightest. In fact a node shrank and doubled Hawaii core would be stronger; and hell have proper 1/2 FP64 compute as well.

 

The wait to Siggraph feels worse than ever now, and AMD is either going to pull out a golden rabbit with Radeon RX Vega; or......

[rasmuskrj]

Saying that increasing the clock on Fiji to 1600Mhz would make it a competitive card doesn't really make much sense to me. I'm sure that if they have had been able to do that, they would have done so.

It's like saying that if they just increased the clock speed on Ryzen then it would be a much better processor. It's not wrong, but it's not really that simple.

 

Architectures are limited by how their pipelines work when finding out how fast a clock they can run at. If you have a long pipeline with small, fast to compute stages then you can run at a higher clock than if you have a short pipeline with more complex stages.

 

My point is that you cannot assume that it would have been faster, or even possible, to optimize the Fiji architecture to run at higher clocks than it was for them to make a new architecture that is capable of higher clocks.

[tiggar47]

1 hour ago, L.Lawliet said:

u are so funny man..

I know GPU its the chip itself but people call graphics card a GPU nowadays just dont get confused with it.

Yes The PCB is also important which consist of Vrms,Capacitors,Etc and the driver is also important.

 

I'm not confused, I know what they refere to but calling Graphics Card a GPU is common but it's wrong

[tiggar47]

Just now, L.Lawliet said:

Its because GPU is the most important parts from a graphics card.

Yes, but it's still wrong

[Prysin]

5 hours ago, LAwLz said:

But that is the case...

When I go out shopping I don't look at the release dates of products to determine which one was the best at the time of release. What I will do is look at what's in store the day I am out shopping and then decide which products to get out of the ones available.

 

If Vega is released today, and Volta released next month, and then we don't get a new graphics card for 24 months then I sure as hell wouldn't go "well AMD is still competitive because at the time of release, Nvidia did not have Volta!". I would say Vega is competing against Volta because at the time of shopping those are my two options.

 

What I am worried about is that AMD will release a card that's only as good as what Nvidia have had on the market for over a year now, and then Nvidia will be able to crush AMD shortly after with Volta.

It just means AMD is one step behind Nvidia. If you're only competitive 6 months out of 12 then you're in a really bad spot.

Of course AMD could drop prices to stay competitive in price:performance, but then we are yet again in this situation where AMD's margins are low and they are seen as a mid-range offering.

 

 

Anyway, leadeater put it way better than me when he said:

That's what my original point was. Vega and Volta will be on store shelves at the same time, and if Vega is just barely matching Pascal then AMD will have a bad time.

 

 

 

Love you too. <3

i never mentioned your name or hinted at it being you, but ok. Still not going out with you, i don't date wimminz iv'e never seen the real face of.

[Prysin]

1 hour ago, L.Lawliet said:

Its because GPU is the most important parts from a graphics card.

no, that would be the PCB... without the PCB, that GPU wouldnt connect to anything  

[Prysin]

3 hours ago, Valentyn said:

Exactly, scaling Fiji to 1600Mhz makes it a very strong GPU; and it wouldn't have taken them over 2 years to simply manage a Die Shrink of Fiji.

The performance for specs make no sense, even given Fiji's horribly bottlenecked front end.

 

I get the doom and gloom from many, the first viewing of Retail Vega is not good in the slightest. In fact a node shrank and doubled Hawaii core would be stronger; and hell have proper 1/2 FP64 compute as well.

 

The wait to Siggraph feels worse than ever now, and AMD is either going to pull out a golden rabbit with Radeon RX Vega; or......

Fiji has too much leakage to shrink. David Kanter and The Tech Report had a good discussion on the flaws of the Fiji Architecture, along with some very informative tweets by Kanter and Scott Wasson.

[leadeater]

4 hours ago, rasmuskrj said:

Saying that increasing the clock on Fiji to 1600Mhz would make it a competitive card doesn't really make much sense to me. I'm sure that if they have had been able to do that, they would have done so.

It's like saying that if they just increased the clock speed on Ryzen then it would be a much better processor. It's not wrong, but it's not really that simple.

 

Architectures are limited by how their pipelines work when finding out how fast a clock they can run at. If you have a long pipeline with small, fast to compute stages then you can run at a higher clock than if you have a short pipeline with more complex stages.

 

My point is that you cannot assume that it would have been faster, or even possible, to optimize the Fiji architecture to run at higher clocks than it was for them to make a new architecture that is capable of higher clocks.

Fiji was mainly limited in clock speed by the node size and power that used. Going from 28nm to 14nm FinFET would have allowed them to increase the clock speed quite a bit and even though Fiji didn't OC well it responded rather well to any increases you could give it. Sure you can't just simply increase clock speeds and not do anything else but don't forget that is basically what GPU development was 2004-2010, node shrinking and minor optimizations.

[rasmuskrj]

1 hour ago, leadeater said:

Fiji was mainly limited in clock speed by the node size and power that used. Going from 28nm to 14nm FinFET would have allowed them to increase the clock speed quite a bit and even though Fiji didn't OC well it responded rather well to any increases you could give it. Sure you can't just simply increase clock speeds and not do anything else but don't forget that is basically what GPU development was 2004-2010, node shrinking and minor optimizations.

I'm not saying that it is not possible, I'm just saying that we cannot assume that it is. Correct me if I'm wrong, but I believe that Fiji itself was an optimization of AMDs older architectures. Ultimately, you will reach a point where you will have to change something fundemental in order to improve.

[Valentyn]

26 minutes ago, rasmuskrj said:

I'm not saying that it is not possible, I'm just saying that we cannot assume that it is. Correct me if I'm wrong, but I believe that Fiji itself was an optimization of AMDs older architectures. Ultimately, you will reach a point where you will have to change something fundemental in order to improve.

It's all GCN architecture.

[JediFragger]

35 minutes ago, rasmuskrj said:

Correct me if I'm wrong, but I believe that Fiji itself was an optimization of AMDs older architectures.

 

No it wasn't, it was literally just a double-sized Tonga with HBM.

[rasmuskrj]

12 minutes ago, JediFragger said:

 

No it wasn't, it was literally just a double-sized Tonga with HBM.

Wouldn't that be in the realm of "optimization"?

 

I'm not really that knowledgeable regarding the AMD achitectures, but what I'm saying is that whether or not it was possible for AMD to simply improve the Fiji architecture clock rates by 60%, call it a day and compete with a 1080 ti is something that only the AMD engineers know.

But I'm pretty sure that If they'd been able to do that, then they would have done so. Probably much earlier too. 

[leadeater]

1 hour ago, rasmuskrj said:

I'm not saying that it is not possible, I'm just saying that we cannot assume that it is. Correct me if I'm wrong, but I believe that Fiji itself was an optimization of AMDs older architectures. Ultimately, you will reach a point where you will have to change something fundemental in order to improve.

Vega is still yet another optimization of GCN, but yes I agree due to the time between node shrinks and the need to stay competitive you don't just simply shrink what you have and not really do anything else.

[leadeater]

15 minutes ago, rasmuskrj said:

Wouldn't that be in the realm of "optimization"?

 

I'm not really that knowledgeable regarding the AMD achitectures, but what I'm saying is that whether or not it was possible for AMD to simply improve the Fiji architecture clock rates by 60%, call it a day and compete with a 1080 ti is something that only the AMD engineers know.

But I'm pretty sure that If they'd been able to do that, then they would have done so. Probably much earlier too. 

Do remember both Nvidia and AMD were stuck on 28nm for a long time, that was a technology limitation to do with silicon fabrication that wasn't under their control. If they could have shrank to 14nm sooner they would have, how and what they would have done if so I don't know.

 

Look at Nvidia's Tesla architecture for an example of this, that same architecture shrank through the following node sizes with only small optimizations: 90 nm, 80nm, 65nm, 55nm, and 40nm. The reason they could do this was silicon fabrication technology was improving so fast.

[rasmuskrj]

3 minutes ago, leadeater said:

Do remember both Nvidia and AMD were stuck on 28nm for a long time, that was a technology limitation to do with silicon fabrication that wasn't under their control. If they could have shrank to 14nm sooner they would have, how and what they would have done if so I don't know.

 

Look at Nvidia's Tesla architecture for an example of this, that same architecture shrank through the following node sizes with only small optimizations: 90 nm, 80nm, 65nm, 55nm, and 40nm. The reason they could do this was silicon fabrication technology was improving so fast.

That is true, but I would assume that what came out when AMD put the GCN architecture on 14nm was the Polaris cards. What limited the Fiji's max clockspeed might have something along the lines of its powerdraw.

 

I might be wrong though.

[leadeater]

15 minutes ago, rasmuskrj said:

That is true, but I would assume that what came out when AMD put the GCN architecture on 14nm was the Polaris cards. What limited the Fiji's max clockspeed might have something along the lines of its powerdraw.

 

I might be wrong though.

28nm uses more power than 14nm FinFET so there is a natural power usage drop in that.

[cj09beira]

10 minutes ago, rasmuskrj said:

That is true, but I would assume that what came out when AMD put the GCN architecture on 14nm was the Polaris cards. What limited the Fiji's max clockspeed might have something along the lines of its powerdraw.

 

I might be wrong though.

your max clock speed is dependent on a few things,

1 is the longest path a signal would have to make in a single cicle,

2 how fast can the transistors switch,

3 how much power the transistors need to switch that fast

and with fiji being such a large chip, i bet there was a few paths that just were too long, 

my 480 can reach 1500+(like 1560) if its cold enough for a short period of time, my 7870 have reached almost 1400mhz, 

amd in the past few gens didn't give max frequency too much importance, and they should have.

 

[cj09beira]

to clarify, polaris clocks better mostly because of the node not design, vega clocks better because of design

[rasmuskrj]

4 minutes ago, cj09beira said:

your max clock speed is dependent on a few things,

1 is the longest path a signal would have to make in a single cicle,

2 how fast can the transistors switch,

3 how much power the transistors need to switch that fast

and with fiji being such a large chip, i bet there was a few paths that just were too long, 

my 480 can reach 1500+(like 1560) if its cold enough for a short period of time, my 7870 have reached almost 1400mhz, 

amd in the past few gens didn't give max frequency too much importance, and they should have.

 

I agree, this was kind of what I was getting at in my first post

[JediFragger]

10 hours ago, rasmuskrj said:

Wouldn't that be in the realm of "optimization"?

 

Fiji was still only GCN 1.2, no no architectural optimisations.

[laminutederire]

13 hours ago, leadeater said:

Vega is still yet another optimization of GCN, but yes I agree due to the time between node shrinks and the need to stay competitive you don't just simply shrink what you have and not really do anything else.

Nvidia is doing it with Pascal and is allegedly thinking about doing it too another time for the next generation.

Amd was planning to node shrink Vega to have an incremental improvement before Navi as well.