AMD 7nm GPU News and Rumours 2018/2019

[LePhuronn]

Polaris 30 looks to also be made at Samsung:

https://www.techpowerup.com/249669/samsung-amds-second-foundry-partner-for-polaris-30

So it appears AMD is trying to free up 14NM capacity?? It might make sense since the IO chips for Ryzen 2 are made on 14NM,and Ryzen+ is already made on GF 12NM.

Well this is interesting. You wouldn't need to free up capacity if Polaris 30 was just a legal box-ticking exercise with GloFo, but if Polaris 30 is to be used in sufficient volume to warrant 2 foundries, I'm disappointed it didn't get an update, like a GDDR6 controller.

I do wonder if Navi could also go the chiplet way,ie,some of the logic like the memory controllers,and other blocks made on 14NM or 12NM,and shader blocks on 7NM??

I don't think Navi will, but it's a possibility for Arcturus.

 

[CAT-THE-FIFTH]

Well this is interesting. You wouldn't need to free up capacity if Polaris 30 was just a legal box-ticking exercise with GloFo, but if Polaris 30 is to be used in sufficient volume to warrant 2 foundries, I'm disappointed it didn't get an update, like a GDDR6 controller.



I don't think Navi will, but it's a possibility for Arcturus.

It could be because 14NM needs to be used for Ryzen 2 - the IO die for Rome is over 400MM2 alone,and even if the one for Ryzen 2 was just for two chiplets,that alone will be over 100MM2. 12NM is also currently used for Ryzen+ CPUs too.

 

[LePhuronn]

True, my point though was Polaris 30 looked like a legal requirement to fill wafers rather than an actual intended product: literally just the 15% performance boost that 12nm brings, rather than more shaders or a GDDR6 controller, y'know actual improvements. And then to only use it as the new top card in the existing 500 series felt like it's not intended to be a high volume product or the start of a refreshed 600 series.

But bringing in Samsung to produce Polaris 30 too makes me think there's more to come from it; wafer agreements can surely be satisfied with GloFo's 14nm kit busy with Zen 2 components and Zen+ until it's EOL but AMD want to make more Polaris 30 chips without risking GloFo getting overloaded?

Pondering...

 

[Poneros]

But bringing in Samsung to produce Polaris 30 too makes me think there's more to come from it

Perhaps AMD are thinking the card will sell successfully.

 

[bru]

All this talk of a GDDR6 controller, AMD don't even use GDDR5X yet, I really thought they would with the 590, but alas it was not to be.

Don't forget they have already showed us a 7nm Vega card with 25-30% more performance than the currant Vega 64. Yes I know it is a compute orientated card, and not necessarily good for gaming, but so was Vega and Fuji before it. It would be daft for them not to bring it to the gaming market, I mean it's not as if they are swimming in cash to be able to make several different chips for different markets.

 

[4K8KW10]

All this talk of a GDDR6 controller, AMD don't even use GDDR5X yet, I really thought they would with the 590, but alas it was not to be.

Maybe HBM2 and HBM3 will become mainstream, so Radeons will never go with 5X/6.

 

[melmac]

I'm disappointed it didn't get an update, like a GDDR6 controller.

All this talk of a GDDR6 controller, AMD don't even use GDDR5X yet, I really thought they would with the 590, but alas it was not to be.

what would the point of going for an updated memory controller on the 590? It would only raise the price of the card for no performance increase.

 

[LePhuronn]

what would the point of going for an updated memory controller on the 590? It would only raise the price of the card for no performance increase.

Well that's kinda what I'm saying. There is no point in an updated controller for the RX 590, but if Polaris 30 was used in a full RX 600 series then it would make more sense. The 12nm respin has shown about a 15% increase in performance. Throw in the increased performance and lower power draw of GDDR6 and we could have a chunky 15%-20% uplift across the entire range.

 

[LePhuronn]

All this talk of a GDDR6 controller, AMD don't even use GDDR5X yet, I really thought they would with the 590, but alas it was not to be.

Don't forget they have already showed us a 7nm Vega card with 25-30% more performance than the currant Vega 64. Yes I know it is a compute orientated card, and not necessarily good for gaming, but so was Vega and Fuji before it. It would be daft for them not to bring it to the gaming market, I mean it's not as if they are swimming in cash to be able to make several different chips for different markets.

GDDR5X would need a new PCB to accomodate the pin increase.

You say AMD aren't swimming in cash so they can't afford to make different chips for different markets, but it's OK to sell Vega 20 to gamers at a ridiculous loss?

 

[LoadsaMoney]

They had to use HBM on Vega, as its a compute card.

 

[bru]

GDDR5X would need a new PCB to accomodate the pin increase.

but it's OK to sell Vega 20 to gamers at a ridiculous loss?

And you think that the fact the 590 is 12nm didn't require a new PCB, hey maybe it didn't, I really don't know. What I do know is there is no reference version so any money spent wouldn't be AMD's anyway it would be down to the AIB's.

Why on earth would they sell it at a loss? You would be looking at Vega 64 +25-30%, that puts it in the 1080ti-2070 bracket, easily worth a decent asking price.

But of course it seems that even AMD fans expect AND to sell at cheap prices even when they do have a good product.

 

[LePhuronn]

Why on earth would they sell it at a loss? You would be looking at Vega 64 +25-30%, that puts it in the 1080ti-2070 bracket, easily worth a decent asking price.

Um no, Vega 20 hasn't even had a price released because it's a datacenter compute card and if you have to ask then you can't afford. And as far as I know, part of the 7nm shrink incorporates all of the new datacenter compute features and 4 stacks of HBM. There is no 7nm equivalent of Vega 10 to produce a gaming card with, so yes any use of Vega 20 for gaming would be ridiculously expensive. Everybody cried foul when AMD asked 1080 Ti money for a 1080 performing card, you think it would be any different if AMD asked 2080 Ti money for a 1080 Ti performing card? Because that's what Vega 20 would necessitate.

But of course it seems that even AMD fans expect AND to sell at cheap prices even when they do have a good product.

Oh, so you're now reducing this to fanboiism? Who said anything about me being an AMD fan? Or how about a little common sense in the market? Nvidia have been dragged over the coals as of late for their exorbitant RTX prices. The 2080 Ti for instance is an exceptional piece of hardware, but it doesn't validate its £1,200+ price tag. A Vega 20 gaming card could be a great product too, but the only price point that would make any sense in the market would require AMD selling at a loss.

 

[bru]

Um no, Vega 20 hasn't even had a price released because it's a datacenter compute card and if you have to ask then you can't afford. And as far as I know, part of the 7nm shrink incorporates all of the new datacenter compute features and 4 stacks of HBM. There is no 7nm equivalent of Vega 10 to produce a gaming card with, so yes any use of Vega 20 for gaming would be ridiculously expensive. Everybody cried foul when AMD asked 1080 Ti money for a 1080 performing card, you think it would be any different if AMD asked 2080 Ti money for a 1080 Ti performing card? Because that's what Vega 20 would necessitate.



Oh, so you're now reducing this to fanboiism? Who said anything about me being an AMD fan? Or how about a little common sense in the market? Nvidia have been dragged over the coals as of late for their exorbitant RTX prices. The 2080 Ti for instance is an exceptional piece of hardware, but it doesn't validate its £1,200+ price tag. A Vega 20 gaming card could be a great product too, but the only price point that would make any sense in the market would require AMD selling at a loss.

So you really think that AMD have produced a chip with built in HBM that can only work and be sold as a very expensive compute card. Well good luck with AMD ever getting back the development costs if that really was the case.

 

[LePhuronn]

So you really think that AMD have produced a chip with built in HBM that can only work and be sold as a very expensive compute card. Well good luck with AMD ever getting back the development costs if that really was the case.

Of course, selling compute cards at multiple thousands of dollars each that are based on an existing, yet refined, technology won't ever recoup costs...

OK, so what leads you to believe Vega 20 can be chopped down then? If you take out all of the new compute features you're just back to where you started with a Vega 10, albeit on a 7nm node. The only gaming benchmark we've seen with Vega 20 was that dubious Final Fantasy one and that showed zero gaming improvement over RX Vega 64. Polaris 20 to Polaris 30 strikes me as a small and easy investment to get a notable return. I really doubt investing time and money to get a 7nm gaming Vega would be cheaper than just maintaining existing 14nm production, especially if there is zero performance gain.

But who knows. That rumoured Vega 12 hasn't showed up yet. Maybe it is just Vega 10 without any changes shrunk to 7nm, or a Vega 20 that didn't have the new compute bits added on. That would certainly free up more 14nm capacity for Zen 2.

Oh hang on...

 

[melmac]

So you really think that AMD have produced a chip with built in HBM that can only work and be sold as a very expensive compute card. Well good luck with AMD ever getting back the development costs if that really was the case.

You know Vega 20 isn't a new process right? It's a node shrink but not a new architecture. It's the same as Vega 10 with improvements for the professional market, larger memory, better DP performance etc. What development costs would there be that wouldn't have already been taken care of by the Vega sales during the mining boom?

 

[bru]

Of course, selling compute cards at multiple thousands of dollars each that are based on an existing, yet refined, technology won't ever recoup costs....

You know Vega 20 isn't a new process right?

You know as well as the rest of us that the MI60 is the worlds first 7nm GPU, quite how that can equate to an existing technology or not be a new process, I'm not sure.


As for will they use it in a gaming card, we will have to wait and see, although I cannot think of a single chip that they have ever produced that they only used for the professional arena, so please if they have done that before do let us know.

 

[melmac]

You know as well as the rest of us that the MI60 is the worlds first 7nm GPU, quite how that can equate to an existing technology or not be a new process, I'm not sure.


As for will they use it in a gaming card, we will have to wait and see, although I cannot think of a single chip that they have ever produced that they only used for the professional arena, so please if they have done that before do let us know.

Why selectively misquote me?

You know Vega 20 isn't a new process right? It's a node shrink but not a new architecture.

I don't know if there is going to be a 7nm Vega gaming card, personally I doubt it, but, the reasons you are giving are questionable. Development costs? It isn't a new Architecture, they just refined it for the Professional market. None of the additions make a difference for the gaming market. You would be looking at a card that's, what? 20% maybe? faster than the Vega 64. And it would be a lot more expensive. If they strip out the additions to make it cheaper, then what are you left with? A more power efficient but also more expensive Vega 64?

Two things, AMD stated themselves they wouldn't be releasing any high end cards until their new arch in 2020 and Navi is been released for the mid range, so where exactly would a Vega 7nm gaming card fit in? Second, AMD would have learned a lot from the 7nm Vega node shrink. That knowledge will surely help with Navi, the Navi cards that are coming next year. I would suggest that AMD have enough on their plate with Zen 2 and Navi. Not sure they would waste their production time on 7nm with another version of Vega for gaming, especially with no mining to save them.

 

[bru]

Why selectively misquote me?

Err I didn't, you asked a question as shown by the question mark.

I mealy pointed out that being the world's first 7nm card, it is indeed a new process.

 

[D.P.]

So you really think that AMD have produced a chip with built in HBM that can only work and be sold as a very expensive compute card. Well good luck with AMD ever getting back the development costs if that really was the case.

Yes, that is exactly what they have done.

Nvidia Developed Volt;a that si basically only ever sold for compute (Titan V sales are completely irrelevant).


These kidns of cards sell for $6-10,000 a time, and liekly customers by 50-2000 at a time.

 

[LePhuronn]

You know as well as the rest of us that the MI60 is the worlds first 7nm GPU, quite how that can equate to an existing technology or not be a new process, I'm not sure.

So is Polaris 30 new technology then? By your logic it must be because it's built on a new process to its predecessor. But of course we all know Polaris 30 is just Polaris 20 constructed on GloFo's tweaked 12nm node; there is nothing new on it. Vega 20 is the same: it's Vega 10 with extra stuff bolted on and constructed on a 7nm node. @melmac must've mistyped because 7nm is a new process, but there is no new technology in Vega 20, just more of the existing stuff.

It's not outside the realms of possibility that Vega 20 didn't actually cost AMD that much to develop. Granted we're not talking £50 and a pack of cigarettes cheap, but Vega 20 has happened at the same time as Zen 2 and Navi, which is where the actual investment is being made; 7nm design for Zen 2 and Navi pays the lion's share of the R&D cost, and Vega 20 rides the coat tails. So with data center compute cards selling at least 4 figures a pop, and usually in big batches too, it's very probably any Vega 20-specific development costs can be recouped without ever touching the gaming market.