You have to keep in mind that GPU designs have been stuck at 40nm for quite a while due to problems at TSMC. AMD couldn't do much more than refine and improve efficiency with the 6X00 series since their transistor budget wasn't going up. The next jump, since they seem to be skipping 32nm strait to 28nm, should be enormous. I don't expect the PS4 and 720 will be likewise hamstrung on 40nm production when released in 2013/14.
Predict: The Next Generation Console Tech
- Thread starter Acert93
- Start date
- Status
You have to keep in mind that GPU designs have been stuck at 40nm for quite a while due to problems at TSMC. AMD couldn't do much more than refine and improve efficiency with the 6X00 series since their transistor budget wasn't going up. The next jump, since they seem to be skipping 32nm strait to 28nm, should be enormous. I don't expect the PS4 and 720 will be likewise hamstrung on 40nm production when released in 2013/14.
The next Xbox will have dual boot option...
You'll be able to do a normal "GM" boot as well as a new "PC" boot mode.
In PC mode you'll be able to use as a desktop computer running Windows XB..(special version of Windows compiled for Xbox CPU processor). In PC mode you'll be able to use both USB/wireless mouse and keyboard.
You'll be able to do a normal "GM" boot as well as a new "PC" boot mode.
In PC mode you'll be able to use as a desktop computer running Windows XB..(special version of Windows compiled for Xbox CPU processor). In PC mode you'll be able to use both USB/wireless mouse and keyboard.
Ninjaprime
Regular
The logical targets for eDRAM if they use it again would be either 64MB or 128MB I think. 64 would allow 1080p in 32 bit formats with x4 msaa in a single tile, 64 bit formats in 2 tiles, 128MB would allow it in all in one tile. Of course, I would guess lower cost shader based AA would become more popular in the future, so they might skip eDRAM altogether.
While I respect Carmac as a programmer he often makes rather weird speculations. Also, that 10x doesn't mean taking all console specs and adding a zero to them.
DuckThor Evil
Legend
Not entirely accurate. AMD 5870 on 40nm has 2.15B transistors on a 334mm2 die. whereas 6970 on 40nm has 2.64B transistors and a die size of 389mm2. AMD had to push slightly out of their comfort zone with the 6970 to get some difference between it and the older 5870. I don't have all the data at hand or in my head right now, but I'd say that with equal die sizes there is no performance differential between those two chips.
I'd say the absolute best case scenario for the next shrink is 2x6970, but I think it will more likely be something like 2x5870, not architechture wise, but transistor count etc. So something like 4.3-4.5B transistors for their single chip top GPU. Well see, but 5870 was closer to their sweet spot strategy than 6970, but perhaps nVidia is putting so strong pressure on them, that they have to adapt to bigger dies for good? maybe they'll come up with a 5.x B refresh later on? once the process has matured, We'll see.
The more important question imo is, how long until another 2x jump from this upcoming 28nm shrink? 28nm mass production looks to be in early 2012. 40nm was at the end of 2009 with poor yields. I don't think anything smaller than 28/32 will be available for the launch of Xbox next, but those processes should be close to as mature as 90nm was at the launch of X360. So if the next box launches in 2013, it'll probably be a system with quite hot chips at 28/32nm and get's a shrink to 20ish nm within the first two years of it's life. A quess from my cornholio says that the GPU power could be around 1.2-1.4 x 6970, not taking mem bandwith into account there though.
In fact the problem occurred TSMC delaying development of GPU, but probably Maxwell * (and its similar AMD / APU in performance) scheduled for reach 10 times the performance 22nm might not even be possible to launch time for the next consoles 2013/2014, probably because will have 4 billion or more ransistors with high TDP for closed box console levels below 200 watts.
So unless there are processes 3D gate transistors ** that provide enough jump allows low-wattage, I believe it is unlikely that we have something like maxwell a game console.
* http://hothardware.com/News/NVIDIA-Exposes-GPU-Roadmap-Kepler-Arrives-2011-Maxwell-in-2013/
** http://www.theengineer.co.uk/sector...nto-high-volume-manufacturing/1008529.article
(I'm praying here my thoughts are wrong)
Last edited by a moderator:
You right,cause even now Batlefield 3 using 158MB frame buffers * for 1080P,4*MSAA,maybe large EDRAM not enough for futures games.
* Page 40 and 41(FXAA)
http://www.slideshare.net/DICEStudio/directx-11-rendering-in-battlefield-3
Last edited by a moderator:
Surely if developers did fine with 10MB on the Xbox 360 targeting 720P then they should be happy with 40-80MB in 2012/13/14! If they are targeting only double the pixels Remember it's primarily a means of alleviating bandwidth constraints.
One thing to note is if they launch on a mature process they may be able to combine the ED-Ram onto the main die because IIRC TSMC has it on their roadmap for 28nm bulk.
One thing to note is if they launch on a mature process they may be able to combine the ED-Ram onto the main die because IIRC TSMC has it on their roadmap for 28nm bulk.
Edram could still make sense if they allow full read/write to it this time, that would allow for crazy stuff especially in post process. The 360's edram was comparatively hobbled because it allowed limited use of it.
That's what I was considering in one post but though it could prove expansive. The bus between tXenon and the daughter die offers 32GB/s of bandwidth. Assuming more EDRAM you need a faster bus. If you want R/W for the GPU it means more logic on chip. The thing could end big and costly, NEC is still stuck @65nm I believe for its EDRAM process.
JOKER454 assuming you have two chips and assuming a compute oriented GPU (even more than what AMD revealed yesterday) what do you think makes more sense: a CPU a GPU or 2APUs.
Clearly as AMD introduced FSA programming model, MS is to follow suit, the Asynchronous Compute Engine on the hardware side (which should like independent GPUs for developers, right?) coherent memory space with the CPUs and more is likely to come (the GPU in next system should newer/ more evolved) wouldn't it be better to have CPUs cores closer to GPUs cores?
On top of it you it can somehow help solving the memory bandwidth problem. Say APU 1 is connect to 2GB of GDDR5, same with APU 2 and there is a fast interconnect between the two chips. You end with twice the bandwidth usable either by CPU or GPU cores and the memory space is still flat. What your pov about it? I remember in an interview between Tim Sweeney and Andrew Richard whereas they were pretty much disagreeing on on everything they both agreed that if they have to deal with two chips they prefer to deal with twice the same chip (and they were considering heterogeneous chips as APU ).
More edram doesn't mean more bandwidth is needed. That's determined by how much data you need to write per clock.
Put a good tiling setup in there and keep it limited (say, 20-30MB) for cost reason and that might actually work. Could give the platform a vast bandwidth edge the others cant match. The transistor count cost might be comparatively small in the overall picture.
Well my info is pretty dated, I've been out of the console game for over two years now. But from what I've seen I'm not sure I see any advantage to using apu's in the console space. Personally I'd prefer using a smidgen of the alloted wattage on a bog simple cpu, then spend the lions share of wattage on a killer gpgpu. That way rather than get locked into what current apu's provide, you can allocate your priorities across cpu/gpu exactly as you want.
In fact it's likely not just an extra zero, but maybe he knows something more concrete and already working (on pcs...He loves work with openGL) with tools and estimate levels of performance( (fill rate,texel rate,flops,shaders ops,flops,geometry/tesselation,bandwidth etc) indicated by Sony and Microsoft for next generation,cause we know ID develops your engines over 5 or 6 years at least, so they need a more concrete informationas possible about the next consoles coming 2013/2014(if not my mistake he says 2 years from now).
Last edited by a moderator:
I modestly agree with your opinion, it is best left in the hands of developers where to put your sources and who produces consoles using fully customized and specialized processors for its native functions (CPU = general porpose,physics,AI, GPU = graphics) for maximum performance.
However once I read in one of Sony's personal interview* at the time of the launch of ps2slim, the junction between EE + GS (there was a final version with EE + GS + EDRAM ..) on same die tests showed performance gains 27% in tests against them separate chips.
I have seen recent tests/benchs on the APU llano with their respectives CPUs and GPUs "separate die" under the same clock etc, there are clear advantages in terms of graphics (up to 50% more performance) and relative low performance in cpu aspect (60% performance against quad cores "separate die" on the same clock),you already answered but in-depth discussion what would be better for developers about APUs and "separate processors' more performance on CPU or GPU?
*
(perhaps the first APU released ... unfortunately I lost the link)
Last edited by a moderator:
Did you get the opportunity to listen the Charlie's (from semi-accurate) interview of Tim Sweeney and Andrew Richards?
As I asked you this question yesterday I listened up to it again. Was not that easy to find, it's in the site archives and the vids are not listed on youtube. Here they are:
http://www.youtube.com/watch?v=tnogwO84O0Q
http://www.youtube.com/watch?v=oABs_HCnMBg
http://www.youtube.com/watch?v=8yoHsVfoCH0
http://www.youtube.com/watch?v=UIByXVphoM0
http://www.youtube.com/watch?v=GmwzrHmDwSQ
All in all I find it really interesting and I believe they discuss pretty critical things on hardware and software (human, power consumption but also communication overhead, the need for a unified platform actually it's a software or hardware platform). Now that I've read report about AMD last gen GPU and where there heading (I believe that in at least one console we will have the V2 of this new architecture) it sounds to me that their POV are a lot less exclusive than they look especially on software rendering.
Richards is really concerned about power consumption and efficiency whereas Sweeney is about convenience for programmers and not TCO but total cost of supporting an architecture or more precisely of improving your code over years/decades. From my understanding GPUs are getting dangerously close to what Sweeney wants, a software platform is close to emerge and to win imho. The hardware is most likely way more efficient than for example larrabee. it's getting close to the best of both worlds, the link will software language allowing to using both CPUs and GPUs in a sane (not easy) fashion.
In this context I believe that we could be close to a rebirth for software rendering, the power and the flexibility to do so is almost there, with some radical design choices it could be there in time for the next gen systems. I'm not interested in software rendering for the sake of it but because I believe Sweeney is right, one rendering method doesn't fit it all. Something that concern me too is bandwidth requirement. AMD hints at "compute approach to graphics" implementing a software pipeline (on an efficient an specialized hardware as up coming GPUs) may allow you to use an approach that makes the most of data locality, a bit like what larrabee did. It could for example like sending bins to CU (compute units) and then the data don't move from the CU everything from vertex workload to rasterization to shading to post processing. It allow this but it allows more standard approach too. It may also allows to implement other approaches (could be with halp from CPUs cores if latency overhead is low).
I start to believe that this is the way to go for next gen especially as I expect their life spawn will be as long if not longer as the one of our consoles. Devs have some years to get ready and a decade to improve, once the software language is ready the code will be more perennial it could prove worse the investment. Editors did not want new systems due to huge investment made but they started stating this some years ago, system won't be there before fall 2013 at best, events bought them extra time and they may actually deal with a solution that will allow them to design something more perennial once again might prove worse the investment
One thing I wanted to ask for quite some time: Supposedly, developers like unified memory for the increased flexibility. The problem, however, seems to be that latency is a huge issue in that case, as you want high bandwidth, high latency RAM like GDDR or XDR for the GPU, while the CPU wants low latency RAM. Now, neither Microsoft nor Sony seemingly cared about that with 360 and PS3, while Nintendo decided to continue using 1T-SRAM even though it's very expensive. They also used expensive pseudo-static low latency RAM on DS, DSi and 3DS. They seem very concerned with memory latency on the CPU side.
If the press release concerning the WiiU CPU IBM issued last week and the Engadget rumor are anything to go by, whatever CPU Nintendo will use will employ significant amounts of eDRAM, which probably means IBM/ Quimonda eDRAM based lv3 cache. I assume the idea is to balance out latency issues in a unified memory environment, correct? It should also mean that the CPU wouldn't need access to the unified memory pool as often, increasing the bandwidth available to the GPU? If my assumptions are correct, I wonder if developers on here consider that a good approach, and if Sony and Microsoft might follow suit with their next generation systems?
If the press release concerning the WiiU CPU IBM issued last week and the Engadget rumor are anything to go by, whatever CPU Nintendo will use will employ significant amounts of eDRAM, which probably means IBM/ Quimonda eDRAM based lv3 cache. I assume the idea is to balance out latency issues in a unified memory environment, correct? It should also mean that the CPU wouldn't need access to the unified memory pool as often, increasing the bandwidth available to the GPU? If my assumptions are correct, I wonder if developers on here consider that a good approach, and if Sony and Microsoft might follow suit with their next generation systems?
Ninjaprime
Regular
I think you're setting yourself up to be sorely disappointed. PCs will be lucky to be 2-3x as fast by the time the next consoles launch. Those are 500-600w beasts. With a third that power as a ceiling, I'd say the next gen consoles are going to be hard pressed to hit 20x PS360 levels, more likely 10-15x or so. Furthermore, reguardless of your like or dislike of diminishing returns, its a real thing, and the closer you get to realism, the harder it is to tell the difference between power levels. Many polygons spend their time below the size of a pixel now, whats more polys gonna add? How much smoother is a 200k poly model compared to a 20k model with todays techniques? Not a lot.
- Status
Similar threads
- Locked
