Predict: The Next Generation Console Tech

[upnorthsox]

Depending on what it actually is, one could easily fit such an APU in probably < 350mm² (even < 300mm² would be possible, if there are not more than 20 CUs). That shouldn't be a problem for 28nm yields mid/end 2013. If you want to fit some DRAM stacks together with the GPU die on an interposer, that interposer needs to be at least twice as large (probably closer to 3x the area). Last number I've read somewhere, is that an 300mm interposer wafer goes for 500 to 600$ (can be done with old tech), that's roughly an order of magnitude cheaper than a 28nm wafer. That means the interposer costs may be about a third of the APU die. But that is just a very rough guess and the assembly costs come on top of it.

My feeling is that 8 cores with 20 CU's may be possible to be around 275mm2, esp. if the bulk of the MC is moved to the dram stack(s). That may be too optimistic though. Still < 300mm2 imo is realistic.

I think your estimates are too high for what would be needed for dram stacks:

Hynix presented a 1.2V 38nm 2Gb DDR4 DRAM with a die size of only 43.15mm². The power consumption was roughly 50% lower than that of the chip’s DDR3 counterpart, even though the new chip’s core frequency is the same as the DDR3 part. Hynix also presented a DDR3 DRAM that uses a 23nm process to produce a 30.9mm² chip -- the smallest die size ever reported for this density. Hynix brought also a special clocking circuit for through silicon via (TSV) DRAM interfaces that uses sophisticated logic to align the clock with data from all the chips in a multi-chip stack.

Even with ddr4 you wouldn't need much more than 100mm2 of interposer area for 2 stacks. That then is an interposer in the 425-450mm2 range and a mobo around the size of an index card (10cm x 15cm).

 

[upnorthsox]

APU+GPU never made sense in my opinion (I always said this) and one can fit 4 or even 8 GByte with some DRAM stacks on an interposer.
4x4GBit per stack for instance, each DRAM die would omit the IO part of DDR3/GDDR5 (or the logic layer of a HMC), the DRAM array of a 32bit GDDR5 device accesses 256 Bit per address anyway (prefetch 8). One just needs to connect everything directly to the DRAM controller on the APU (or some logic on the interposer?). This would make for a 1024bit SDR interface per stack (running at let's say 750MHz, bandwidth would be equivalent to a 128bit GDDR5 6 GT/s interface [96 GB/s] per stack with 2 GB). Two stacks would be 4 GB and 192 GB/s, three stacks 6 GB and 288 GB/s, and 4 stacks would result in 8 GB and 384 GB/s.
I'm not saying this is what we are going to see. It is probably expensive and risky for a 2013 introduction, but in principle doable in my opinion.

Seems we've heard that bw number before. :smile:

 

[upnorthsox]

I'm not sure what your sources are, or where you got your information.

1st PS4 dev kit: ~1.25TFLOPS (6650D = 480gflops, 7670 = 768 gflops)
2nd PS4 dev kit: >1.38TFLOPS (7660D = 614GFLOPS, No mention of discrete GPU, at least 7670 though)

1st Durango dev kit: ~1.4TFLOPS (rumored to be nvidia GTX 570)

That info is only with regards to the rumors of the early dev kits for each system. I think that both will end up around 2.5TFLOPS when they release.

Is that the best you've got? From the same link you posted:

The HD 7670 is a rebranded version of last year's HD 6670, sporting identical specs across the board. If you recall, the HD 6670 is the same card that we reported will be used in the Xbox 720. The implication, of course, is that the two systems will effectively go toe-to-toe as far as discrete graphics are concerned.

The eurogamer article doesn't even mention a GPU let alone 1.4TFlops and quoting Theo isn't going to endear you to anyone here.

 

[N2O]

What if the consoles are like this?:

PS4:
- APU with 4 core steamroller at 3.2 Ghz and weak GPU
- 8 GB of RAM part (2 of which GDDR5) or 4 GB of GDDR5
- Discrete GPU AMD 8870

Xbox:
- APU with 8 Jaguar cores at 1.6 Ghz and AMD 8850 with eDRAM
- 8 GB of DDR3/4 RAM

The #2 AMD china guy also mention nextbox will be little better,but both will likely release at 2014
And he said don't need to care Wii U(performance)
http://club.tgfcer.com/viewthread.php?tid=6577631&page=14#pid15719863
"下代主机,是世界杯那年开买
明年,应该会有媒体开始吹风
两主机的性能360占优
WIIU就当打酱油的吧,此机莫提"

Btw #1 AMD china guy open a new thread
http://club.tgfcer.com/thread-6593085-1-1.html
He ask people why think APU=weak,and if AMD want we can make a 7970+8 core CPU in 1 APU,but will come with:heat,yield and size issue,that's why Durango delayed(don't worry(yet) he did not said delayed to 2014,maybe 2012->2013)

He also don't understand why MS so obstinate to make it in 1 APU

 

[MrFox]

I think we're getting very close to an amazing milestone, having one rumor for every possible combination of chip vendor, architecture, die size and memory configuration in the last 12 months. Most of them are only an anonymous post in a forum that was blown out of proportion by gaming news media looking for a scoop.

So far we've got: Larrabee, PowerPC, AMD x86, Intel X86, ARM, 32bit, 64bits, Nvidia Kepler, nvidia maxwell, AMD 6xxx, 7xxx, 8xxx, small APU, medium APU, huge APU, APU plus small GPU, APU plus big GPU, many small cores, few large cores, UMA, DDR3, DDR4, split pool, small edram, huge edram, GDDR5, 2GB, 4GB, 8GB, 16GB.

(but there's no rumor about HMC or HBM, that's weird, if I were making up rumors I would have used that)

 

[Proelite]

The #2 AMD china guy also mention nextbox will be little better,but both will likely release at 2014
And he said don't need to care Wii U(performance)
http://club.tgfcer.com/viewthread.php?tid=6577631&page=14#pid15719863
"下代主机,是世界杯那年开买
明年,应该会有媒体开始吹风
两主机的性能360占优
WIIU就当打酱油的吧,此机莫提"

Btw #1 AMD china guy open a new thread
http://club.tgfcer.com/thread-6593085-1-1.html
He ask people why think APU=weak,and if AMD want we can make a 7970+8 core CPU in 1 APU,but will come with:heat,yield and size issue,that's why Durango delayed(don't worry(yet) he did not said delayed to 2014,maybe 2012->2013)

He also don't understand why MS so obstinate to make it in 1 APU

The bolded part makes me suspect his validity.

 

[N2O]

The bolded part makes me suspect his validity.

I think both try to aim it holiday 2013,but also prepare if they really need delay it to 2014(like MS,i don't think they want RROD again)

 

[Albofighter]

Maybe they can fit an 8970 gpu. They will be designed for better efficiency right? Plus better performance and they can modify it to make sure it doesn't overheat. Why use an older gpu when they have plenty of time to test the new ones. Either way, I don't see why they can't easily do 3+ TF. The high end mobile 8xxx series alone will go over 2 TF.

Would also love SSD so that OS can be extremely fluid. Or at least have some flash memory just for the OS.

 

[JasonLD]

I think both try to aim it holiday 2013,but also prepare if they really need delay it to 2014(like MS,i don't think they want RROD again)

It will be really interesting if Sony actually makes 2013 release date while MS misses.

 

[N2O]

Maybe they can fit an 8970 gpu. They will be designed for better efficiency right? Plus better performance and they can modify it to make sure it doesn't overheat. Why use an older gpu when they have plenty of time to test the new ones. Either way, I don't see why they can't easily do 3+ TF. The high end mobile 8xxx series alone will go over 2 TF.

Would also love SSD so that OS can be extremely fluid. Or at least have some flash memory just for the OS.

He just use 7970 for example,he mean they can make any GPU in APU,but high-end GPU may come with some problems

 

[MrFox]

It will be really interesting if Sony actually makes 2013 release date while MS misses.

I'll be more interested if they both make it, it would be the best E3 ever, and a big needed boost to the whole industry. Both sides saying afterward that they "won" E3 whatever happens, it's going to be epic.

 

[JasonLD]

I'll be more interested if they both make it, it would be the best E3 ever, and a big needed boost to the whole industry. Both sides saying afterward that they "won" E3 whatever happens, it's going to be epic.

Agreed, lol. Sony and MS have never launched their consoles in the same year. It will be an epic battle if they both make it in 2013.

 

[MrFox]

APU+GPU never made sense in my opinion (I always said this) and one can fit 4 or even 8 GByte with some DRAM stacks on an interposer.
4x4GBit per stack for instance, each DRAM die would omit the IO part of DDR3/GDDR5 (or the logic layer of a HMC), the DRAM array of a 32bit GDDR5 device accesses 256 Bit per address anyway (prefetch 8). One just needs to connect everything directly to the DRAM controller on the APU (or some logic on the interposer?). This would make for a 1024bit SDR interface per stack (running at let's say 750MHz, bandwidth would be equivalent to a 128bit GDDR5 6 GT/s interface [96 GB/s] per stack with 2 GB). Two stacks would be 4 GB and 192 GB/s, three stacks 6 GB and 288 GB/s, and 4 stacks would result in 8 GB and 384 GB/s.
I'm not saying this is what we are going to see. It is probably expensive and risky for a 2013 introduction, but in principle doable in my opinion.

From what I understood lately, this sums up the whole dram problem. The major part of the die which is the dram arrays is pretty much the same things for GDDR5, DDR3, DDR4, WideIO, etc... it's the IO, buffer, interface, and signaling which are different, power hungry, inefficient, and needs to be a standard in order to have millions of identical chips produced, competition kicks in, and the cost goes way down. If you change or remove the IO part of GDDR5 it's basically an expensive custom chip.

HBM is similar in performance to what you describe, but it's a better choice because it's a standard. A single layer will be 1024bit wide, and 256GB/s per chip (1066 DDR), or lower power at 128GB/s per chip (would be either 1066 SDR or 533 DDR). There's no reason to make a custom memory chip, the best solution is right there in a JEDEC standard.

 

[Blazkowicz]

What you call a JEDEC standard, I thought it was a draft standard at best. The standard could be adopted, reach version 1.0 so to speak, after the consoles are released.

For now we don't even have a single name for this kind of new memory : Hypercube, three-letter initialism, interposer memory, 2.5D stacked memory, wide I/O memory.. I always wonder what to write when speaking about it.

 

[arijoytunir]

the recent sony patent of gpu context switching is also one of the capabilities of HSA . If the recent rumors of sony releasing a ps4 in 2014 is true then APU + GPU is a realistic prediction . By then the gpu s will be fully hsa compatible .

 

[MrFox]

What you call a JEDEC standard, I thought it was a draft standard at best. The standard could be adopted, reach version 1.0 so to speak, after the consoles are released.

For now we don't even have a single name for this kind of new memory : Hypercube, three-letter initialism, interposer memory, 2.5D stacked memory, wide I/O memory.. I always wonder what to write when speaking about it.

So what? designs were already in progress for DDR4 when it was just a draft. We do have a name for this new kind of memory, it's called WideIO. It exists now, but it's the low power version. The high power variant is a standard which is expected late 2012 or early 2013.

WideIO is what's in the Vita, and upcoming phones and tablet (probably the next iPhone). It's a JEDEC standard.
HMC is Memory Cube, which is not a JEDEC group (like Rambus, it's an independent standard). Final specs were expected for 2012, they should be here soon.
HBM is an extension of WideIO, developped by JEDEC, the preliminary specs they gave publicly are similar to WideIO but 1024bit, 8 banks, 128 to 256GB/s. Final specs are expected late 2012, early 2013. Some literature call it WideIO-2, other call it HBM (because it's the name of the task group which was created exclusively for it).

I agree the timing is a difficult one, but remember that the "final" specs of WideIO were released December 2011. That was about at the same time the Vita was released. The Vita worked on a draft specs.

 

[bkilian]

So why should anyone be excited over dev kit GPU if retail one is crap and you know it? Seriously, you don't know it and anybody on this board who is speaking of it doesn't either. All we know is what is in year old dev kit, nothing else.

You don't even know that. All you have is guesses based on a outdated backplane image taken from a setup doc.

 

[(((interference)))]

XB360 and PS3 were bitten by a forced transition to lead-free solder. Those issues should be understood and engineered around now. Any thermal issues would be down to cheap engineering. We may need expensive/loud cooling, but hardware failure shouldn't be the calamity this gen was.

Exactly, the myth keeps being perpetuated that the 360 had terrible hardware quality and the RROD debacle shows MS are bad at system engineering (usually to claim that the Wii U isn't actually so bad after all).

MS may not be Apple but they are certainly not cheapskates when it comes to hardware, unlike Nintendo.

My Dell XPS laptop and PS3 both died due to the exact same issue of lead free solder failing and unseating the GPUs from the motherboard, pretty much all nVidia mobile GPUs of the time had the same issues - including the MacBooks.

In the 360 it was just exacerbated as the casing was significantly smaller than the PS3 and only had a single fan.

 

[N2O]

#1 AMD china guy another new hint
http://club.tgfcer.com/viewthread.php?tid=6593085&page=2#pid15931245
"嘿嘿，我家有块APU，可是384BIT总线的哦"
He said Durango APU will be 384bit

 

[jlippo]

I think it is going to be a while until we hear about dx12. Heck, crysis 3 is probably thr first game that really uses dx11 in full scale and its coming out in next year lol.

Which has nothing to do with hardware capabilities and release date of an API.
2014 is currently the late date for 'DX12' or Shader Model 6 and what ever else the new hardware brings with it.