Okay then. Bit of a wandering post here, basically collecting together bits and pieces of theories I have on exactly what AMD’s 2018 and 2019 plans are. Sony is getting mixed up in here because, from a product standpoint, AMD is tightly tied to Sony on what is known as AMD’s next generation Navi part. So I’m going to try and speculate on what I think AMD and Sony are up to.
Now, I have been over much of this already on my Twitter Account and on a post to Massdrop’s, well, community forum and blog system. So if you’ve been following me none of these speculations should be much of a surprise.
As a specific note: I have not been in any contact with anybody from AMD or Sony in any official capacity and this editorial references events that have transpired between the original posts to Twitter and Massdrop. I have not yet been contacted by AMD or any AMD hardware partner to review or have hands on time with the the incoming products. This should be considered, as of this posting, as speculation on my part.
So, let’s work with what with what we think we know. We think we know that the work on Navi for Sony PS5 was one of the reasons AMD’s Vega Architecture was a flop. We think we know that Navi will be a monolithic design, but it will also be scalable but not in terms of Vulkan MultiGPU support explicitly. We think we know that Navi, if configured to be used in a card with an MSRP of ~$250(US), is capable of outperforming Nvidia’s GTX 1080 in DX12 and Vulkan rendering scenarios. We think we know that there will be a 7nm part from AMD shipping this year. Forbes corrorabates a 7nm part before 2018 is out. We do know that Dr. Lisa Su will be unveiling quote/unquote High Performance GPU and CPU designs on 7nm during the CES2019 keynote.
We also know that out of, well, nowhere, AMD added a Polaris30 chip ID to it’s Linux drivers.
We also know that Vega 56 Prices are being pushed down to the $300 mark.
We also know that a RX 590 branded gpu has been leaked in benchmark tests.
So here’s what I think is going on.
Polaris30 :: aka RX 590 / RX 595
I think AMD’s Polaris30 is not a 12nm chip as potentially reported by other sites. I think that new Polaris 12nm chip design is the Radeon RX 580 2048SP that is being exclusively sold in China.
I think Polaris30, this new RX 590, is a 7nm chip.
Now: admittedly; later on in this editorial I am about to add significant caveats to the expectation of the RX 590 as a 7nm release only. So just keep the RX 595 nomenclature that I referred to on Twitter in mind as you read on.
I think the Polaris30 was AMD’s first pass at working with 7nm fabbing leveraging a simpler architecture than the significantly more complex Navi or Vega archicturers. There is also ample precedent for AMD running die-shrinks on the simpler mid-range architecture to dial in their internal tooling. Just a quick glance over on TechARP and we see examples such as the RadeonHD 2×00 line where the Mid-Range was on 65nm versus the High-Range on 80nm. The legendary RadeonHD 4770 clocked in 40nm while the 48xx+ sets were still on 55nm.
So this would not be an unusual move on AMD’s part to fab out an older or less complex graphics architecture on a smaller die-size before running the next high-end performance part on that die-size.
I also don’t think the that the Polaris30 is just a refresh of the Polaris10 as used in the RX 580.
I think that Nvidia’s failure to even mention a RTX 2060 GPU, or to cut prices on the GTX 1070/1080 products, or to do anything with the GTX 1060 besides offer some expanded memory options has opened up a huge window of opportunity for AMD. Right now there just Is No Mid-Range Graphics card.
Nothing in the Mid Range
Let’s recap what we think we know on this subject. Polaris10 was a huge energy hog with the RX 580 carrying a Thermal Envelope of 185 Watts. That’s higher than the stock GTX 1080 with… well… the practical performance less than that of the 120 Watt GTX 1060. The GTX 1060 and RX 580 are both relatively capped to 1080p screenspaces. The GTX 1060 generally handles QHD resolutions a bit better… but runs into caps elsewhere in the hardware. So anybody looking to spend between $200(US) to $300(US) on a graphics card can finally get one now that the Bitcoin Mining Craze has crashed… but they’re stuck with either artificially capped hardware… or well… The RX 580.
With the rise of affordable QHD-class freesync capable monitors, such as Massdrop’s Vast Monitor, the lack of a mid-range card that can tackle those resolutions is huge opportunity for AMD because well, there is a precedent here too. The Legendary R9 290x. When AMD and their partners dropped the asking price to ~$300 the benchmark was set. That was the performance/price that Vega was supposed to hit. A $300 card that could outperform the R9 290x in DirectX 11, OpenGL 4, DirectX 12, and Vulkan rendering scenarios.
The R9 290x’s legacy was established by it’s 64 Raster Operation Count… which allowed the card to be one of the first, if not the first, graphics cards that could realistically render into 4k/UHD resolutions… and by AMD’s willingness to break all the rules. The R9 290x was almost as obnoxiously loud as the Nvidia GeforceFX series on a number of dual-fan coolers at full load, and would run into thermal temperatures well beyond anything seen before… but it was still trading blows with much more expensive graphics cards from the Geforce family as resolutions increased… and I think AMD’s going to leverage that same strategy on Polaris30.
I think AMD is going to hit Nvidia with a mid-range broadside by unleashing a superclocked Polaris chip with a doubled ROP count from the RX 580. On a 14nm to 7nm die-shrink, and based on known air-cooled stable overclocks to ~1.5ghz for the RX 580, I’m expecting Polaris30 / RX 590 / RX 595 to potentially clock in with a ~1.8ghz maximum speed with 64 ROPS.
Now, both on Twitter and on Massdrop I tried to make a point of why I believed the rumors about Polaris30 being a 7nm version with a 64ROP count. I think Dr. Lisa Su is done with half-measures inside of AMD. It’s either knock the ball out of the park; or get out of the company. One of the big failings with the Polaris architecture from my perspective is that it was too limited from the start. It was always hobbled because any better performance and it would be stepping on the toes of Raja’s pet Vega Architecture.
The Polaris design team then has something to prove with a Polaris30 product release. They are out to prove that they had a fundamentally good mid-range chip design on their hands and that Raja’s ego is what ultimately killed Polaris as a performance competitor to the GTX 1060. With Dr. Lisa Su wanting Home Runs and not infield-pop-flys; it almost seems too simple to envision her signing off on letting the Polaris team make Polaris as good as it could be while getting in time on 7nm production.
My suspicions seem to be confirmed by the unveiling of benchmarks of a RX 590 labeled GPU in Final Fantasy XV. Notably a RX 590 sample punches out the GTX 1060 in 4k rendering resolutions; and that’s keeping in mind that the Windows FFXV client is heavily invested in Nvidia’s software optimizations infrastructure. While there have not been any screenshots or raw-frame-video-footage releases to date showing that there was an equivalent image quality between the test runs; we can only assume that either in the absence of, or even with the benefit of, cheats in FFXV’s rendering that favor Nvidia’s hardware… the RX 590 brings Polaris out of the 1080p screenspace limitations.
I’m also not of the opinion that this performance is entirely reached on clock speed gains alone. We know from RX 580 overclocks that the Polaris10 architecture routinely ran into screenspace rendering caps long before it ran into graphical detail or compute caps. Case in point would be the mining craze where overclocked RX 580’s showed much better percentages of diminished returns on compute values compared to the falloff of screenspace performance. I’m also of the opinion… that if AMD was going to be sneaky about the RX 590 launch… they’d probably only issue out pre-release versions with clocks close to that of the RX 580. In theory the massive gains in performance could just be down to fixing the artificial caps Raja’s Crew put into Polaris to keep it from competing against Vega.
So, let’s be clear what the RX 590 is not. This is not going to be a High Performance Part. This is not going to be a Power Efficient Part. This is not likely going to be the $300 R9 290x Killer that gamers have been waiting for. It may very well trade blows with the R9 290x and win a few of those fights… but this won’t be a Doors Blown Off The Van – You Had ONE JOB! scenario. Or at least I don’t think it will be one.
What the RX 590 could be is a part that successfully occupies the performance space of Vega 56… at a $300 price point. While it might not be able to handle pushing most games into 4k resolutions with appreciative levels of detail the RX 590 should handle UWHD and UWQHD 21:9 aspects in Vulkan rendering scenarios without a problem and perhaps in potentially better graphical fidelity than the GTX 1060.
What RX 590 is… or looks like it could be… is an actual $300 Mid-Range part that will be available to actually purchase in stores without leaving gamers feeling like they got screwed. That alone could captivate the gamer’s wallet during the remainder of 2018 and thus buy time for AMD to get to Navi.
Navi : PS5
So… Navi. The GPU that ties AMD to Sony. I think I may be able to sum up what I think I know in one sentence:
Navi’s Purpose is to Deliver the same level of Graphical Fidelity that a PS4Pro running in a 1080p screenspace can deliver running on the bare-metal Vulkan API into a 4k/UHD screenspace, while maintaining a similar total power consumption profile to the PS4Pro, while also maintaining an entire system physical hardware cost profitable at $500(US), while maintaining ABI/API compatibility with PS4Pro.
(edited to try and clean up the wording)
I’m not sure if this is the most accurate description of what I think Navi is, but I think it is one that best fits the rumors. A very important catch in the qualifications here is that Navi isn’t just about the hardware package itself. Navi is as much about the software packages that AMD and Sony provide end-developers.
We kind of saw this focus on software packages at the launch of the Playstation 4 when then Sony owned Sony Online Entertainment was supposed to provide the Forgelight Game Engine as the default internal PS4-optimized game engine. That never happened and we witnessed Sony hurriedly strike up deals with other game engines to fill the void left by SOE’s managerial incompetence.
Going into PS5, as I understand the outlined scenarios, Sony is much more self-conscious about making sure developers have functional dev kits in their entirety on day 1; and that there won’t be a focus on backing on specific engine or another. Maintaining ABI/API compatibility with the PS4Pro should make it easier for developers to hit the ground running with optimized games.
The point of maintaining ABI/API compatibility also links into other rumors that have been circling around. We know that in 2016 that Kaz Hirai saw Pokemon Go as a game changer. We also know that Kenichiro Yoshida has an interest in 5g mobile deployment and on streaming technologies. Need I say more than Playstation Now. We also know that various executives within Sony have muted that a return to Mobile Gaming was probably in Sony’s future.
I think the reason that the rumors and comments about Navi’s portability and scaling have been all over the place is that I think AMD has made three different Navi models. I think those models are as follows:
- Navi+Zen Base: PS4 1080 fidelity (mobile product?)
- Navi+Zen Mid: PS4 Pro 4k checkerboard fidelity
- Navi+Zen Full: PS4Pro 1080 fidelity @ 4k freesync fidelity
While perhaps a gross oversimplification of what AMD and Sony have been up to; if we consider that there may be up to three different Navi chips in play for Sony the specific comments about scalability come into focus. What if Sony wasn’t just looking at one console SKU for the PS5 launch? What if Sony was getting back into mobile gaming? What if Sony wanted to keep PS4 owners happy and continue to release PS5 titles on PS4? What if Sony had a consistent ABI/API across a new mobile gaming device that is based on the launch PS4’s ABI/API capabilities?
What if Nintendo was aware that Sony and AMD had prototyped such a mobile device and let slip that there was a new Nintendo Switch in the works?
If we accept this range of potential performance levels… I think this may also put into perspective specific comments from developers that Navi was capable of generating completely new visual environments. Quite simply… I don’t think we’ve actually seen developers push the PS4 Pro to it’s absolute limits in 1080p alone. While we’ve seen a number of developers offer additional graphical improvements while running 1080p on a PS4 Pro… there is a vast difference between improving the performance and optimizing for that performance. That only leaves a version of Navi+Zen offering the launch PS4’s performance as an outlier… but if that base model only had to worry about a 720p HDR or 1080p HDR panel… it’d put itself right in line for the visual fidelity that third party developers working on a competitor’s system would be optimizing for. As shown on the Nintendo Switch mobile gamers will happily sacrifice a bit of visual quality to take their game on the go.
So how does AMD’s work with Navi influence the PC release schedule? If Navi silicon is ready now, why wait for 2019? Why even release a Polaris RX 590 stopgap?
Let’s talk about actual chip construction for a moment:
Taiwan Semiconductor Manufacturing Company : Capacity
Simply put: AMD and Sony have a big problem with their 7nm chip plans. That problem is the limited capacity of TSMC’s production lines. Sony wants a significant number of Playstation 5 consoles in stock and ready to ship on launch day. AMD wants a significant number of 7nm Zen and Rome processors in channel to leverage Intel’s manufacturing shortfalls. AMD also wants a significant number of 7nm GPU’s in channel to meet GPU demand after multiple years of gamers getting screwed by bitcoin miners.
TSMC only has a certain amount of manufacturing capacity on it’s 7nm production line. I think this is why most people assume that RX 590 will be a 12nm chip… and it could very well actually be one by the time this editorial posts because Apple has laid down a significant fiscal outlay to secure TSMC’s 7nm capacity for a 2019 Apple Iphone within the last month.
So I’m going to make a bold prediction for AMD’s New Horizon Event. I think one of the subjects could be the joint licensing of TSMC’s 7nm process to Globalfoundries or Samsung Semiconductor. GloFo and Samsung have signed such a deal between themselves before; so there is precedent for such a licensing deal when fabs get backed into corners on orders they can’t fill. However, I’m more inclined to believe an even more earth shaking news bite. I think that AMD, Sony, and other hardware vendors who contract with TSMC are looking to build a new 7nm fabrication facility… that isn’t in Taiwan. Since I’m laying this guess out there; I think TSMC could be about to make a massive upgrade to the WaferTech facility in Camas, Washington.
If we assume a 6month to 8month facility upgrade time… we now have a pretty solid reason for AMD to delay Navi until 2019 when TSMC will be better positioned to meet 7nm demand. Incidentally, delving into modern politics, opening a 7nm chip facility within the US would certainly help TSMC land a few contracts that it might not otherwise be able to bid for. This also may simplify or eliminate potential tariffs that TSMC, Sony, AMD, and AMD’s PC GPU board partners could encounter.
So what about the RX 590? Could AMD have secured enough manufacturing capacity on TSMC’s existing 7nm production lines to actually have inventory on store shelves?
Yes, I think so. Again, the Polaris design is not really that large or that complicated in terms of modern fabrication technologies. A smaller GPU size means more GPU’s per wafer, and AMD and TSMC probably have a very high yield percentage on that larger number of GPU’s per wafer. Again, AMD’s GPU teams have done this kind of die-shrink and mass-production on previous cards. It’s nothing new.
Now… is that actually what AMD is going to do? Or will AMD play it safe and pump out RX 590 on a 12nm process which can be obtained from a variety of fabrication facilities?
On Twitter I intentionally gave the name of the potential Polaris30 product as an RX595. Part of my reason for using that nomenclature is that while I’m fairly confident that 7nm Polaris30 exists… in volume for shipping… I also am aware of how hungry the gaming market is for that quote/unquote Legendary R9 290x Killer. Vega didn’t deliver… Nvidia didn’t deliver… and AMD’s history strongly suggests that many employees will accept an inside-the-park-double over an actual home-run.
My guess… as this article actually goes to post… rather than a month or so ahead of AMD’s Next Horizon November 6th briefing is that if AMD is stuck on the supply side… the RX 590 probably will be a 12nm part and we could still see a 7nm part as an RX 595.
Vega : AMD Arcturus
One of the claims I made over on Massdrop is that gamers should probably expect the Vega20 series to be dead as an actual standalone product. While we will probably still see the Vega20 architecture show up in CPU designs that are already being taped out… I don’t think AMD is going to be announcing Vega20 @ 7nm when CES2019 hits.
The reason for that is the recently leaked AMD Arcturus. Not much is known about this chip other than it’s some kind of successor to Navi.
There is little argument that Navi is long overdue, and while it may not have been as much of a victim of Raja’s mismanagement as was Polaris, it still has some big shoes to fill. Everybody is expecting Navi to lay waste to Nvidia’s GTX 1080 at a $250 price point. Everybody is expecting Navi to do what Vega failed to do. While I realize this is probably driving the knife in deeper; Vega is a Flop.
The Vega architecture is, in general terms, not thermally efficient for it’s computational capability. Discrete Vega graphics cards relied on exotic and expensive technologies that had not matured into profitable parts. While parts built on Vega for Zen APUs certainly perform well compared to their competitors there are questions as to just how much of the Vega 8 and Vega 11 architectural disadvantages are disguised by the complications of an APU design.
So let me phrase this as a theoretical exercise.
If we once again assume that the design of the Navi GPU was built around delivering the graphical fidelity of the PS4Pro’s Polaris GPU in the Vulkan API with a 1080p screenspace into a 4k/UHD screenspace, while maintaining a similar power consumption profile to the Polaris GPU, while under a strict cost constraint of all of the GPU components…
Then Arcturus is likely a GPU designed to deliver the performance of the PS4Pro’s Polaris GPU @ 1080p into an 8k screenspace with a similar power consumption profile and no cost constraints on silicon or support
I think AMD hit their target. I think Arcturus is as much a successor to the Navi family as it is an outright replacement for Vega20. I think without Raja mismanaging the graphics division AMD already has an architecture that can realistically drive Vulkan rendering in 8k resolutions.
That being said, if the theoretical exercise is accurate… Arcturus products are not likely to be generally affordable. We’re probably talking a graphics card that could actually command the price of a GTX 2080… while delivering the performance you’d expect for making your wallet cry out in pain.
So, that’s what I have to speculate on the subject of where I think AMD and Sony are going, or could be going.
Okay, I have lots more to speculate, but that’s just speculation on top of speculation.