"AMD may be headed back to the well with Polaris, this time launching the new GPU on a 12nm version of GlobalFoundries 14nm process node. There’s been rumor of this coming, though there have also been questions about whether AMD would attempt a 12nm refresh given the reality of Nvidia’s Turing launch.
According to Phoronix, a new PCI ID, 0x6FDF, has been added to Linux. It’s not mentioned in any database of Radeon drivers or GPU models anywhere . It’s listed as being part of the Polaris 10 family — in Linux, Polaris 20 GPUs (RX 500 series) are part of the Polaris 10 family as well, which increases the chance that this is a 12nm die shrink. What might we expect from such a part?
We only have one data point for comparison for this shift: the transition from Ryzen 7 1800X on 14nm to Ryzen 7 2700X on GlobalFoundries 12nm (optimized 14nm). In this case, non-gaming benchmarks are actually a better comparison point than gaming tests, since they highlight CPU performance in areas where the GPU isn’t a bottleneck and we want some perspective on what the practical improvements of moving from 14nm to 12nm were for AMD. Our slideshow from the Ryzen 7 2700X review is embedded below if you care to refer back to it.
The Ryzen 7 2700X is ~11 percent faster than the Ryzen 7 1800X in the majority of tests, though some isolated benchmarks may show larger results. To look at how much this would matter, we’ve also included our slideshow from the original RX 580 launch. Our comparison point is the Nvidia GTX 1060.
A straightforward 10 percent clock increase would put AMD’s RX 580 ahead of the GTX 1060 in every test we ran at launch. Power would still be a significant advantage for Nvidia unless AMD managed to rearchitect the chip along with the die shrink, but AMD would have a stronger position to challenge Team Green in the midrange.
The idea that AMD would pull a refresh like this seemed unlikely back when Nvidia was thought to be launching a top-to-bottom stack refresh. But today, that doesn’t look to be the case. The performance hit from enabling features like RTX is so heavy, it’s not clear if Nvidia can even launch the feature below the RTX 2070, which means Turing may effectively be a refreshed GPU family with a lot of Pascal cards hanging on. Nothing is stopping NV from eventually launching new “2060” cards that are basically rebadged 1060s, and both AMD and Nvidia have gone down this road before.
But Nvidia may not want to take that step this time around. The question for the company is whether it makes more sense to position the new RTX family as entirely devoted to ray tracing and DLSS with the old Pascal numbering system reserved for non-RTX cards or if bringing these other GPUs forward into the Turing product line and labeling them with the new naming convention would confuse buyers. I’m inclined to think it would; buyers are unlikely to grasp the distinction between an “RTX 2070” and a hypothetical “GTX 2060” if Nvidia tries to make the “G” alone the distinguishing factor for whether new features are available. Regardless of whether or not Nvidia keeps Pascal in market or attempts to re-brand older cards as part of the Turing family, AMD may have an opportunity to better position itself in the mid-market.
Of course, all of this is supposition, based on some PCI ID codes for a product family, not actual hardware. But if AMD can snag a win across the midrange for a die-shrunk version of its Polaris family, the company may well go for it. Improving its competitive midrange position is a good way for AMD to win back market share, even if it isn’t competing with the RTX family.
This is why the Mainstream Tech Media just frustrates me, as they're constantly grasping at straws.
If they'd bothered to actually pay attention., they'd have noticed that 0x6FDF appeared at the same time the Radeon Pro WX 5100 (Polaris 28CU) Workstation Card was added to the Radeon Pro Family back in August.
Given the ONLY reason for an additional Device ID is to support a NEW Device Configuration., such-as a different Compute Cluster Structure.
GCN (as a keynote) is designed to operate with (up to) 4 Compute Units operating as a Shared Cluster … however there is no Hardware Component designed to either verify the Cluster Configuration, in-fact most of the Polaris Series are "Odd" Cluster Configurations., albeit this is strictly limited by the Microcode.
And this isn't just set by "Good" and "Bad" CU, but as all are *required* to operate at an Identical Frequency to maintain Data Synchronisation to the Cache, any that are unable to maintain within the Cache Timing Table have to be disabled; otherwise they can cause in a Best Case additional Cycles Per Instruction and Worst Case R/W Conflicts that result in an Architecture Hard Fault; or to Determine the Total Cluster Count.
You can of course run a Challenge and Response on Compute Units until you receive "No Response", meaning there is no Compute Unit there (Disabled Compute Units will report back Condition Flags, one of which is a Disabled., so the Hardware Scheduler knows if it can dispatch Instructions) but doing this interrupts Data Flow … again would increase the Cycles Per Instruction.
As such it's a better (and easier) approach to handle this via Pre-Programmed Profiles within the Driver that are assigned to the Device ID., as this should always have the same configuration.
This as a note is why the RX 470 and RX 570 have the same Device ID to each other, but this Device ID is difference from the RX 480 and 580.
For all intended purposes these are the same GPU., however instead of 4 Device IDs; we only have 2 Device IDs... and even if these were to be on 12nm., if functionally they're the same GPU, then they'd retain the same Device ID.
It's only if their Computer (Core) Configuration Changed, they'd receive a new Device ID... such-as 6FDF being the 28CU Variant.
Now we can of course question why there hasn't been either a 24CU or 28CU Variant during the Lifecycle of the RX 400-Series or RX 500-Series., but this is likely a simple case of Volume. All of them produced were almost certainly being purchased by Sony for the PlayStation 4 Pro.
It's likely a simple case that Sony now have enough (or what they consider enough) Stock to cover 2019., and as such are not ramping down their orders / production... there's also a possibility that the PlayStation4 Pro isn't selling as well as they'd hoped, given the main reason for getting one is PlayStation VR, which again has grossly undersold what Sony were hoping.
The Standard PlayStation 4 however continues to sell quite well.
It wouldn't make much sense to sell the "Puma" 8-Core CPU with 28CU Polaris as a limited run A12 9900., as well no one would buy it and they likely couldn't make a profit on it... plus it would just be bad publicity as people would complain "Why isn't this with the Ryzen Architecture?!"
Yet, disabling the CPU Component and just selling it as a Workstation Card., which already comes at a notable premium... well that makes sense to get rid of excess stock.
Actually would be interesting to see if that's true., but they're £410 … so eh., I'm not willing to drop that just to take it apart and see if it's basically the PlayStation 4 Pro APU. Still if you look at the Power Consumption., that it's using the RX 550/560 Reference Design instead of the WX Reference Design., well it looks like a last-minute "Hodge-Podge" rather than an intended addition to the WX Line-Up.
The same is true in regards to it being release in Q3 2018... keep in mind the last "New" Polaris was released in June 2017.
Does any of this point to a "Polaris 30" Series being produced at Global Foundries? Erm… No.
If we look at this as a whole, considering that the WX5100 was available to purchase (i.e. in Retail Stores, like Overclockers UK) almost the same day that the Linux Drivers were updated to support it's Device ID., well what make anyone thing we'll know anything (officially) about the Next-Generation Radeon until basically they're available to Pre-Order or Purchase?
I mean the same is true for the Athlon 200-Series., that essentially appeared out of bloody nowhere.
No one was expecting it... it hadn't been rumoured... no leaks... it was essentially being sent out to Reviewers within days of the news breaking.
The same is true for the Ryzen with Graphics GE Versions., HE Versions., as well as the Ryzen 2nd Gen Line Up; the AMD website itself at present appears to be the BEST place for "Leaks" as they appear to be updating it with Unavailable / Unreleased Products a few weeks Reviews / Launch / Announcement / etc.
Honestly I don't think (regardless of how "Big" a deal the RX "600" Series is) we're going to hear much about it until it's basically available to buy.