Maybe it's an SOC?
I'm getting confused LOL. With all this talk of multiple CPU/GPU's would they do two 128 bit pools?
Next gen RAM choices *spawn
- Thread starter aaronspink
- Start date
Maybe it's an SOC?
I'm getting confused LOL. With all this talk of multiple CPU/GPU's would they do two 128 bit pools?
Well that would be nice, I've been ranting for a long while that the perfect system for me would be SoC in the + 300 sq.mm range fed on a 256 bit bus
Wrt to the rumors of late it's getting impossible to make sense out of it. After what appeared a mixed up of Sony annd MS system specs and the early "design guidance" from MS which looks legit, I would not dare a bet at this point especially on MS system.
There are too much discrepancies between what we hear now :S
The throughput and power consumption of 8 or 4 jaguar cores and 4 or 8 BD cores are not in the same ballpark.
The amount of memory is troublesome too, 4 or 8 GB. I still keep somewhere in my mind that devs kits can have twice the amount of memory than shipping system for convenience.
Aaron Pink is a blessing for us and from his hints discarding inter-posers, memory stacking, etc. sounds safe. It's not ready for prime.
Then there is edram. It has a lot of benefits and introduce limitations too. Looking at the state of PC development I would say that if Edram is part of the design it can't be implemented as it was in the 360. Shortly I would bet that to meet modern engines requirements it would need to use a plain scratch pad memory. So as I see things, the edram (if there is edram) has to be on the same die as the GPU. It would accessible to both read and write for both the ROPS and the Shader ALUs.
There is only one company that can produce such a chip @ 32nm: IBM. I wonder about their willingness to produce something that doesn't include PPC tech (a gpu in that case) if they are, being the only option, I wonder about their willingness to let that happen for cheap.
A nice thing with what we heard about the ps4 so far is that it should use pretty "cheap" (if 28nm qualifies as cheap) process. All the chip APu + GPU or a big APU should be made on the same process.
Looking at MS and the leaked early guidance for the system, it's really tough to guess how it could have evolved since then.
Plenty of thing could make sense lol.
For example something like this:
3 chips
CPU: 8 jaguar cores or 4 piledrivers cores (2 modules). +100 sq.mm
GPU/north bridge/scratchpad memory: 8/12 CUs between 32 and 64 MB of edram, 200 < x < 300 sq.mm?
128 bit bus to 4GB of ddr3 or ddr4.
OS CPU: Arm SoC with stacked RAM (1 GB) handling low power/OS operation both the main CPU and GPU can be power killed. (could be a tegra3 or whatever MS plann to support for Win8 RT at the time).
Last edited by a moderator:
french toast
Veteran
What if the os cpu is something like tegra 4...which controls the os, and also kinect? Multitasking?...there would be more than enough power there to super charge kinect and system.
I'm thinking the gpu will be more powerfull than that...I was hoping for 3 flops...that would also give them a sizable marketing muscle..who's got the most flops..etc etc.
What kind of edram would we need to run 1080p and lots of goodies?
Would love the idea of 8 jaguar cores with 128bit simd attached...there would be plenty of room for devs to get to grips with all those threads!...how much l2 cache to make the most of them...8mb shared??
Would it be possible to weld on a 256bit simd to those jaguar cores?
I'm thinking the gpu will be more powerfull than that...I was hoping for 3 flops...that would also give them a sizable marketing muscle..who's got the most flops..etc etc.
What kind of edram would we need to run 1080p and lots of goodies?
Would love the idea of 8 jaguar cores with 128bit simd attached...there would be plenty of room for devs to get to grips with all those threads!...how much l2 cache to make the most of them...8mb shared??
Would it be possible to weld on a 256bit simd to those jaguar cores?
The HD7770 is 123 mm^2.
http://www.tomshardware.com/reviews/radeon-hd-7770-7750-benchmark,3135.html
That appears to be far too small to small for a 256-bit bus, just as the Xbox GPU revisions were. If you intend to use a chip that is too small for a certain size of bus, or intend to shrink to a size at which the chip will be too small, it's probably a pretty bad idea to use that bus. The reason being that it wouldn't fit.
The reason for using Edram in this situation wouldn't be to "waste money" or "waste people's time" as you suggest, but for a reason such as massively increasing bandwidth more cheaply than massively swelling the chip up so you can fit a 256-bit bus (which would also require 2X the minimum number memory chips, larger mobo, lower yields, no shrinks, limited SoC potential, etc etc etc).
Well that would be nice, I've been ranting for a long while that the perfect system for me would be SoC in the + 300 sq.mm range fed on a 256 bit bus
Well if the 192GB/s is correct then it's hard to see how they could be aiming at a small chip! It takes a pretty fast GPU to justify that kind of bandwidth - at least that's the impression you'd get from looking at PC GPUs.
Yeah, read/write access seems essential this time, but you could still do that off-chip. An off-chip bus would seem to place more limitations on BW and increase latency though, which might be an issue if you wanted fast access to it from the CPU?
I think Global Foundries have some kind of agreement to co-fab some stuff with IBM, which presumably includes making none PowerPC stuff at competitive rates. More of an issue might be a roadmap for a smaller node to move production to over a 10 year product lifespan. But if IBM can offer some kind of guarantees about that (and presumably they'll need some kind of
2nm edram friendly process for their POWER line of server chips) then maybe IBM would be a goer for GPUs and SoCs.That IBM 32nm press release made it sound like they were confident of making console stuff on it...
Oh come on Rangers, we have been posting back and forth for years, you can take a like smack talk
And you were not getting jawed at for suggesting a Cape Verde / 4core CPU / 8GB DDR4 system (a simple, well balanced lower end console), I was giving you some smack because you called it TASTY! Tasty?? Come to your sense partner!
Don't drink the Kool-Aid "we scared you with the 6670 rumors, now sing our praises for a 7770!" Oh, did I leak the secret methodology to the MS leaks. opps I would not put much faith in that number, it's more than what the gtx 680 provides and never before saw a chip connected to 256bit bus achieves such bandwidth (south of 180GB/s).
If I look at Pitcairn, that's 212 sq.mm for the GPu + memory interface so my idea was 100 sq.mm or more devoted to the CPU
By the way at the time I was considering IBM CPU.
Well I think one has to do trade off the edram can't be on both the GPU or the CPU and the link between the 2 can only be that fast.
As far as CPU perfs are concerned I wonder if actually the CPu could act as the north bridge and the GPU would access memory through it. I don't know the bandwidth requirements but looking at the 360 I would not be surprise if ultimately the same memory organization makes more sense this time again.
Honestly I would hope they pass. Depending on their target perf I still think a SoC +a 256 bit bus is the best possible choice.I think Global Foundries have some kind of agreement to co-fab some stuff with IBM, which presumably includes making none PowerPC stuff at competitive rates. More of an issue might be a roadmap for a smaller node to move production to over a 10 year product lifespan. But if IBM can offer some kind of guarantees about that (and presumably they'll need some kind of
That IBM 32nm press release made it sound like they were confident of making console stuff on it...
I've read Aaron Pink comment on DDR4 and its price advantage against gdd5 but for now I see gddr5 going nowhere no GPU manufacturer is going to sit on +50% bandwidth (and more) even it comes at a premium (vs DDR4) So for me Gddr5 is here to stay till inter-posers, wide IO interfaces, memory stacking, etc. are ready for prime time. As he said himself it could be with 3/5 years.
I would not be that surprised if MS, Sony Nv and AMD ask that the GDDR5 road map is pushed a bit further. So we may see even fasster memory and why not 4Gb memory chips?
Overall whether it's a SoC or an APU+GPU I'm pretty much liking what I heard so far about the ps4.
It sounds straight forward from a developer perspective, I'm confident they would pull the shit out of it even though it were to come with only 2GB (still hope for 4 though).
On a bit different topic and looking at MS guidance and Sony offering, I've wondered at job today if it could make sense for Sony to also include another SoC to the system to run low power and OS operation.
I got this idea even though I don't expect it to happen it sounds great to me. I don't know how much the Vita SoC (if it's a Soc I'm not sure actually though I would hope) + the RAM would cost next year but I believe it could be a great addition to the ps4. It could run the OS, and more (why not the same games as the vita may be in higher res if the GPU can be over clocked).
They could have (like MS if they indeed end up with an Arm SoC) PSN games developed for both the vita and the ps4, easy cross platform gaming looking forward, etc.
Overall I wonder if it would make them more good than 2 extra GB of ram.
If they are even serious about going with their own OS and trying to carve a niche from them selves they have to start somewhere, PS vita hard (at different clocks, extra ram) either in TV, ps4 or the psv could set the basic specs for some years to come.
Any opinion on the matter?
Oh come on Rangers, we have been posting back and forth for years, you can take a like smack talk
And you were not getting jawed at for suggesting a Cape Verde / 4core CPU / 8GB DDR4 system (a simple, well balanced lower end console), I was giving you some smack because you called it TASTY! Tasty?? Come to your sense partner!
Heh, as I explained it's tasty compared to Cape Verde, the working assumption of whats in Durango...
Just noted a little hypocrisy there on your part
Yeah, it seems unlikely, but it's nice to be optimistic once in a while! Only a beast of a system would require that much bandwidth, unless the BW was so cheap that there's no real reason make it available.
I guess there's four possibilities:
1) The 192 GB/s comes from very fast GDDR5 on a 256-bit bus
2) The 192 GB/s figure comes from a dev kit (GPU + main ram) and does not represent the final system
3) The 192 GB/s comes in part from edram
4) The 192 GB/s is LIES and should not be considered as confirmed, as I think another poster had indicated
If I look at Pitcairn, that's 212 sq.mm for the GPu + memory interface so my idea was 100 sq.mm or more devoted to the CPU
By the way at the time I was considering IBM CPU.
My "best likely PS4" was about the same, Pitcairn + 2 AMD CPU modules. With a bit of luck we'll get something like that, although a Cape Verde like 123 mm^2 + two modules would be closer to the 200 ~ 250 mm^2 figure that has previously been the biggest stuff we've seen for consoles.
I think traditionally the memory controller has gone with the GPU because of design and bandwidth reasons (e.g. the GPU designer also did memory controllers, and the GPU requires the most bandwidth) but seeing how GPU's are supposed to be the most latency tolerant and how CPU vendors now integrate memory controllers, maybe this could change in the future...?
I guess the idea of a SoC makes this irrelevant. A SoC with a slab of fast edram (like the 360S) could maybe make sense in the future. That was you get to shrink (full node or half node) while continuing to make your edram on an increasingly cheap process.
Probably for developers (especially if it's GDDR5 we're talking about) but maybe not for the platform vendor, depending on their long term cost plans.
I see things the same way. Given how hard high end GPU's push die sizes and power/thermal limits I can't see anyone giving up 50 ~ 100% additional bandwidth. For the mainstream and console markets things might be very different though - this is where cheaper memory and/or smaller buses and/or edram have previously been used. I still think smaller buses and possibly cheaper memory, paired with edram, will make sense for some player(s).
GPUs don't seem to have hit the limit of the existing GDDR5 roadmap yet, so maybe limits of the technology and manufacturing processes are starting to hold them back? I guess 4Gb GDDR5 chips will come along once the money is there, or are larger chips at high speeds likely to bin even worse than current top end GDDR5 memory chips (larger chip for the signal to propagate over)?
I don't know about these things, but it does seem there has to be a reason for the holdup?
I'm quite excited about everything next generation! As long as there isn't a large OS reservation and as long as we have a HDD or flash cache I don't think 2GB will be disastrous. If everything has to stream from a 6X Bluray drive and 512MB is reserved for the OS, however, then that could be a bit disappointing ...
I think the idea of having a lower power mode for service provision is a great idea, but surely you could downclock the main processors and shut off whatever you're not using? AMD APUs seem pretty power efficient, and cross compilation and/or some kind of VM/emulator could possibly be a cheaper way to handle a common platform. Maybe you couldn't get quite the same power efficiency, but given he cost of the processors, motherboard complexity and possibly other stuff (like duplicating memory or using hardware to emulate memory access) I think that maybe the less elegant solution is more practical and cost efficient?
Prepared to be disappointing?
iirc the PS3 was closer to 50GB/s aggregate between XDR and GDDR3. 8x would be 400GB/s!
Memory bandwidth has not scaled as quickly, as you now, as other parts in chip design. That is why stacked memory -- and waiting until 2014 -- seems to me to be the "duh" position because, as some much more informed than myself have noted, it offers a significant "generational" leap in one of the biggest design bottlenecks. Of course technology availability is NOT the timeline consoles are developed on but stacked memory or biting the bullet for a large eDRAM cache bandwidth is a premium -- especially with GDDR5 being about 4x the cost of DDR3 and the cost of a 256bit bus -- or multiple busses.
160GB/s would take a 256bit bus with high end, power hungry, expensive GDDR5 or possibly a 384bit bus. I wish it was something we could realistically hope for but the rumors point toward this being VERY unlikely.
I chatted a bit with the friendly neighbourhood EE yesterday and he told me that the reason the DDR4 spec is so delayed is that they want to have a stacking interface for it from the get go, so that: "products with large bandwidth demand can start with a wide bus and cost-reduce to stacked ram later."
Is it just me or does this sound very much like a wide DDR4 interface?
Is it just me or does this sound very much like a wide DDR4 interface?
Would that mean a 256-bit bus with traditional chips could be scaled down to a 128-bit bus with double stacked chips? I would have assumed the bus width would have remained the same, but I haven't read any specifics on how stacked memory works.
If that is the case it makes DDR4 all the more viable for next gen consoles.
If that is the case it makes DDR4 all the more viable for next gen consoles.
Hmm, could be pretty relevant to Durango.
Not anymore than it can with just bigger chips. It would drastically change the performance charasteristics, so it would probably not be done.
Well, yes, the bus width is essentially visible to programs so it should probably be kept as is.
The idea isn't just stacking the memory chips on themselves, it's stacking them on the *logic*. So that you can draw your traces as short, narrow, cheap ones on silicon instead of having to drive the signal out onto the board.
This can be done by stacking the chips directly on the APU, but that probably has heat dissipation issues. A more realistic way is to build a cheap, large chip (called the silicon interposer) using an obsolete and cheap process, and then stacking everything on it.
________________________
My post from the other thread
" PS3 had 128bit main memory bus, DDR3 would have only been able to manage about 12GB/s in the PS3 & if the PS4 is 256bit DDR3 would only get about the same main memory bandwidth as the PS3 using 4 64bit 1GB DDR3 chips.
So it would be 4GB of memory in the PS4 but only about the same bandwidth as the PS3.
(That's if I'm doing this right.)
& with that same 256bit bus the PS4 could have 224GB/s (about 10 X the bandwidth of the PS3 ram) using 8 32bit 256MB GDDR5 chips for 2GB of ram.
or 409.6 GB/s (about 20 X the bandwidth of the PS3 ram) using 8 32bit 512MB XDR2 chips for 4GB of ram. "
_____________________
But I'm not sure if I have the right formula for adding all this up so feel free to correct my mistakes.
Intel to do away with DRAM in PCs
I think this will be a lot like what PS4 will be doing with it's Memory.
I think this will be a lot like what PS4 will be doing with it's Memory.
