Rikimaru
Veteran
No they are advertizing flicker free monitors without PWM.
Predict: Next gen console tech (9th iteration and 10th iteration edition) [2014 - 2017]
- Thread starter Shifty Geezer
- Start date
- Status
How about this? @otobori @otobori @otoboricorrectIt's interesting because using @name doesn't actually notify me to come to this thread
lol no. But I feel like I'm missing a joke I should understand =(
@Beetlejuice @Beetlejuice @Beetlejuice
It's maybe a little off topic but I wanted to put it somewhere, AMD announced Fury XT GPU (releasing soon) is 8.6 teraflops. Busting out the calculator, that's 4.7X PS4 and a whopping 6.6X Xbox One, almost 7X.
http://www.fudzilla.com/news/graphics/38013-amd-announces-fiji-based-radeon-fury-lineup
Just that even a single PC GPU is already approaching the transformative order of magnitude more power than consoles (10X=probably what you need to see a major onscreen difference, and is also a typical generation leap for consoles). Happened very quickly. Although I suppose to properly reach it (on a single GPU) will take another year/generation.
To make it a bit on topic, seems Fury XT HBM is very power conscious. With flagship cards being very short. HBM should help 9th gen consoles in many ways (assuming we do get proper high end consoles again).
http://www.fudzilla.com/news/graphics/38013-amd-announces-fiji-based-radeon-fury-lineup
Just that even a single PC GPU is already approaching the transformative order of magnitude more power than consoles (10X=probably what you need to see a major onscreen difference, and is also a typical generation leap for consoles). Happened very quickly. Although I suppose to properly reach it (on a single GPU) will take another year/generation.
To make it a bit on topic, seems Fury XT HBM is very power conscious. With flagship cards being very short. HBM should help 9th gen consoles in many ways (assuming we do get proper high end consoles again).
Last edited:
AMD's Hawaii GPU was already at 5.6 TF before the PS4 launched, so the Fury GPU is actually a smaller step compared to the deficit that already existed at day 0.
It's a nice bit of kit, but the difference in what it takes to get there versus what is acceptable for a console is so vast that it's difficult to project when technology will advance enough to bring it out of the ionosphere. In that regard, Fury is even further away now than Hawaii was then.
It's a nice bit of kit, but the difference in what it takes to get there versus what is acceptable for a console is so vast that it's difficult to project when technology will advance enough to bring it out of the ionosphere. In that regard, Fury is even further away now than Hawaii was then.
Its pretty impressive that we have a 8.6gflop gpu on the same generation process as the 7970 at 3.8tflops all on the 28nm process.
The true sea change will be next year with sub 20nm process and hbm 2. I would predict that 5.6tflops will be on a low end gpu that sips power and we will be far beyond 8.6tflops. Perhaps in the 14tflop range for a single gpu
Taken from the Dreams thread about Media Molecule going fully compute instead of rasterizer.
How much APU space those fixed functions need? Because they could remove these and replace them with pure ALU hardware in future APUs, notably consoles APUs, right?
How much APU space those fixed functions need? Because they could remove these and replace them with pure ALU hardware in future APUs, notably consoles APUs, right?
Dreams is highly stylised. I don't think compute-only is a suitable alternative for rasterisation in all cases.
I know (and I expected this kind of response), but if you remove those fixed hardware functions in those APUs, you could put more compute and so maybe it would become suitable in most cases because you'd have more pure compute power which also will have others advantages for the others stuff to process.
So overall, it could be a win win, depending of how much space those fixed function need.
So overall, it could be a win win, depending of how much space those fixed function need.
The rasterizer itself is pretty small, so while there may be implications as to how the system is laid out, just removing it would probably provide no real opportunity for additional compute. It draws little power.
The texture pipeline is a complicated case, since it is part of the vector memory section of the CU that the compute shaders would use. A significant chunk of that area is going to be there regardless, and without it you would be devoting a significant fraction of ALU time to calculating limited stride lengths, format decompression, and low-precision math, which AMD has figured there are better things for those SIMDs to do. Even Larrabee found that its CPU cores could be better employed than that. Also uncertain at this point is whether there aren't some functions in that domain being used, since the two functions share hardware space and there are in some workloads compute uses for those blocks.
The ROPs are perhaps the one block most clearly not used if going pure compute. Their area cost is probably rather modest, although I've never been entirely sure which blocks are involved in that subsystem.
To follow through with this question, there are APU shots in the following:
http://www.extremetech.com/gaming/1...ered-reveals-sram-as-the-reason-for-small-gpu
I'll go with the the PS4 APU.
Everything to the left of the CU array is dominated by CPU modules, PHY, and northbridge and control hardware. Most of it is unlikely to change. A significant fraction of the hardware to the right of the array is special-purpose hardware, but may include the various DSPs and media blocks that are going to be there for non-graphics reasons.
The texture/memory units themselves are about 1/4 of the CUs, and are probably oriented along the midline. A significant chunk of that area simply has to be there, so perhaps it might shrink without going away. The rest of the CU array would not change much.
ROPs might be somewhere in the area above and below the CU array, perhaps some hardware is to the right. Much of that area is again dominated by PHY or other hardware.
Maybe you could add another ~4 CUs? (edit: maybe 6 if these were shrunken CUs with compromised texturing?)
Fully emulating what was lost could cost more in peak performance and power efficiency, and for all the software that doesn't use this method the design's utility would be compromised. The TDP ceiling would not change, and for a lot of this the general-purpose replacement for the same amount of peak specialized silicon is going to cost more in area and power.
I look forward to some of the details about the implementation of the distance-field algorithm. The other thread linked to a paper discussing the concept, and perhaps someone with more insight on how this might be implemented could comment on some of the phrases I noticed when it came to generating acceleration structures, with conservative collision testing and lists constructed after a coarse rasterization step. There are behaviors there that I have seen mooted in other functionality demonstrations, just not for GCN hardware. Perhaps if there were a bit more fixed function hardware, such as a future full DX12 GPU, a variation of the method could find some hardware assistance.
Last edited:
It could be that Fury level performance could be attainable already at Samsung/GF 14nm FF, say somewhere around the middle of 2016 (after having run in the process for over a year of volume production). 250-300mm2 die area, + dual stacks of HBM2. Assume the CPU part of the APU is negligeable in size comparatively. Cost and power draw should be acceptable for living room use and a $400 launch price.
However, it wouldn't bring the factor of ten or so level of improvement that was postulated above to be required to give an clear lift in capabilities (I broadly agree with that).
For that we would need to reach at least the 10nm lithographic node, and a factor of 10 increase in performance over the PS4 seems a stretch even there. Assuming TSMC and Samsung achieve their goal of starting volume production at the 10nm node at the end of 2016, we might see large size APUs somewhere around the middle/end of 2018. Lets speculate a factor of 6-8 performace relative PS4 at reasonable die sizes and power draws.
Trying to predict when it's possible to produce large dies at acceptable yield and at acceptable cost at 7nm is pretty much impossible. If you want a console APU at that node, you would have to accept not being able to predict when your product could make it to market.
If I were Sony, and wanted to play it safe, I'd target a 2019 release of a console made on a 10nm process. Of course, predicting that far into the future, other issues such as whether AMD can be trusted to be around as a supplier are valid concerns.
AMD Zen should be out sometime next year, and that could coincide with the launch or announcement of a new Nintendo system, likely a true successor to the Wii U. A quad-core Zen with anywhere from 640 to 1280 SPs is plausible. It could be equipped with HBM2 only (if costs fall quick enough) or an HBM2 + DDR3 approach to leverage the benefits of HBM, without the total cost a full commitment may require. I'm already expecting a mobile or desktop quad Zen w/ ~768 SPs w/ HBM capability next year, so AMD might have a product that is ready for Nintendo to exploit, with enough CPU power to emulate Wii U titles without issue, and enough GPU horsepower to make it worth porting games to from the Xbone and PS4. Nintendo has a history of using exotic memory, and HBM2 would be their ticket to high-bandwidth heaven, even with only 2 GB of it. You could easily pair it with 4 or 8 GB of external DDR3L.
Nintendo shouldn't play too conservatively, as they could actually take the lead and make themselves an attractive platform for developers, especially by adopting standard CPU and graphics architectures. That doesn't mean going 2000+ SPs, it means getting into Xbone or even PS4 territory, which is not that difficult.
Nintendo shouldn't play too conservatively, as they could actually take the lead and make themselves an attractive platform for developers, especially by adopting standard CPU and graphics architectures. That doesn't mean going 2000+ SPs, it means getting into Xbone or even PS4 territory, which is not that difficult.
Last edited:
Rikimaru
Veteran
Nintendo said nothing like this. On the contrary they said N could offer more hw forms if they have unified software SDK for various devices like Android.
Nintendo will not go for HBM in near future. Too costly, unproven tech.
Most likely they'll go 28nm ARM or AMD Jaguar with EDRAM as usual.
I figure expecting Nintendo to do anything high-tech at all is a non starter let alone all that. Let them prove me wrong but I'll believe it when I see it. Until then I dont expect NX to even remotely challenge Xbox One/PS4 let alone anything higher. Nintendo always disappoints to an amazing degree. Like, it's hard to imagine how low tech the Wii U was after many people honestly thought it was something quite powerful. Last best guess turned out to be 160 shaders...far far lower than anybody could have dreamed.
Yep, I agree with @Rangers ,Miyamoto has recently done an interview where he reiterated that Nintendo were (still) not interested in High end specs:
This could belong in the NX thread:
http://www.npr.org/sections/alltech...on-the-origins-of-nintendos-famous-characters
This could belong in the NX thread:
http://www.npr.org/sections/alltech...on-the-origins-of-nintendos-famous-characters
- Status