AMD: Speculation, Rumors, and Discussion (Archive)

[sheepdogexpress]

The 7870 upon release was only 10% faster than a 6970 which had a $350 MSRP at the time. The price of the 7870 upon release was $350. The die size of the 7870 was 212mm2.

Pricing a 14nm card that is 232mm2 for $200 dollar is not profitable. The cost of wafers nearly doubled this time around meaning per transistor, a 232mm2 die cost as much to manufacture as a 438mm2 with some saving with yields for the 14nm one. AMD is not a benevolent company that wants to bring the world rainbows and help cure world hunger. Launching cards that low would kill margins to unprofitable levels and kill them. They have to recover their cost on R and D and pay for overhead, etc.

Because of the cost of manufacturing dies on the 14nm/16nm process, it will be in both AMD and Nvidia best interests to raise videocard pricing. Launching 14/16nm cards with pricing 40% lower than last gen's 28nm initial pricing, which gave a transistors savings cost of 80%, would be a sure fire way to crater the GPU industry into a unprofitable market.

 

[Jawed]

My impression from last night's stream/slides/Ryan Shrout's video interview:

Polaris 10, "480X", is going to be between R9 390x and Fury X perforrmance and be priced at up to $250 with 256-bit, 8GB memory (4GB would be a fatal mistake). Polaris 11 looks like the lowest performance chip and I dare say we're hoping they'll both be available to buy in June. There may be a third Polaris: 256-bit GDDR5X-based for autumn "back to school" launch with say 40%+ performance over Fury X.

I expect none of the Polaris based cards to have any variant of HBM.

Many people see Vega on that slide and conclude that it's the "enthusiast" or "ultra enthusiast" chip: it's higher up the slide, it must be the fastest! Except that slide shows performance per watt.

To me this implies that Polaris is a tweaked GCN which relies heavily upon the change to FinFET for PPW gains. To get another big step up from Polaris to Vega in terms of PPW would seem to require a more thorough change in architecture. NVidia changed ALU architecture amongst the things it did to get PPW gains - they didn't just magically appear.

Vega appears to be HBM2. Raja said to Ryan that HBM2 will only come when it's ready for the mainstream. This implies to me that Vega, a range of 2 or 3 chips, will depend entirely on HBM2. HBM2 will, itself, lead to PPW gains. But HBM2 on its own won't deliver all the gains required. And somewhere in the Vega range will be the true successor to Fiji, perhaps as far off as 1 year from now.

Double precision looks like a victim (it hurts PPW for consumers): I am tempted to see this as an indication that AMD is now chasing VR in preference to compute. I get the feeling (perhaps it's just their marketing doing what it's meant to do) that AMD sees VR as its CUDA moment. A completely new market, beyond gaming, blah blah.

 

[Kaotik]

My impression from last night's stream/slides/Ryan Shrout's video interview:

Polaris 10, "480X", is going to be between R9 390x and Fury X perforrmance and be priced at up to $250 with 256-bit, 8GB memory (4GB would be a fatal mistake). Polaris 11 looks like the lowest performance chip and I dare say we're hoping they'll both be available to buy in June. There may be a third Polaris: 256-bit GDDR5X-based for autumn "back to school" launch with say 40%+ performance over Fury X.

I disagree, it would simply kill Radeon Pro Duo (from both consumer & AMD point of view - why use expensive Fijis when you can use cheaper Polaris-somethings and be faster too?)

Any possible Polaris outside 10 and 11 could really only fit under 11 (replace oland/iceland) or between 10 & 11 (is the gap big enough?)

 

[Silent_Buddha]

I disagree, it would simply kill Radeon Pro Duo (from both consumer & AMD point of view - why use expensive Fijis when you can use cheaper Polaris-somethings and be faster too?)

Any possible Polaris outside 10 and 11 could really only fit under 11 (replace oland/iceland) or between 10 & 11 (is the gap big enough?)

Inventory clearing. This way they don't have to worry about single GPU Fiji going against Polaris. If Polaris is as good as they claim, they are likely to sell as many as they can manufacture. And 2x single cards have always performed better than dual GPU cards and generally been cheaper as well, so that's nothing new. As well, the Radeon Pro Duo with three power connectors and what looks to be a rather large radiator (looks to be larger than 2x fury x radiator) could mean they aim to push each core harder than they pushed it for Fury X.

People that are in the market for a Dual GPU card have never been shy about paying more for it than dual cards that would offer better performance. I'm not convinced this will sell many units, but perhaps it doesn't need to. If they've stopped production of Fiji chips they just need it to clear out as much inventory as possible.

Regards,
SB

 

[Kaotik]

To me this implies that Polaris is a tweaked GCN which relies heavily upon the change to FinFET for PPW gains. To get another big step up from Polaris to Vega in terms of PPW would seem to require a more thorough change in architecture. NVidia changed ALU architecture amongst the things it did to get PPW gains - they didn't just magically appear.

AMD already stated that it's Polaris that will tweak the architecture, though.

Also, just watching the video, Koduri doesn't say there will be third (or more) Polaris-chip, just that if they would build another Polaris-chip, it would be Polaris 12 regardless of how it performs compared to 10 & 11.

 

[Grall]

HBM DualLink = 8GB

Ehm... Clamshell HBM!

 

[Jawed]

I disagree, it would simply kill Radeon Pro Duo (from both consumer & AMD point of view - why use expensive Fijis when you can use cheaper Polaris-somethings and be faster too?)

Pro Duo isn't a consumer card. Also AMD seems intent on giving away lots of these to places that run the Crytek VR course.

AMD already stated that it's Polaris that will tweak the architecture, though.

I didn't say otherwise. Merely that Vega may have to be a substantial architectural change, because it doesn't have the benefit of the introduction of FinFET. (Though it apparently does have the benefit of being solely HBM2 based.)

Also, just watching the video, Koduri doesn't say there will be third (or more) Polaris-chip, just that if they would build another Polaris-chip, it would be Polaris 12 regardless of how it performs compared to 10 & 11.

I didn't say there would be a third chip.

I merely pointed out that GDDR5X provides the opportunity to make such a card. It might be fast enough to persuade people who wanted something faster than Fury X.

Another question: will Polaris fully implement 12.1? Or will we be waiting for Vega to see that?

 

[msia2k75]

I suspect Polaris 10 & 11 are both designed to be less bandwidth hungry compared to previous GCN cards. AMD will fit them with bog standard GDDR5 memory bus. A bit like Maxwell.

EDIT:

HBMv1 on these cards doesnt make any sense to me. Too costly and unneccesary for the performance targeted.

 

[eastmen]

I suspect Polaris 10 & 11 are both designed to be less bandwidth hungry compared to previous GCN cards. AMD will fit them with bog standard GDDR5 memory bus. A bit like Maxwell.

EDIT:

HBMv1 on these cards doesnt make any sense to me. Too costly and unneccesary for the performance targeted.

depends on how much HBM1 memory AMD had to buy and has left over lol

 

[silent_guy]

Pricing a 14nm card that is 232mm2 for $200 dollar is not profitable.

It should be totally fine.

The cost of wafers nearly doubled this time around meaning per transistor, a 232mm2 die cost as much to manufacture as a 438mm2 with some saving with yields for the 14nm one.

16nm and 14nm should be nicely past their immature stages. And the presence of competition between TSMC, Samsung, and GF should have helped too with the 'nearly doubled' problem.

 

[bgroovy]

Ehm... Clamshell HBM!

I'm not sure you understand how interposers work!

 

[CarstenS]

Vega confirmed for late 2016 / early 2017

Captain Obvious noting two [edit: actually, three] things here:
1st - This is a marketing slide
2nd - No metric is specified (could as well be DGEMM as it could be 3DMark Score)
3rd - 28nm GPUs are placed and described in a way that it could (see 1st) refer to pre-HMB [edit: HBM] as well.

That in turn could mean, Polaris with GDDR5 gets 2.5x the efficiency of an unspecified pre-HBM GPU by virtue of architectural changes and FinFET while Vega adds to that with the added efficiency of HBM.

Captain Obvious of course playing devil's advocate here.

 

[jacozz]

Yes, I'm with CarstenS on this one.
The slide probably indicates that the 28 nm gpu = some model from the 300-series, (Hawai or Tonga), not Fiji. Now the slide makes sense. Polaris with GDDR 5, and Vega with HBM2.
Polaris 10 is made to be good enough to handle VR but at a much cheaper price than 390/x. HBM1 is probably to expensive for such a "VR to the masses" kind of product.

 

[CSI PC]

Raja Koduri mentions in the interview with Ryan Shrout that they will not bring HBM2 out until

it is ready for the mainstream, when it is ready for a larger range of gamers

I think one can exclude them using Samsung as I am sure they are pretty much aligned/tied into SK Hynix.

Cheers

 

[Grall]

I'm not sure you understand how interposers work!

I'm not sure you understand how jokes work.

 

[AlNom]

Is it like the 3 sea shells?

 

[Deleted member 13524]

I think one can exclude them using Samsung as I am sure they are pretty much aligned/tied into SK Hynix.

Where are all of Samsung's HBM2 chips being sent to, then?

Is it like the 3 sea shells?

Lol only a moron from last century wouldn't know how those work.

 

[jacozz]

Where are all of Samsung's HBM2 chips being sent to, then?

Nvidia?
According to the interweb, GTX 1080 will be out in a couple of weeks!!!
Pascal Titan Gigantus will be out with HBM2 in a few months as well

 

[CSI PC]

Where are all of Samsung's HBM2 chips being sent to, then?

Well not AMD after I specifically quoted Raja Koduri who mentions they are not going with HBM2 until it is mainstream
Also AMD will have a contract with Hynix, I doubt they can afford or make it cost effective to have a contract with them and also Samsung for the 1st batch; which is emphasised by Raja mentioning cost and not until "mainstream".

I do think NVIDIA will have the Samsung contract, but then I also feel NVIDIA like AMD are having problems with upper tier GPUs and their larger size/power requirements and why there is a lot of noise from news and rumours regarding mobile and low power/efficiency models for now.
Cheers

 

[Entropy]

The 7870 upon release was only 10% faster than a 6970 which had a $350 MSRP at the time. The price of the 7870 upon release was $350. The die size of the 7870 was 212mm2.

Pricing a 14nm card that is 232mm2 for $200 dollar is not profitable. The cost of wafers nearly doubled this time around meaning per transistor, a 232mm2 die cost as much to manufacture as a 438mm2 with some saving with yields for the 14nm one. AMD is not a benevolent company that wants to bring the world rainbows and help cure world hunger. Launching cards that low would kill margins to unprofitable levels and kill them. They have to recover their cost on R and D and pay for overhead, etc.

Because of the cost of manufacturing dies on the 14nm/16nm process, it will be in both AMD and Nvidia best interests to raise videocard pricing. Launching 14/16nm cards with pricing 40% lower than last gen's 28nm initial pricing, which gave a transistors savings cost of 80%, would be a sure fire way to crater the GPU industry into a unprofitable market.

I do not agree either with the basic premise (double cost per wafer) or the conclusion.
Just about all the material I have seen are quite clear that the cost increase at the 16/14nm node are mostly connected to verification, software and to some extent design, i.e. fixed costs (that will drop over time).
Thus, what these early designs need is volume to amortize the fixed costs. Relatively small dies not only improves yields a bit, but above all, improves marketability and thus projected total volume. Of course cream will be skimmed from early adopters which will help deplete inventories of previous generation products, but this has to be balanced against getting production volumes up. AMD is not operating in a market vacuum either. If nVidia or Kaby Lake (fat chance) are too competitive, then they risk not getting their volumes up which would be a disaster. So it might be a good idea to instead price their new product to sell while it is fresh, and sell old parts at cost to clear inventories that way. With volume and hopefully market share up, they are in a much better competitive position going forward. In mobile space design wins begets design wins, and gains in mind share can only benefit the market impact of their HBM2 based products.