Vega/Navi Rumors (Updated)

[T1beriu]

tldr; There's no faster RX (consumer) Vega coming than air-cooled Frontier Edition.

I've read too many posts stating that Vega consumer cards are going to be clocked higher because FE is a workstation card because these types of cards are always clocked lower than consumer cards because TDP, power consumption and stability are more important in this sector. Optimistic posts say that RX Vega is going to be faster clocked, 1700-1800 Mhz, thus 15-20% more powerful that the Ti.

Let me remind you that there's going to be a water cooled FE that's probably the highest clocked Vega card we'll see, more performant that standard consumer cards (that's why some slides show 12.5 and some ~13 TFLOPS - ~4% OC). This water cooled FE is mostly similar to Fury X, clocked to the limit with 50-75 Mhz extra OC capabilities. The only reason to push water-cooling, in my opinion, is to provide a fighting chance against 1080 Ti by pushing a very aggressive OC. Why provide AIO cooling if you're not pushing the card to its limits? Why make it so hard to mount this card because its radiator and make it almost impossible to use more than 2 cards in a system because of this?

My prediction is that the top tier RX Vega is going to have the performance of air cooled FE (12.5 TFLOPS) with just 8GB of HBM2, with some brave AIBs pushing for ~3% OC.

Stop dreaming and be realistic.

 

[maddie]

But it does seem like it's almost a pride thing with AMD and HBM/2 at this stage. I'd guess after the Fury HBM1 shortage it was too late to do much about it, but doesn't HBCC imply they have at least somewhat abstracted memory?
So if HBCC is a sort of abstraction layer away from the actual memory, then swapping the HBM2 and GDDR5/X should be possible. And surely the lower end Vega is not going to be using HBM2. If it is not possible to use HBCC with GDDR5/X, then someone back at the initial design stage of Vega didn't do proper risk management and believed Hynix's optimistic roadmap.

AFAIK, the memory access properties of the two are completely different. HBCC needs a fine grained, low latency, high bandwidth memory to work efficiently. It would be fairly ineffective with GDDR5/X.
Yes, a big risky bet that's causing problems. The problem with loosing so much ground is that you have to bet big to recover and that is VERY risky. Think of chaotic complex systems and tipping points.

Having said that, to me it has become obvious that Navi is the real core of the comeback play, comparable to Zen in the CPU division. Infinity fabric, HBCC, HBM2 and scalibility [using small die on 7nm] will be a true game changer.

Polaris, whilst quite an attractive product and Vega, TBD, appear to be holding actions for Navi, the spoiler.

 

[w3rd]

People are trying to push stock prices into a technology thread... do so typically, because they do not want to talk technology.


Let be real here, AMD owns a lot of patents, that both Intel and Nvidia would love to have.
We are starting to see AMD exploit those patents and Dr Su has something special for gamers in mind. We are going to see three bits of technology come together under the AMD brand, later this year in APU. ANd in many forms using their fabric tech. I know as much as anyone can, but it is easy to spot where this is going.



I am around people who see things and since joining here I have been telling people, the RX Vega x2 is real.
And that Dr Su is trolling the industry right now, because very few people have actually seen big vega yet. That is why Dr Su & Friends are always grinning away from the cameras and don't care what the forum tards are crying about.

Work this out in your head, if this is what Dr Su is dealing with:

• After taping out and many early ES showed that baby vega is nearly equal to GTX1080, but on average 10% more powerful. (& leading the industry in power efficiency)
  
• While going in a slightly different direction of performance and uarch, big vega is 30% more power than Titan Xp. (& leading the industry in power efficiency)
• Sitting there with those two realizations, then (internally knowing & sitting on the fact), you know that using your infinity fabric you can use multiple GPUs on one package. Or even mix & match GPU on a socket.

How do you go to bed at night, knowing this? How do you roll out it out, so that is smacks the industry on it's face. I put all the pieces together in my head and my only technical concern the infinity fabric ability to clock & mature.



Knowing the contracts and if HBM2 yields are low and AMD can not release Vega in volume, and wants to wait 40 days instead to build up stockpiles, it is OK. It doesn't matter and predict that many people complaining about Vega and AMD untrustworthiness now, will be them complaining about not being able to find any for sale. Cuz they want one.


So for Dr Su, I don't think she is in a hurry, because baby Vega (in x2 sku) will destroy 4k. And AMD can pump out a ton of those at low cost. Problem is nobody has seen HDMI 2.1, or Display Port 1.3/4. So in essence why would AMD debut big vega, or NVidia release volta for that matter..?

Ironically, NVidia & AMD both have high-end cards for science, that have no need for graphic ports. We need 32" 4k OLED gaming panels yesterday... they don't exist at any price, because HDMI2.1 is a myth.
I want my 4k @ 120 w/HDMI 2.1 please, thnx.


I am a gamer & a dreamer. (Also a realist.)

 

[Cookie Monster]

As has been repeated innumerable times in this thread, adding a GDDR5(X) PHY to the die would massively affect die size. For sufficient performance, you'd need at least a 384-bit bus, if not 512 - that's 6 or 8 64-bit PHYs, each 3-4x the size of one of the current two HBM2 interfaces. If Vega without GDDR5(X) is 500-550mm2, it would easily be 600-650mm2 with GDDR5 support. Not only is it a complete waste for the highest end SKUs, but it would significantly affect yields and production costs.

Wait how do you know what the die size difference between a HBM2 memory controller vs a GDDR5/5X controller?

Its well documented that the PCB size is significantly reduced but the actual GPU die?

Having a GDDR5/5X VEGA wouldn't have been so bad because well it would have it out by now. No supply problems. The only issue i could think of is that this solution would be more power hungry (vs HBM) so the actual chip might have a lower headroom for its core clock.

Evidently we dont see any issues with GP102 and its die size is under 500mm2.

 

[Cookie Monster]

On a side note this HBCC feature somewhat reminds me of TurboCache or whatever marketing name they use to call it back in the old days.

Only difference now is that its more effective perhaps with faster non volatile memory alternatives etc?

 

[Glo.]

I see that there was a debate over RX Vega Frontier Edition, and wheter it has 16 GB of HBM2 or 8 GB HBM2 + 8 GB GDDR5 memory on chip. Well, it appears that it has 16 GB HBM2.

 

[otinane]

On a side note this HBCC feature somewhat reminds me of TurboCache or whatever marketing name they use to call it back in the old days.

Only difference now is that its more effective perhaps with faster non volatile memory alternatives etc?

I think it works as a scalar. You give it 2, it makes it 6.

 

[piesquared]

So how does HBC work without HBM? Since Vega is designed around HBC, HBCC, and HBM im not sure how people think just swapping out HBM and G5 is going to save TTM since that would seemingly require an entire redesign of the GPU. I like what they are doing with this GPU and dont care if it takes a little longer to get to market, ill buy it when it gets here.

 

[Valantar]

Wait how do you know what the die size difference between a HBM2 memory controller vs a GDDR5/5X controller?

Not the controller, the physical bus interface. There have been annotated die shots showing the difference. My memory was apparently mixing up a few numbers and details, though. A 64-bit GDDR5 interface isn't 3-4x the size, but a HBM interface is ~3-4x the density (if not more). Even if it's from WCCFTech, this image is AFAIK correctly annotated:

That is HBM1, but HBM2 shouldn't be radically different. No more than 1.5x the size, yet 16x the bus width. In other words: two HBM interfaces take up less space than three GDDR5 interfaces, yet perform far, far better (dependent on memory clock speed, of course).

 

[T1beriu]

I see that there was a debate over RX Vega Frontier Edition, and wheter it has 16 GB of HBM2 or 8 GB HBM2 + 8 GB GDDR5 memory on chip. Well, it appears that it has 16 GB HBM2.

Direct links to Videocardz's images are broken.

This is the image Glo. posted above to debunk that silly rumour that Vega FE will have 8 GDDR5 + 8GB HBM2, clearly stating that it has 16GB HBM2.

 

[OatisCampbell]

People are trying to push stock prices into a technology thread... do so typically, because they do not want to talk technology.


Let be real here, AMD owns a lot of patents, that both Intel and Nvidia would love to have.
We are starting to see AMD exploit those patents and Dr Su has something special for gamers in mind. We are going to see three bits of technology come together under the AMD brand, later this year in APU. ANd in many forms using their fabric tech. I know as much as anyone can, but it is easy to spot where this is going.



I am around people who see things and since joining here I have been telling people, the RX Vega x2 is real.
And that Dr Su is trolling the industry right now, because very few people have actually seen big vega yet. That is why Dr Su & Friends are always grinning away from the cameras and don't care what the forum tards are crying about.

Work this out in your head, if this is what Dr Su is dealing with:

• After taping out and many early ES showed that baby vega is nearly equal to GTX1080, but on average 10% more powerful. (& leading the industry in power efficiency)
  
• While going in a slightly different direction of performance and uarch, big vega is 30% more power than Titan Xp. (& leading the industry in power efficiency)
• Sitting there with those two realizations, then (internally knowing & sitting on the fact), you know that using your infinity fabric you can use multiple GPUs on one package. Or even mix & match GPU on a socket.

How do you go to bed at night, knowing this? How do you roll out it out, so that is smacks the industry on it's face. I put all the pieces together in my head and my only technical concern the infinity fabric ability to clock & mature.



Knowing the contracts and if HBM2 yields are low and AMD can not release Vega in volume, and wants to wait 40 days instead to build up stockpiles, it is OK. It doesn't matter and predict that many people complaining about Vega and AMD untrustworthiness now, will be them complaining about not being able to find any for sale. Cuz they want one.


So for Dr Su, I don't think she is in a hurry, because baby Vega (in x2 sku) will destroy 4k. And AMD can pump out a ton of those at low cost. Problem is nobody has seen HDMI 2.1, or Display Port 1.3/4. So in essence why would AMD debut big vega, or NVidia release volta for that matter..?

Ironically, NVidia & AMD both have high-end cards for science, that have no need for graphic ports. We need 32" 4k OLED gaming panels yesterday... they don't exist at any price, because HDMI2.1 is a myth.
I want my 4k @ 120 w/HDMI 2.1 please, thnx.


I am a gamer & a dreamer. (Also a realist.)

You might be right. Lisa Su would be grinning if there is a mystery chip that blows everything else away.
Might be a million other reasons too, I've been known to grin.

 

[Valantar]

(Also a realist.)

That made me laugh a little.

All joking aside, the scenario you're drawing up is compelling, even if we (unfortunately) don't have any evidence to back it up. The power efficiency stuff you're saying is especially out there - but again, would be amazing were it to be correct. If AMD manage to pull off a Ryzen-style improvement with Vega, I can see quite a few people shouting TAKE MY MONEY at them. I mean, there are quite a few already, even if what we've seen of Vega so far is competitive at best.

 

[krumme]

Lisa said 5 month ago that zen was going to be fast and vega to be very fast.

Well zen was a lot faster than most expected.

She seems very technical competent to me so either vega will be darn fast or either there is some drama comming up internally if its slow because then she have been really misinformed.

Dont know.. Raja seems like the type that drives technology forward (and my take is he will replace papermaster eventually) but also a bit optimistic and talking a lot ..so. lol.

After all we have to take her words for face value. She is the ceo after all and that does account for something. Zen is fast. Vega will be very fast.

 

[Maverick177]

Lisa said 5 month ago that zen was going to be fast and vega to be very fast.

After all we have to take her words for face value. She is the ceo after all and that does account for something. Zen is fast. Vega will be very fast.

Vega will be very fast, but compared to what? Fiji? Polaris? or GP102?

 

[maddie]

Wait how do you know what the die size difference between a HBM2 memory controller vs a GDDR5/5X controller?

Its well documented that the PCB size is significantly reduced but the actual GPU die?

Having a GDDR5/5X VEGA wouldn't have been so bad because well it would have it out by now. No supply problems. The only issue i could think of is that this solution would be more power hungry (vs HBM) so the actual chip might have a lower headroom for its core clock.

Evidently we dont see any issues with GP102 and its die size is under 500mm2.

Most of the memory interface logic for HBM resides on the HBM stack [the lowest die at present], not on the GPU die.

AMD is shifting the design concepts for use in the Navi generation, when all the synergies happen. HBCC, HBM2, infinity fabric, interposer, small core complex. Small core complex is what Navi finally introduces, completing the building blocks needed.

The Zen design philosophy is this exactly. A CCX core that is replicated as needed to achieve the desired performance. It's a really smart design methodology for when you're limited in resources and need to maximize your returns.

Points to note:
The interposer size is not limited by reticle size restrictions. Only the individual exposures on the interposer are reticle limited.
A small die on 7nm per mm^2 will be magnitudes cheaper than a large die, especially at first.
The sum total GPU logic area of a segmented GPU will not be restricted by the reticle limit on that process. They can assemble huge GPUs.
Interposers containing no logic are cheap per mm^2, contrary to what some here claim.

"Sesh Ramaswami, managing director at Applied Materials, showed a cost analysis which resulted in 300mm interposer wafer costs of $500-$650 / wafer. His cost analysis showed the major cost contributors are damascene processing (22%), front pad and backside bumping (20%), and TSV creation (14%).

Since one can produce ~286 200mm2 die on a 300mm wafer, at $575 (his midpoint cost) per wafer, this results in a $2 200mm2 silicon interposer."
http://electroiq.com/blog/2012/12/lifting-the-veil-on-silicon-interposer-pricing/

 

[GoodRevrnd]

Anyone else get the sense that AMD might be just starting to leverage its unique position as a CPU and GPU developer to deliver extra performance gains by designing synergies into its products? It seems like there are hints of this from infinity fabric, and now HBCC tapping into addon nvme drives... Makes you wonder what else they might be moving toward. Will we see a point where having a nearly all AMD platform yields substantial gains over traditional mix and match? It's a potentially massively profitable business strategy, but this is the type of move we all hate nVidia for and AMD theoretically has the ability to take it to a new extreme.

 

[Bacon1]

Anyone else get the sense that AMD might be just starting to leverage its unique position as a CPU and GPU developer to deliver extra performance gains by designing synergies into its products?

Not just you

 

[maddie]

Anyone else get the sense that AMD might be just starting to leverage its unique position as a CPU and GPU developer to deliver extra performance gains by designing synergies into its products? It seems like there are hints of this from infinity fabric, and now HBCC tapping into addon nvme drives... Makes you wonder what else they might be moving toward. Will we see a point where having a nearly all AMD platform yields substantial gains over traditional mix and match? It's a potentially massively profitable business strategy, but this is the type of move we all hate nVidia for and AMD theoretically has the ability to take it to a new extreme.

I feel that this is exactly what we are witnessing. The end of an age and beginning of another.

To digress a bit, we have a technological tour de force in V100 resulting in a 800mm^2 monster die on a custom process. Not something a small company can replicate, so design around it. Don't go head to head. That's AMD future.

 

[GoodRevrnd]

Ha, okay, I was at work during the analyst call and was pulled from my desk for most of the CPU presentation. Did not realize they finally got that explicit with infinity fabric scalability and interconnectivity.

 

[Despoiler]

I feel that this is exactly what we are witnessing. The end of an age and beginning of another.

To digress a bit, we have a technological tour de force in V100 resulting in a 800mm^2 monster die on a custom process. Not something a small company can replicate, so design around it. Don't go head to head. That's AMD future.

V100 is kinda stupid though. Nvidia's strat is an expensive, monolithic, one size fits all chip. AMD's is scalable solutions. Google's TPU 2.0 is faster, more power efficient, and more flexible given that it's an ASIC. EPYC + Vega + GTPUs allows a customer to tailor the system top to bottom for their calculation needs. They won't have too much or too little of anything. AMD will do better in the long run.

 

[zinfamous]

Anyone else get the sense that AMD might be just starting to leverage its unique position as a CPU and GPU developer to deliver extra performance gains by designing synergies into its products? It seems like there are hints of this from infinity fabric, and now HBCC tapping into addon nvme drives... Makes you wonder what else they might be moving toward. Will we see a point where having a nearly all AMD platform yields substantial gains over traditional mix and match? It's a potentially massively profitable business strategy, but this is the type of move we all hate nVidia for and AMD theoretically has the ability to take it to a new extreme.

eh, don't forget the Sony hate, too, for driving this very model into the ground, years ago. The difference here, though, is if this represents an actual improvement in performance, and the superior option compared to others. If it does, then so what? Go with the superior option or don't. Sony rarely had a superior option with their internal proprietary formats (OK--I'm talking about cables and storage and certain hardware--not Beta or BD which actually were/are superior to their competitors).

But yeah, this was AMD's goal all along with buying ATI, it just doesn't seem to have worked all that well up until now (maybe).

 

[krumme]

Vega will be very fast, but compared to what? Fiji? Polaris? or GP102?

It was mentioned in the same sentence as zen. I guess the only thing that makes sense is compared to the competition ie gp102.

I was very much surpriced of zen perf and will be here too if its the same kind of uplift. Look at amd history. But hey we can do with zen and some half weak gpu. Cant get it all at the same time. My guess is Raja needs 2 years more.

 

[CatMerc]

Wait how do you know what the die size difference between a HBM2 memory controller vs a GDDR5/5X controller?

Its well documented that the PCB size is significantly reduced but the actual GPU die?

Having a GDDR5/5X VEGA wouldn't have been so bad because well it would have it out by now. No supply problems. The only issue i could think of is that this solution would be more power hungry (vs HBM) so the actual chip might have a lower headroom for its core clock.

Evidently we dont see any issues with GP102 and its die size is under 500mm2.

https://i.imgur.com/A3DH7rr.jpg

I made the comparison a few weeks ago.

 

[Mopetar]

V100 is kinda stupid though. Nvidia's strat is an expensive, monolithic, one size fits all chip.

NVidia has a CUDA lock-in though. Even if there are better options, they can't run CUDA and it's probably more expensive to replace that code than it is to eat some extra cost.

 

[Glo.]

https://videocardz.com/69662/raja-koduri-explains-where-radeon-rx-vega-in-reddit-ama

Elmnator: Will the consumer RX version be as fast at the Frontier version?

RK: Consumer RX will be much better optimized for all the top gaming titles and flavors of RX Vega will actually be faster than Frontier version!

Most likely - increased performance from higher bandwidth.

nas360: Why the official Frontier page shows renders of the card with 2x8pin connectors and the one you were holding at the presentation has 1×8 + 1×6 connectors?

RK: I grabbed an engineering board from the lab on the way to the Sunnyvale auditorium, and that boards works well with a 6 and an 8 pin. We decided to put two 8 pin connectors in the production boards to give our Frontier users extra headroom