Vega/Navi Rumors (Updated)

[PeterScott]

Its obvious that GCN has hit its limits and a new architecture is badly needed for AMD to stand any chance of competing in the high end GPU market in the future. AMD does not need new blood as they have the talent. They need strong leadership and bold decisions. Going for a clean sheet design is the only option for the long haul. AMD needs to try and emulate what they did with Zen.

I have a hard time reconciling "They have the Talent", with the existence of Vega.

Also for Zen, didn't they bring in one of the industries best CPU design gurus?

 

[Elixer]

Its obvious that GCN has hit its limits and a new architecture is badly needed for AMD to stand any chance of competing in the high end GPU market in the future. AMD does not need new blood as they have the talent. They need strong leadership and bold decisions. Going for a clean sheet design is the only option for the long haul. AMD needs to try and emulate what they did with Zen.

Isn't that what Navi is supposed to be?

We still don't know what is wrong with Vega though, could be hardware bugs that they had to resolve in software or... or... so, it still is too early to say that GCN hit a wall.

 

[EXCellR8]

is there actually anything wrong with Vega though? i mean it's inefficient and falls short of what everyone was expecting but...

how is the power draw compared to an OC R9 390X or Pro Duo?

 

[PeterScott]

Isn't that what Navi is supposed to be?

We still don't know what is wrong with Vega though, could be hardware bugs that they had to resolve in software or... or... so, it still is too early to say that GCN hit a wall.

I think Navi is just GCN in smaller "scalable" dies joined together.

I remember AMD claiming Navi would tape out before the end of the 2017, but AMD claims...

 

[raghu78]

I have a hard time reconciling "They have the Talent", with the existence of Vega.

Also for Zen, didn't they bring in one of the industries best CPU design gurus?

Yeah AMD brought Jim Keller to lead the CPU team. But a modern high performance CPU core is not designed by one person. Keller's experience in high performance CPU design would have been invaluable to AMD. I am sure he contributed at a very high level on Zen design but his role as Chief Architect of Microprocessor cores was more of leadership and of a facilitator to make sure the necessary processes, tools and structures were in place so that the team can deliver on such a big project. Keller in this 2014 interview talks of a strong and capable team

https://www.youtube.com/watch?v=SOTFE7sJY-Q

Isn't that what Navi is supposed to be?

We still don't know what is wrong with Vega though, could be hardware bugs that they had to resolve in software or... or... so, it still is too early to say that GCN hit a wall.

Both Fiji and Vega are not area and power efficient for gaming. Vega has regressed on Fiji in area and power efficiency for gaming and thats very a clear sign that GCN atleast in its current form has hit some major issues with improving performance and efficiency.

Nvidia is meanwhile churning out GPU generations like clockwork with outstanding perf/watt and perf/sq mm. Nvidia has no problems improving perf/clock/cuda core (Maxwell) or increasing clocks massively without regressing on efficiency (Pascal). AMD has a trouble doing both of these and whatever the reason for this they have to come up with a solution. Nvidia's Volta will bring huge efficiency gains again especially in perf/watt vs Pascal.

 

[maddie]

Its obvious that GCN has hit its limits and a new architecture is badly needed for AMD to stand any chance of competing in the high end GPU market in the future. AMD does not need new blood as they have the talent. They need strong leadership and bold decisions. Going for a clean sheet design is the only option for the long haul. AMD needs to try and emulate what they did with Zen.

What's the current estimate for a new design?

Always felt and have posted that Navi was when things changed substantially. To me, an early 7nm scalable design is very bold. Design costs being now spread across both the CPU and GPU divisions, namely the Infinity fabric, is a further productivity enabler.

Isn't that what Navi is supposed to be?

We still don't know what is wrong with Vega though, could be hardware bugs that they had to resolve in software or... or... so, it still is too early to say that GCN hit a wall.

Good luck making that argument in this environment.


All crying
1) Fire RTG leadership
2) Abandon high-end gamer market
3) Divest RTG

I can't help remembering posts before we started getting good info on Zen. Pretty much the same about the CPU section as are now being said about the graphics division and Bulldozer was much more behind than Vega is to the competition.

 

[maddie]

I think Navi is just GCN in smaller "scalable" dies joined together.

I remember AMD claiming Navi would tape out before the end of the 2017, but AMD claims...

Good one.

Actually, great one.

 

[Snarf Snarf]

I have a hard time reconciling "They have the Talent", with the existence of Vega.

Also for Zen, didn't they bring in one of the industries best CPU design gurus?

They did the same for Radeon with Raja Koduri, he is an expert in GPU design and has been a part of some extremely successful projects before. When Lisa Su brought back Keller and Koduri she made the right call. We have no idea what kind of awful budget he was given, or how much of the design was already in place when he started trying to fix it. Dr Su has so far made all the right calls for the business side of things, which unfortunately for us, isn't the high end gaming market right now. RTG is going to try to penetrate data center and HPC market where the margins are much better and their architecture performs pretty well still.

If you look at the full picture the TBR system looks to be intended to help bandwidth starved chips the most (APU's, or smaller laptop GPU's). Packed math is entirely a compute feature, and the HBCC system with infinity fabric seems to have been designed with the SSG principle to begin with. While some of these things we hoped *might* improve gaming performance, if you take a step back, it seems none of these features were ever designed with gaming in mind, but to meet other needs in different markets that AMD wants to penetrate. It's a shrewd business move that upsets quite a few people around here, but it's most likely the wisest choice given what limited resources the company has been operating with for the last 8 or so years. Here's to hoping their long play is going to pay off, the goal for 2017 was supposed to be single digit market share growth with gross margins in the ~35% range to allow profitability and so far they look on track to meet those goals. We can only hope as enthusiasts that AMD can start pouring more R&D into RTG to allow them to return next generation with a more competitive design.

 

[caswow]

I have a hard time reconciling "They have the Talent", with the existence of Vega.

Also for Zen, didn't they bring in one of the industries best CPU design gurus?

vega has transistors wasted for things other than gaming. ssg comes to my mind. ssg can be huge for amd.

 

[PeterScott]

They did the same for Radeon with Raja Koduri, he is an expert in GPU design and has been a part of some extremely successful projects before. When Lisa Su brought back Keller and Koduri she made the right call. We have no idea what kind of awful budget he was given, or how much of the design was already in place when he started trying to fix it.

Didn't he return in early 2013. That's more than 4 years ago. I really don't think you can lay Vega's issues at the feet of his predecessor. He should have been there for a big part of Fiji, let alone Vega.

 

[Snarf Snarf]

Didn't he return in early 2013. That's more than 4 years ago. I really don't think you can lay Vega's issues at the feet of his predecessor. He should have been there for a big part of Fiji, let alone Vega.

The average is about 4-5 years from paper to product launch, so Polaris and Vega are the only projects he's had meaningful impact on. Even then both were already on the table when he got hired on, the consensus is that Vega has his fingerprints on it, but how much budget he had and manpower he had access to is a complete unknown. Navi will likely be his brainchild that he has had 100% control over from start to finish, this will likely be a high clocking slimmer design with modules that will be able to be tacked onto the underlying uarch to address multiple markets with a "core" design so to speak fused on an interposer using next generation infinity fabric.

 

[Stuka87]

All crying
1) Fire RTG leadership
2) Abandon high-end gamer market
3) Divest RTG

Thing is the current RTG leadership took over AFTER Vega design work was started. During the live stream there was a whole bunch of people saying to fire Raja, but Vega isn't his fault.

Not to mention he most likely had such a small budget to work with starting over was completely out of the picture.

 

[tviceman]

I can already see a #waitfornavi movement forming. The mountains Navi will have to climb are going to be absurd. Koduri is not a Good and it's obvious from Vega's many faults he can't work miracles. Volta is going to be >3x more efficient than current Polaris and Vega. AMD going to have to baby step themselves back into competition, just like they did with the HD 3000 series after missing the boat with the HD 2000 series and hope Nvidia doesn't keep their release cadence up.

 

[Elixer]

is there actually anything wrong with Vega though? i mean it's inefficient and falls short of what everyone was expecting but...

That is tricky to answer, since we don't have actual cards + drivers that "enable" everything.

On paper, Vega seems to have been a good design, for both compute & gaming.
We know they pushed Vega to the limit, thus ruining any performance/watt benefit. Which means, that it most likely is somewhat efficient, as long as you don't pump so much voltage into it.

Can't really say anything more, except we will know more in ~2 weeks.

 

[IEC]

I can already see a #waitfornavi movement forming. The mountains Navi will have to climb are going to be absurd. Koduri is not a Good and it's obvious from Vega's many faults he can't work miracles. Volta is going to be >3x more efficient than current Polaris and Vega. AMD going to have to baby step themselves back into competition, just like they did with the HD 3000 series after missing the boat with the HD 2000 series and hope Nvidia doesn't keep their release cadence up.

You said basically the same things with the R9 series and the Polaris series vis-a-vis nVidia. It wasn't true then, and it still isn't true now. The gap is far less than people make it out to be.

Does nVidia have superior GPUs for gaming, and perf/W in gaming? Yep. Is RTG that far behind? I really don't think so. A day (year) late and a dollar (half tier) short, maybe, but GTX 1080 level performance in gaming is nothing to sneeze at. Especially with a chip that was apparently designed for datacenter applications rather than gaming.

Vega's most important characteristics are:
1) AMD finally has a GPU that will sell well outside the gaming market (read: fat margins)
2) AMD has finally returned to the high end

In the long run it doesn't matter much that they didn't retake the performance crown - it matters that they will make it to their goal of IF-linked MCM GPUs. Because that will be the "Ryzen" moment for RTG.

 

[PeterScott]

You said basically the same things with the R9 series and the Polaris series vis-a-vis nVidia. It wasn't true then, and it still isn't true now. The gap is far less than people make it out to be.

Does nVidia have superior GPUs for gaming, and perf/W in gaming? Yep. Is RTG that far behind? I really don't think so. A day (year) late and a dollar (half tier) short, maybe, but GTX 1080 level performance in gaming is nothing to sneeze at. Especially with a chip that was apparently designed for datacenter applications rather than gaming.

Vega's most important characteristics are:
1) AMD finally has a GPU that will sell well outside the gaming market (read: fat margins)
2) AMD has finally returned to the high end

In the long run it doesn't matter much that they didn't retake the performance crown - it matters that they will make it to their goal of IF-linked MCM GPUs. Because that will be the "Ryzen" moment for RTG.

It isn't that good either. AMD can't even build a top tier competitor (1080 Ti) and they can only "compete" in the lower ranges by throwing excess resources at them(smaller margins to go with AMDs lower volume). AMD uses a bigger die size, and consumes more power at every tier, and at the Vega/1080 Level they are at a Titan Sized Die, and using expensive HBM memory to equal a device with a much smaller die and cheaper memory.

Linked MCM GPUs is no panacea, and there are clear signs NVidia is also working on similar solutions.

Also people need to stop treating infinity fabric like it is magical. It's just an interconnect. Everyone has them, they just dont make a big deal about it.

Interconnects are trivial. Making a good GPU is hard.

 

[4K_shmoorK]

so my dear forum members. I hope more info was out until today.
So I got one little question

1070 or vega? 1080 or vega? Vega no?
monitor resolution 3440x1440
Thank you

Bench for wait marks.

 

[tviceman]

You said basically the same things with the R9 series and the Polaris series vis-a-vis nVidia. It wasn't true then, and it still isn't true now. The gap is far less than people make it out to be.

It was true then, and it's more true now. I don't need to look any further than market share, GPU use cases (data centers, pro, consumer, mobile), and profits. Polaris did nothing to change AMD's position with any of those categories. The one and only market Polaris is dominating is mining; a very niche market based on speculation and which can crash at any moment. Without mining Polaris would be at selling at reduced prices just to maintain that 25-30% discrete market share.

Does nVidia have superior GPUs for gaming, and perf/W in gaming? Yep. Is RTG that far behind? I really don't think so. A day (year) late and a dollar (half tier) short, maybe, but GTX 1080 level performance in gaming is nothing to sneeze at. Especially with a chip that was apparently designed for datacenter applications rather than gaming.

A GTX 1080 is a great performing card, but it's been out for 15 months and has been eclipsed by 3 faster cards. AMD is just now almost reaching that level of performance, but with 60% higher power draw, and a new record for stock power draw with a single GPU card. Die size wouldn't matter so much if it was a much higher volume card or lower power. But the trifecta of last place in all 3 categories seals the deal for showing the world just how far behind AMD is.

Vega's most important characteristics are:
1) AMD finally has a GPU that will sell well outside the gaming market (read: fat margins)
2) AMD has finally returned to the high end

AMD's Vega put pressure on Nvidia to unlock artificially locked features / performance on Titan X. Today Nvidia responded and now Vega RX looks way, way less compelling. It loses in most pro benchmarks now. Vega quite literally compelled Nvidia to unlock pro features on it's prosumer card, and because of that RX Vega is no longer compelling looking in most prosumer situations.Hooray for competition, but it seems like Nvidia ALWAYS has an answer for anything AMD does.

In the long run it doesn't matter much that they didn't retake the performance crown - it matters that they will make it to their goal of IF-linked MCM GPUs. Because that will be the "Ryzen" moment for RTG.

You're right; the performance crown isn't that important. Perf/$ is important and perf/w is important. AMD isn't winning at one and is losing VERY bad at the other. Notebooks are entirely a foregone conclusion. SFF is nearly a foregone conclusion. Lead time on market is long gone. Performance advantage is long gone. Price advantage is gone. And costs are up - Vega is trying to sell at the same price as much smaller chip that has a cheaper power delivery system and cheaper vram.

Worst of all? Consumer Volta's are very likely 5-8 months away. Nvidia is on a very steady, very predictable cadence. GV104 will be ~50% faster than GTX 1080, and GV102 will be ~50% faster than GV102. Nvidia's GV206, probably a ~250mm2 chip will be about as fast as Vega 56 at 115w of power consumption!

To drop the final bomb, Nvidia is moving to MCM's too. It's nothing unique or special to AMD and MCM designs won't matter if AMD doesn't massively improve their architecture because power walls will get hit just as hard with MCM's as they will with single chip designs.

 

[Elfear]

I'm not, at all. I was being sarcastic. It's the reality of what Vega is and my prediction on the impact it will have. I wish Vega was either 20% faster or 50+ watts more efficient because then it'd be in a way more competitive position.

Glad to hear it. Strong competition is good for all of us. And I agree with you, Vega should have at least been on par efficiency-wise with Pascal after coming out much later.

is there actually anything wrong with Vega though? i mean it's inefficient and falls short of what everyone was expecting but...

how is the power draw compared to an OC R9 390X or Pro Duo?

Something doesn't add up with Vega right now. According to AMD's own comparison it's basically on par with a 1080FE or ~28% faster than a Fury X according to TPU's latest benchmarks.

It appears that DSRB is enabled in those AMD comparison charts as they even show the benefits in one of the slides. So DSRB is providing some kind of performance increase and I would estimate somewhere in the order of 5-10% depending on how much memory bandwidth a game requires. If that is the case, without DSRB, Vega is ~18-23% faster than FuryX with ~50% higher clocks.

I sure hope we learn one day what that bottleneck is because it's rather mind-boggling right now.

 

[Paratus]

Glad to hear it. Strong competition is good for all of us. And I agree with you, Vega should have at least been on par efficiency-wise with Pascal after coming out much later.



Something doesn't add up with Vega right now. According to AMD's own comparison it's basically on par with a 1080FE or ~28% faster than a Fury X according to TPU's latest benchmarks.

It appears that DSRB is enabled in those AMD comparison charts as they even show the benefits in one of the slides. So DSRB is providing some kind of performance increase and I would estimate somewhere in the order of 5-10% depending on how much memory bandwidth a game requires. If that is the case, without DSRB, Vega is ~18-23% faster than FuryX with ~50% higher clocks.

I sure hope we learn one day what that bottleneck is because it's rather mind-boggling right now.

Figured out the cause of the bottleneck:

I must have put a decimal point in the wrong place or something. I always do that! I always mess up some mundane detail!

 

[Cookie Monster]

You said basically the same things with the R9 series and the Polaris series vis-a-vis nVidia. It wasn't true then, and it still isn't true now. The gap is far less than people make it out to be.

Does nVidia have superior GPUs for gaming, and perf/W in gaming? Yep. Is RTG that far behind? I really don't think so. A day (year) late and a dollar (half tier) short, maybe, but GTX 1080 level performance in gaming is nothing to sneeze at. Especially with a chip that was apparently designed for datacenter applications rather than gaming.

Vega's most important characteristics are:
1) AMD finally has a GPU that will sell well outside the gaming market (read: fat margins)
2) AMD has finally returned to the high end

In the long run it doesn't matter much that they didn't retake the performance crown - it matters that they will make it to their goal of IF-linked MCM GPUs. Because that will be the "Ryzen" moment for RTG.

The RTG imo is very far behind.

They are stuck with an architecture thats basically hitting its limit (or bottle necked significantly in places not disclosed) as evident by the lack of performance increase given its clock speed and power draw.
They are 14 months behind nVIDIA whom have a GPU (GP104) that does what VEGA does in 3D applications while consuming much less power, using less xtors and a much smaller die with a conventional memory interface through superior resource saving techniques. GV104 would most likely be at the levels of 1080Ti, so clearly they are far far behind.
They DONT have a well established software ecosystem which is paramount in the AI/server/HPC market especially if they want to make meaningful inroads. The hardware doesn't need to be at the levels of GV100 but if the library, compilers and toolchains are all there with good documentation/support it shouldn't be a problem. But far too many times have their software features go forgotten (does anyone remember true audio?). Its like a hardware tool without a manual.

They need a Ryzen type change to really move forward to be much more competitive. That'd mean a complete over haul instead of incremental GCN updates (please ignore the marketing buzzwords).
They need to stop with features that the developer must take advantage of and instead have features that benefit the architecture without much developer/game engine interaction.

I also think that VEGA is first and foremost a 3D accelerator. The gaming market is their core market and the bulk of revenue comes from that market just like nVIDIA. To pretend that it isn't and it was designed for GPGPU/HPC is just grasping at straws.

P.s Im in the boat that a bigger Polaris with GDDR5X/larger memory bus or just a Fury shrink with higher clocks more VRAM would have resulted in near same results (or close) for less the money and delays.

Very disappointed in VEGA but very much looking forward to a TR setup for my workstation

 

[beginner99]

is there actually anything wrong with Vega though? i mean it's inefficient and falls short of what everyone was expecting but...

how is the power draw compared to an OC R9 390X or Pro Duo?

Well the 390/390x and Fury X was the first glance at AMDs downward spiral. Remember that the 390x was pretty much just a rebranded 290x with more vram and not reference card. In fact AIB 290x perform the same as 390X and uses the same power. The 290/290x (hawaii) was a pretty good card as it was competitive in performance and price. Yes, it used a lot of power but at least it could keep up and was cheap. The reference card and mining kind of destroyed it but for the patient people it offered best performance/$ (those sub $250 deals) that only polaris manged to beat barley several years later. 390x was already a huge fail IMHO as it was a rebrand and price was raised.

What is wrong with Vega? It is less efficient than Fiji while going from 28nm planar to 14 nm finfet. I actually call that a huge achievement to regress in efficiency while doing a 1.5 node jump. That's why vega is fail.

 

[beginner99]

P.s Im in the boat that a bigger Polaris with GDDR5X/larger memory bus or just a Fury shrink with higher clocks more VRAM would have resulted in near same results (or close) for less the money and delays.

Bulldozer all over where a thuban shrink would have offered better perfromance at lower wattage and that was actually confirmed with the APU (llano) that contained shrinked STARS cores.

 

[Elixer]

They need a Ryzen type change to really move forward to be much more competitive. That'd mean a complete over haul instead of incremental GCN updates (please ignore the marketing buzzwords).

That *is* what Navi is supposed to be. The first GPU that Raja has total control over.

If it will really turn out that way, nobody except AMD knows. If it is indeed going to launch in '18, (and on the same node as Vega) then they already had first test silicon back most likely.

 

[Veradun]

P.s Im in the boat that a bigger Polaris with GDDR5X/larger memory bus or just a Fury shrink with higher clocks more VRAM would have resulted in near same results (or close) for less the money and delays.

You mean a non competitive chip for professional and datacenter space? Nope.

At this point it is clear they addressed the fastest growing, higher margin, market (learning) and started to tackle the professional space (FA, WX and especially SSG).

Unfortunately for us they chose not to push in the gaming direction.
Fortunately they can address that with the right price :>