Predict: Next gen console tech (9th iteration and 10th iteration edition) [2014 - 2017]
- Thread starter Shifty Geezer
- Start date
- Status
D
Deleted member 11852
Guest
They may well be on track now but that can change quickly. We have decades of silicon wafer production experience but that never seems to prevent initial yield problems with newer process changes. Moving from silicon to a new material will be huge challenge. It's not just about layout and design, but new production equipment, new material suppliers. The entire chain from R&D to production and testing will have to change.
Is this removing the command processor, or just supplanting its role for initiating wavefront launches?
Absent more significant reworkings of the GPU, just taking any single sub-processor out and replacing it with a host CPU may lead to there needing to be one or more sub-processors between the CPU and the internal networks and custom queues that the original processor interfaced with.
Significantly more than coherent caches would need to be done, since a good chunk of the CP's interactions with the CUs and fixed function pipeline wouldn't go through a cache, and caches don't do push updates well. Of course, all those clients could be upgraded to have their own little processors that poll for updates to the queues in memory.
From a conceptual point of view, the CPU would be at least as suited.
From a practical aspect, at least right now, the existing CPUs would not given the nature of implementation. They physically do not have the same linkages, so a reworked core and likely reworked GPU architecture would have to come hand-in-hand.
Involving a host CPU is also where I noted the difference in opinion between AMD and others when it came to kernel visibility of the GPU's handling of the worklist. Involving a host CPU would very likely mean involving the kernel, which at least AMD doesn't care to do for the time being. Giving the client domain access to a host CPU without involving the system would reopen the system concerns in a big way.
Watching Nvidia's GTC keynote today and the release of the 7 teraflop Titan X video card and then the information about how Pascal with 3D stacked memory/nvlink etc next year is going to be potentially up to 10x Maxwells performance...I don't see this generation of consoles lasting long. With the 390X out sometime this year and Pascal next year we're already approaching 4K gaming as a feasible option. Xbox One is averaging 900p. The hardware gap is just too far now for these consoles to keep up, especially Xbox One. Also the Xbox One is already priced at $350 (sometimes less on sale) less than a year and a half after launch. Microsoft's tactics now smell of the 6th generation and the original Xbox. I wouldn't be surprised it we get the next version in fall 2017.
My predicted specs of the next Xbox:
10-15 teraflop gpu
16GB of stacked high bandwidth memory
Something like Nvlink built into the board to handle CPU<->GPU bandwidth
1TB of flash storage
$500 price tag
My predicted specs of the next Xbox:
10-15 teraflop gpu
16GB of stacked high bandwidth memory
Something like Nvlink built into the board to handle CPU<->GPU bandwidth
1TB of flash storage
$500 price tag
Just to qualify that, it's going to be up to 10x for "deep learning" - http://i.imgur.com/4wFctzF.png - and he qualified that with "CEO math" beforehand. http://i.imgur.com/25pz1fx.png - 4x the FP16 (I'm assuming double the FP16 rate like in X1 and double the shaders in total, making 4x) and double the interconnects (8 GPUs v 4) with a bit of a clock speed increase makes about 10x. For games that'll be a tad optimistic shall we say.
yeah but everything he was referring too was "for deep learning" lol...so I'm not sure the 10x performance increase for Pascal is only referring to "deep learning" applications...either way even if it's half that it would be impressive
Love_In_Rio
Veteran
Well, wouldn´t in theory a powerful gpu (10-15 tflops?) that allowed to use realtime GI -voxels based for example-in a non limited way make developments easier/cheaper?. Of course, as someone has said, if next gen we jump again in resolution more power will be for nothing.
Prophecy2k
Veteran
With VR on the horizon and pretty much all involved parties agreeing that more than 1080p resolution is ideal, I can certainly see this happening.
upnorthsox
Veteran
For VR, not sure I'd make that distinction for non-VR.
Prophecy2k
Veteran
But the next-gen consoles are almost certainly going to be designed around being able to do high quality VR.
It's the easiest selling point to get people to upgrade, since power alone doesn't seem to keep gamers happy unless its accompanied by [mythical] "next-gen gameplay™".
It's the easiest selling point to get people to upgrade, since power alone doesn't seem to keep gamers happy unless its accompanied by [mythical] "next-gen gameplay™".
Millions and millions (maybe about 35 million now) have already bought PS4 and X1 which are just more powerful traditional consoles...
VR or not to VR is up in the air IMO, actually probably is giving next gen planners fits right now, because a lot depends on if it takes off or not.
I guess just make a powerful console and then you can add on VR later? But those expensive goggles are going to be a problem. Could be just like Kinect again, company A packs in goggles for the all in one solution at the expense of less powerful base hardware. Company B ships more powerful hardware stand alone. Company B will have the edge of course.
VR or not to VR is up in the air IMO, actually probably is giving next gen planners fits right now, because a lot depends on if it takes off or not.
I guess just make a powerful console and then you can add on VR later? But those expensive goggles are going to be a problem. Could be just like Kinect again, company A packs in goggles for the all in one solution at the expense of less powerful base hardware. Company B ships more powerful hardware stand alone. Company B will have the edge of course.
D
Deleted member 86764
Guest
I seriously doubt that'll happen EVER again. They just need to make some games for VR and some for the main console with all available at launch. That'd be the difficult part.
Headset should be optional (like how Kinect 2 should have), or packed in, then the manufacturer can make a tidy profit on both.
The thing is they need to making these decision right now on the next console, before VR actually launches anywhere.
From what I've seen working at semiconductor companies, it will take 3-5 years (or even more) to produce new silicon. Typically the teams are pipelined in such a way that they're always busy and everyone is working on some different chip, but from start to finish the design takes a long time.
If I were to guess on the timeline for a console development: 12-18 months for the architecture specification. Even if both Sony and MS go with AMD and leverage existing (or upcoming) AMD hardware, they'll want to prove out the solution in software through simulation and thorough study. Once that is done and everyone is satisfied, my guess another 12-18 months for RTL design and verification. They'll probably do a couple of spins of the silicon add 6 months for fabrication and another 3 or so in between the revs for actual on silicon testing. So you're looking at close to 4 years for the development time.
If the console developers were to wait to see how VR does (and I don't think you will be able to gauge anything before the end of 2016), you're probably looking at a 2020 or later next-gen. So my guess is that they're going to have to make a gamble on whether to pursue it now or not.
From what I've seen working at semiconductor companies, it will take 3-5 years (or even more) to produce new silicon. Typically the teams are pipelined in such a way that they're always busy and everyone is working on some different chip, but from start to finish the design takes a long time.
If I were to guess on the timeline for a console development: 12-18 months for the architecture specification. Even if both Sony and MS go with AMD and leverage existing (or upcoming) AMD hardware, they'll want to prove out the solution in software through simulation and thorough study. Once that is done and everyone is satisfied, my guess another 12-18 months for RTL design and verification. They'll probably do a couple of spins of the silicon add 6 months for fabrication and another 3 or so in between the revs for actual on silicon testing. So you're looking at close to 4 years for the development time.
If the console developers were to wait to see how VR does (and I don't think you will be able to gauge anything before the end of 2016), you're probably looking at a 2020 or later next-gen. So my guess is that they're going to have to make a gamble on whether to pursue it now or not.
Persistantthug
Newcomer
Well, since i was right about the 8th generation's 8 gigs of ram:
https://forum.beyond3d.com/threads/...ation-console-tech.29254/page-369#post-752850
I now designate myself MASTER OF THE RAM PREDICTOR
128GB of ram for the PS5!
I thus hereby decree it.
https://forum.beyond3d.com/threads/...ation-console-tech.29254/page-369#post-752850
I now designate myself MASTER OF THE RAM PREDICTOR
128GB of ram for the PS5!
I thus hereby decree it.
Well, since i was right about the 8th generation's 8 gigs of ram:
https://forum.beyond3d.com/threads/...ation-console-tech.29254/page-369#post-752850
I now designate myself MASTER OF THE RAM PREDICTOR
128GB of ram for the PS5!
I thus hereby decree it.
Goodness! Nearly everyone vehemently denied the possibility of 8GB of ram.
That's quite a trip back down memory lane. It would be somewhere around 6 months before the capacity debate returned to that thread. The conservative position was reluctant to bet on a capacity of 8, particularly when some were counting on a 2012 launch.
After that seemed ruled out, 4 had more adherents, and it wouldn't be evident for a while longer where Durango's memory architecture would go. GDDR5 remained a question mark since the math couldn't be worked out with any known density at the time, something that Sony itself agonized a bit over until near the end.
After that seemed ruled out, 4 had more adherents, and it wouldn't be evident for a while longer where Durango's memory architecture would go. GDDR5 remained a question mark since the math couldn't be worked out with any known density at the time, something that Sony itself agonized a bit over until near the end.
Initially on B3D even 4GB was rather scoffed at. 8GB would have been deemed unimaginable. 2GB was the expected.
It's one area we got a nice return even if we didn't from the boxes GPU's.
But back to next gen, even 16GB seems like a lot to ask. But I guess Titan X just dropped 12GB GDDR5...and they're talking about 16GB of stacked memory.
It's one area we got a nice return even if we didn't from the boxes GPU's.
But back to next gen, even 16GB seems like a lot to ask. But I guess Titan X just dropped 12GB GDDR5...and they're talking about 16GB of stacked memory.
It took another six months before that capacity clawed back into the discussion. It was considered a nice thing to have on the wish list.
There are diminishing returns. I suppose someone could imagine what they would do with hundreds of gigs of RAM, but I think at that point we would need to consider improving the IO of the platform to match.
The progress for non-volatile memories may be something that deserves more focus.
- Status