Vega/Navi Rumors (Updated)

[kraatus77]

From what I've read, you can't just interchange GDDR5 with HBM.

Source please.

 

[96Firebird]

Source please.

Source

 

[kraatus77]

So instead of providing source from where you read about it. you are trolling ? nice. remember that for the future, it will come handy.

 

[96Firebird]

I will "remember that for the future", as you nicely edited out.

Here, Ryan Smith does a good job of explaining the need for an interposer and memory controllers for HBM. These memory controllers are located on the GPU die (PHY)...

http://www.anandtech.com/show/9390/the-amd-radeon-r9-fury-x-review

 

[kraatus77]

thanks, see you could have posted that link before instead of trying to create unnecessary jibber-jabber.
but why do that right ?

as i said. will remember for future.

 

[raghu78]

The Vega architecture is not tied to any specific memory architecture. We know for sure Vega 10 with 4096 sp will use HBM2. But Vega 11 could very well be similar to Polaris 10 chip with 2304 sp but using the Vega architecture and the associated NCU. Vega 11 is an ideal candidate for G5 and G5X. We could see the fully enabled Vega 11 chip with G5X and the cut down Vega using traditional G5. This would be akin to what AMD did with the HD 4870/HD 4850. Vega architecture will be used even in smaller chips (replacement for Polaris 11) with G5.

 

[96Firebird]

Right, that would be possible (and IMO likely). I was responding to Headfoot's prediction about a full big Vega using HBM and a cut-down big Vega using GDDR5X.

 

[Head1985]

Right, that would be possible. I was responding to Headfoot's prediction about a full big Vega using HBM and a cut-down big Vega using GDDR5X.

you cant have both HBM and GDDR5x in same SKU.
i think small vega(4096sp) will use only HBM2
Full 4096SP with HBM2- 5-10% faster than GTX1080
Cutdown 3584SP with HBM2-20-25% faster than GTX1070
350mm2 SKU
250w TDP

 

[SlickR12345]

Of course they are designed to best use a certain memory, they are certainly not tacking on HBM2 or GDDR5x. I think the big |Vega will use HBM2 and the smaller Vega will use GDDR5 and GDDR5x.

 

[tential]

Fury X came with no premium upon release. It was just 100 dollars more than a regular fury and considering the substantial amount of disabled shaders and the water cooler, it was underpriced relative to it's cost to make. Most AIB cards with a water cooler cost 100 dollars more than standards cards. In AMD case, you got a water and the full chip for 100 dollars more. In the past, AMD charged 100-150 extra to get the full chip alone. No upgraded cooler and on much smaller chip which meant the yields between the full chip and cut chip were more similar. Because there was only 100 dollar difference, unless the form factor doesn't allow it, the fury x is a no brainer over a fury.

There was never going to be an AIB fury X because it doesn't make sense from a marketing standpoint.

A so called none premium fury X would be around 599 and at that point it destroys regular fury sales. There needs to be enough difference between the cards that 1 card doesn't cannibalize the sales of another card. With the standard pricing of fury x, it already cannibalized the sales of fury. Fury x were initially sold out for a while, fury were easy to get. And considering the die size of fury, the yields on a fury x were much worse than a regular fury. Getting 100% perfect die on something that's 600mm2 is not easy. So a 50 dollar price difference is too small, even 100 dollar price difference is too small when you take in the water cooling. But for AMD, the design of the fury x was finalized and AMD probably didn't expect to have to price fury x at 650.

The problem came with Nvidia's gtx 980 ti which upon release was 650 dollars. Remember those rumors of a 800 dollar fury x, there was likely some truth to them. But the gtx 980 ti came in so it got a price cut before it was officially released. The Fury and the fury nano have realistic price point in terms of profitability due to the lack of competition. Fury X on the other hand was underpriced but AMD was forced to sell it at a lower price.

AIB FURY X can be $650... It doesn't have to come at a price cut. Let AIBs do what they want when making their cards. Don't FORCE them to only use a reference design. The price can still be set in stone at whatever it has to be at. I would buy a Sapphire Fury X Tri X (Or whatever they want to call this cooler now). Many people would. Many people simply didn't want a WC card. End of story.
In fact, AMD's audacity to believe they can produce the best cooler, and no one else can improve on it, therefore you have to use their design is a massive joke,


Pretty much invalidates your whole post.

 

[Headfoot]

I believe they may have designed it having both PHY interfaces. HBM Phy is not going to be comaptible with GDDR5X, but if they are already doing 2 dies, I think it's likely one (the smaller, or maybe both) of the two dies is equipped for GDDR5 or GDDR5X AND HBM or at least GDDR5/GDDR5X.

There's some precedent from both AMD and Intel having both DDR3 and DDR4 controllers on the same unit. Look at the latest Xeon Phi for example, this is exactly where I think they may be going. HBM2 and GDDR controllers on the same die would be similar to HMC + DDR4, maybe not simultaneously available like on the Xeon Phi, but it shows it can be done.

It would provide them an upgrade path too as HBM gets cheaper. Initial release with GDDR5X -> refresh with HBM2. But you'd have to have the die already equipped with memory controllers for both.

 

[SpaceBeer]

IIRC, Raja was saying new chips (Polaris, Vega) are designed to use both GDDR and HBM

 

[tential]

I just read the unit demod was 10% faster than a 1080 in doom. I'm guessing I'm super behind but uh.... What is the deal with this.... What chip is that? That can't be big Vega......

 

[kraatus77]

why can't ? they don't have to beat TXP if it's 10% faster than 1080 for 500-550$. because let's be honest the gap will increase in favor of vega by time just like how it happened with every GCN gpu vs it's counterparts.

and unlike 1080, it will have much better tech as HBM2. and actual working hardware for ASC. maybe something more but we don't know yet.

power efficiency wise we didn't know. maybe (most likely) it won't match 1080. but efficiency, it's not the only thing that matters. at least not for those who have previous amd nuclear reactors and are holding for vega because Fiji was not enough of an upgrade.

for others it will bring 1080's price down, that's what they wanted anyway. so a win win situation for both camps. it's not like AMD is forcing you to buy vega, right ?

 

[Arachnotronic]

I just read the unit demod was 10% faster than a 1080 in doom. I'm guessing I'm super behind but uh.... What is the deal with this.... What chip is that? That can't be big Vega......

That's Vega 10, the biggest Vega that will be relevant to the consumer market, IMO.

Should be a solid successor to Fury X. More VRAM (8GB vs 4GB), higher performance, and better efficiency.

 

[Arachnotronic]

why can't ? they don't have to beat TXP if it's 10% faster than 1080 for 500-550$. because let's be honest the gap will increase in favor of vega by time just like how it happened with every GCN gpu vs it's counterparts.

If it's 10% faster than a competing GPU and that competing GPU is priced at $600+, I don't think AMD will price lower than the competing GPU, especially considering that AMD is using neat-but-costly HBM2 tech.

for others it will bring 1080's price down, that's what they wanted anyway. so a win win situation for both camps. it's not like AMD is forcing you to buy vega, right ?

Price drops on competing products not guaranteed. Fury X did not cause a drop in price of its direct competitor.

 

[tential]

why can't ? they don't have to beat TXP if it's 10% faster than 1080 for 500-550$. because let's be honest the gap will increase in favor of vega by time just like how it happened with every GCN gpu vs it's counterparts.

and unlike 1080, it will have much better tech as HBM2. and actual working hardware for ASC. maybe something more but we don't know yet.

power efficiency wise we didn't know. maybe (most likely) it won't match 1080. but efficiency, it's not the only thing that matters. at least not for those who have previous amd nuclear reactors and are holding for vega because Fiji was not enough of an upgrade.

for others it will bring 1080's price down, that's what they wanted anyway. so a win win situation for both camps. it's not like AMD is forcing you to buy vega, right ?

.... If that's all true and amds biggest gpu is only 10% faster than a 1080 a midrange chip then forget it. There is no high end for amd for me. There's no way amd can compete. I can't even coherently think of how mad I'd be if amd pulled a stunt.

Also what in God's name does better tech matter? Hbm2? I don't care about this at all. What does this matter if the chip still performs slow. Next you'll tell me to be happy because the cardboard box the card ships in looks better

Dude I care about how fast this is and everything I'm being lead to believe right now is Vega will be a letdown. And I'm waiting for this to happen because if amd decides to continue to jerk me around as a consumer I have no reason to stay.

You can't advertise your highest end gpu as 4k ready if it only is as fast as a 1080 before a mild oc.

If it doesn't compete with titan xp, with all the extra bonuses like Hbm2 it's a massive failure.

That's Vega 10, the biggest Vega that will be relevant to the consumer market, IMO.

Should be a solid successor to Fury X. More VRAM (8GB vs 4GB), higher performance, and better efficiency.

If it's that speed what's the point of calling it flagship if it can't compete with flagship? That's what I mean why isn't this more relevant? If big Vega is that weak it's in the range of a 1080 isn't that a massive issue? Either ncu is a massive failure or amd didn't show off the fastest gpu.

I just don't want to find I waited 8 months to find out the fastest gpu I can buy from amd is not even close to a titan xp.

Edit :
I'm not ready to mentally consider that amd on a generation past the competition can't compete at the high end. I'll believe it when I see it. Then cry myself to sleep if it happens.

 

[Arachnotronic]

.... If that's all true and amds biggest gpu is only 10% faster than a 1080 a midrange chip then forget it. There is no high end for amd for me. There's no way amd can compete. I can't even coherently think of how mad I'd be if amd pulled a stunt.

Also what in God's name does better tech matter? Hbm2? I don't care about this at all. What does this matter if the chip still performs slow. Next you'll tell me to be happy because the cardboard box the card ships in looks better

Dude I care about how fast this is and everything I'm being lead to believe right now is Vega will be a letdown. And I'm waiting for this to happen because if amd decides to continue to jerk me around as a consumer I have no reason to stay.

You can't advertise your highest end gpu as 4k ready if it only is as fast as a 1080 before a mild oc.

If it doesn't compete with titan xp, with all the extra bonuses like Hbm2 it's a massive failure.



If it's that speed what's the point of calling it flagship if it can't compete with flagship? That's what I mean why isn't this more relevant? If big Vega is that weak it's in the range of a 1080 isn't that a massive issue? Either ncu is a massive failure or amd didn't show off the fastest gpu.

I just don't want to find I waited 8 months to find out the fastest gpu I can buy from amd is not even close to a titan xp.

Edit :
I'm not ready to mentally consider that amd on a generation past the competition can't compete at the high end. I'll believe it when I see it. Then cry myself to sleep if it happens.

Vega 10 is rated at what, 12.5TFLOPs? Polaris 10 is 5.8TFLOPs. So Vega 10 = 2.16*Polaris 10.

2.16 * Polaris 10 should give you a pretty reasonable idea of how Vega 10 will be positioned performance wise relative to competition and AMD's current offerings.

 

[Bacon1]

I'm being lead to believe right now is Vega will be a letdown. And I'm waiting for this to happen because if amd decides to continue to jerk me around as a consumer I have no reason to stay.

You can't advertise your highest end gpu as 4k ready if it only is as fast as a 1080 before a mild oc.

If it doesn't compete with titan xp, with all the extra bonuses like Hbm2 it's a massive failure.

So are you expecting Vega to be faster than Titan XP while costing under half of it's price, otherwise it's a failure? Or would you be happy paying over $800 if it's as fast as Titan XP?

 

[kraatus77]

It's pretty simple, set absurd expectations from vega and call it a failure later it it doesn't meet any.

when amd competed they didn't got any reward so why now ? i actually applaud amd for not giving anything for 1080 from their stable. you treat amd as they are only here as a catalyst for low prices and still badmouth it. well guess what ? no candy for you to waste anymore as no matter what you won't buy that candy.

you people don't really deserve it. those who do, they will appreciate and buy it. just wait and watch.

 

[Arachnotronic]

It's pretty simple, set absurd expectations from vega and call it a failure later it it doesn't meet any.

This, sadly, applied to all unreleased chips. If the chip isn't out yet and not much is known about it, it's very easy to come up with a "dream" chip in one's own mind.

Just take today's chips, imagine that they are X% better in some tasks, have a set of features that today's products don't have, and set the price at something lower than today's products in your mind. Great mental picture, but the technical and economic constraints of the real world don't usually want to play ball

 

[HurleyBird]

It's pretty simple, set absurd expectations from vega and call it a failure later it it doesn't meet any.

I don't think it's absurd to think that a huge die with super expensive memory tech coming out almost a year later should be near GP102 given AMD and Nvidia's historical competitiveness.

That said, I have a feeling that Vega could end up being another Polaris level disappointment. Too many subtle warning signs at this point. I don't doubt that AMD eventually wants to aim for the jugular again, especially with Zen looking so good. Good things would happen to the AMD brand if they were very competitive in both spaces at the same time. Hopefully AMD starts trying again with Navi. Well, hopefully I'm wrong about Vega altogether...

 

[SpaceBeer]

I still don't see why Polaris is failure? Today we are paying $200 (RX 480 4GB) to get the same performance we needed to pay $330 last year (GTX 970). What's the problem with that? Just because nVidia offers similar (better?) products at the same price doesn't mean it's a failure

 

[Arachnotronic]

I still don't see why Polaris is failure? Today we are paying $200 (RX 480 4GB) to get the same performance we needed to pay $330 last year (GTX 970). What's the problem with that? Just because nVidia offers similar (better?) products at the same price doesn't mean it's a failure

Some people are just never happy The amount of GPU power that one can buy today for a relatively modest sum is just insane.

 

[Samwell]

I don't think it's absurd to think that a huge die with super expensive memory tech coming out almost a year later should be near GP102 given AMD and Nvidia's historical competitiveness.

Just that Vega 10 won't be a really huge die. It should be somewhere between 350-400 mm². In between GP104 and GP102 and there we should expect its performance. Then we have Vega 11 below it, which is mostly important for notebooks, because amd is loosing there against pascal. The real big Vega is Vega20, but this gpu won't come in 2017. None of the Vegas for next year have 2:1 Double Precision capability, but Vega 20 should have it and probably also more shader as it will be the next HPC offering (MI25 is just for Deap Learning).