Clues Gather Regarding Possible New AMD Polaris (Re)Revision Launch

[Raevenlord]

Clues have been popping here and there regarding a possible new Polaris revision being launched by AMD in the (relatively) near future. Speculation first reared its head regarding a revised "Polaris 30" silicon, allegedly being built for TSMC's 12 nm process - not unlike AMD's 2000-series Ryzen CPUs. The company has been enamored with trying out and adapting new foundry processes for its products as soon as possible, now that they've found themselves fabless and not having to directly support the R&D costs necessary for process node development themselves.

Some publications are pointing towards a 15% performance improvement being achieved on the back of this process change for Polaris - which, if achieved only via a new process implementation, would require clock speed increases that are higher than that. AMD has already launched their revised Polaris 20 RX 500 series, which built upon their RX 400 series (and Polaris 10) by upping the clocks as well. A smaller node would likely be associated with higher yields and decreased costs per finished chip, which would allow AMD to further reduce pricing/stabilize pricing while introducing a new product generation to tide users over until Navi is finally ready.





Adding to all of this (and the included NaCl), a post via Phoronix has been posted which speaks of a new Polaris Device ID (0x6FDF) that's being added to the latest AMDGPU Linux kernel patch. The new device ID is being added under the "POLARIS 10" family, which includes the Polaris 20 revision. We'll see how this pans out, but if AMD are to in fact revise their Polaris architecture for the 12 nm node, some architectural changes likely wouldn't go wrong to extract maximum value out of that investment.

View at TechPowerUp Main Site

 

[dj-electric]

If AMD can make a product similar to the RX580, with 10%+ of performance and shave a few dozen watts for 249$, that might get some success in the market today for those not wanting 2nd hand Pascal\Polaris for very attractive prices.

 

[Xzibit]

What are the chances of these just being OEM cards RX 500 series. Those get announced a few months before a new series launch.

 

[londiste]

A smaller node would likely be associated with higher yields and decreased costs per finished chip

Except TSMC 12nm is not a smaller node compared to the TSMC/GF 16/14nm.

 

[XiGMAKiD]

A Polaris-sized Vega or straight die shrink? I hope it's the former

 

[Assimilator]

A Polaris-sized Vega or straight die shrink? I hope it's the former

They wouldn't call it Polaris if it wasn't Polaris...

This is going to be another underwhelming die-shrink of Polaris to give AMD's GPU division some sort of method for competing with Pascal/Turing over the holiday season, by launching "new" cards with a larger first digit model number than their "old" cards. If they can drop the price low enough it should be sufficient to steal some of the GTX 1060's thunder.

 

[XiGMAKiD]

They wouldn't call it Polaris if it wasn't Polaris...

That makes sense, and will make the actual product less interesting unless priced very attractive let's say $200

 

[PanicLake]

That makes sense, and will make the actual product less interesting unless priced very attractive let's say $200

Why do people do always exaggerate what would be the reality? A more powerful new product that will cost way less than what it is replacing...

 

[techy1]

it is sad that the fact that RX 480 (from year 2016) node shrink might the most exciting news in GPU market 2018

 

[Fatalfury]

how can a mere 2nm node reduction (from 14 nm to 12nm) give an increase of 15% performance.
maybe 5-10% performance increase with AMD pushing the clock rates beyond the limits thereby having like 220W same as rtx 2080( because rx 580 = gtx 1080 in efficiency) power consumption for card only
thereby making it DOA...

 

[First Strike]

Radeon Graphics: Inter than Intel

 

[ppn]

Why can't AMD just put 3072 processors and GDDr6 in future product soon.

 

[efikkan]

Even if true, this could very well be OEM models.

Except TSMC 12nm is not a smaller node compared to the TSMC/GF 16/14nm.

When will the staff get this?

how can a mere 2nm node reduction (from 14 nm to 12nm) give an increase of 15% performance.

maybe 5-10% performance increase with AMD pushing the clock rates beyond the limits thereby having like 220W same as rtx 2080( because rx 580 = gtx 1080 in efficiency) power consumption for card only

thereby making it DOA...

It's not a node shrink. It's just a refinement, comparable to what Intel did with their 14nm, 14nm+ and 14nm++. We should expect marginally better thermals and yields.

I'm not sure we can expect 10-15% improvement from this alone.

And BTW, GTX 1080 is about 80% more efficient than RX 580.

Why can't AMD just put 3072 processors and GDDr6 in future product soon.

Because that would require rebalancing of the whole GPU. Taking a Vega10 and scaling it down would probably work much better than scaling Polaris up. Still, the greater problem for AMD isn't lack of theoretical performance, it's lack of utilization. RX 580 already have 50% more GFlop/s and 18% memory bandwidth than GTX 1060 which beats it, so there isn't a lack of resources. Adding more cores without also improving the scheduling will result in reduced efficiency for very marginal performance gains.

 

[xkm1948]

Better than nothing

 

[Assimilator]

Taking a Vega10 and scaling it down would probably work much better than scaling Polaris up.

Good luck with that. Remember, the real reason Fury and Vega were coupled to HBM is that its power consumption is so much lower than GDDR, which AMD needed to offset the absolutely terrible power consumption of their own chips so that the overall package power usage was merely terrible, not horrific.

There's also the more pertinent fact that AMD has finally realised that ploughing money into the dead end that is GCN is a waste, so there's no possibility of a Vega redesign... maybe a respin, maybe maybe Vega 20 consumer parts at best.

 

[efikkan]

Good luck with that. Remember, the real reason Fury and Vega were coupled to HBM is that its power consumption is so much lower than GDDR, which AMD needed to offset the absolutely terrible power consumption of their own chips so that the overall package power usage was merely terrible, not horrific.

Not really, a smaller Vega die, bundled with 256-bit GDDR5 (presumably) would not be a major issue, that's what Polaris have after all.

There's also the more pertinent fact that AMD has finally realised that ploughing money into the dead end that is GCN is a waste, so there's no possibility of a Vega redesign... maybe a respin, maybe maybe Vega 20 consumer parts at best.

Well, early roadmaps did show a smaller vega chip. Many assumed this to feature up to 3072 cores. I assume this chip was scrapped due to worse scaling than Polaris, since pushing the clocks on Polaris might have scaled better than presumably unbalanced resources on the smaller Vega chip. I don't see many other reasons to scrap the new design and keep refreshing the old after the new one is made.

 

[lexluthermiester]

My 2 cents; AMD sees RTRT as the next big thing that it really is. They will very likely have an answer to RTX.

 

[T4C Fantasy]

this Device ID is very very old, dates back to 2015.

and is now picking up steam in 2017/18

https://www.techpowerup.com/gpuz/details/aemqn

 

[eidairaman1]

My 2 cents; AMD sees RTRT as the next big thing that it really is. They will very likely have an answer to RTX.

My 2 cents; AMD sees RTRT as the next big thing that it really is. They will very likely have an answer to RTX.

They already have way before nvidia started cowtouting to game studios.

 

[Casecutter]

I considered this a while back as the best low cost/engineering play seeing the pull-back AMD has done toward consumer graphics. I see it strange that this is referencing TSMC's 12 nm process. I'm sure GloFo isn't telling AMD to take-a-hike, they can't loose Polaris business. One would figure GloFo is looking to consolidate around their 12nm, and more than willing to work with AMD. AMD would be smart just to "stay the course" at GloFo and re-spin Polaris 20 with some refinements and a new GDDR6 memory controller. As holding to GDDR5 is probably going to see waning inventories, and then not great pricing. Given the strong uptick in bandwidth while getting some return on lower energy it's time in so many ways. Sure this node shrink is more about a TDP improvement vs. increase clock, or balance of both. AMD could use the up-tick in memory/power savings to balance out at the existing "power envelope" and not see big push back from gamers.

Deliver a full "Polaris 30" that's a 1440p darling for all the 27-32" FreeSync monitors and price it the same MSRP at $230. For the 570 replacement make the 4Gb (GDDR6) as the "end-all-be-all 1080p card" most gamers need (kind of what a 7770 did for anything less than 1080p) and promote how it runs most anything silky-smooth with a FreeSync 1080p monitor and price it at $160.

There's a ton of gamers that have been making due with cards that they bought after the first mining bust of 2014-15 and have of yet seen prices/performance be good enough to take the leap, especially given both sides are still hawking tech that's now approaching 2-1/2 years.

 

[ppn]

12nm is 16++ node. 7nm or skip.

 

[DeathtoGnomes]

so we are prolly looking at RX585/590/595

 

[Mistral]

I already have a 480 and a 580, might as well get this too and collect them all... Seriously, if they put it around $200 to $250 it'll sell.

 

[robot zombie]

Nobody really knows quite what AMD is gonna do right now. I think the price/performance figures people are tossing out right now are optimistic. It just wouldn't make sense for them to give their cards away while nvidia charges so much.

I find the general silence telling, though. Probably smart to play the waiting game right now. I'm not convinced they're even looking to compete right now. I mean, no matter what they were to build people up to right now, it wouldn't be enough. I'm betting they're not hyping things up because they know they have no answer to RTX at the moment, though no doubt we will know when they do. Wouldn't be surprised if they dropped something with no fanfare between then and now, as they have done before.

I dunno, not thinking anything of them one way or another right now.

 

[Casecutter]

It just wouldn't make sense for them to give their cards away while nvidia charges so much.

Their after the "mid-range enthusiast" needing hold market share that creep'd-up under mining. Plus if they can sneak in before Christmas and Nvidia doesn't have something, lots of those X-mas dollars kids get will come to them. While AMD continues to take CPU sales I think lot of those builds need/want Radeon GPU's that provide full potential of pixel pushing to both existing monitors, while upping the desire to some nice FreeSync panel that are both plentiful and pricing better.