Predict: The Next Generation Console Tech

[3dilettante]

The cautionary statement when bringing up a reference that discusses NTV is that going for the very high efficiency range comes at a significant penalty in straightline performance.

Intel's demo chip pulled several mW at 3 MHz.
Extreme power savings are great, but there are lower bounds of what is acceptable.

 

[hoho]

Intel's demo chip pulled several mW at 3 MHz.

It might be a "little" bit offtopic but any idea at what frequency might the Cortex-A8 usally idle that's inside phones? Sure, there is the problem of keeping the connection with network up but I can't see it being that big of a power hog considering we have dumbphones lasting for weeks on batteries much smaller than what smartphones use. According to some battery measurement device my n950 idles at around 3-4mW when I don't have any background tasks running, I'm in area with good coverage, the screen is completely turned off and aren't connected to mobile internet.

 

[anexanhume]

The cautionary statement when bringing up a reference that discusses NTV is that going for the very high efficiency range comes at a significant penalty in straightline performance.

Intel's demo chip pulled several mW at 3 MHz.
Extreme power savings are great, but there are lower bounds of what is acceptable.

Yeah, I was more going for the relationship between efficiency and operating voltage as the question was asked what high end was possible. I simply remembered the graph from having read about NTV recently. It's a more relevant topic for handset discussions, but the graph still applies.

 

[3dilettante]

It might be a "little" bit offtopic but any idea at what frequency might the Cortex-A8 usally idle that's inside phones? Sure, there is the problem of keeping the connection with network up but I can't see it being that big of a power hog considering we have dumbphones lasting for weeks on batteries much smaller than what smartphones use. According to some battery measurement device my n950 idles at around 3-4mW when I don't have any background tasks running, I'm in area with good coverage, the screen is completely turned off and aren't connected to mobile internet.

I'm not certain of the absolute lower limit. Anandtech said the A8 in the OMAP 3630 could drop to 300 Mhz.

The next rung would be stopping the clock, then power gating is the next big step.

 

[pjbliverpool]

25x the performance at what kind of code?

As for power usage, from PS2 -> PS3 didn't the first-release console's power usage rise several times? I'm fairly certain original PS2 consumed considerably less than 100W.

Shader code presumably. The 20-25x figure comes from 17.5x more theoretical performance * 1.5 for efficiency gains of GCN over Xenos.

Texture fill rate is also around 15x the theoretical performance but GCN had shown 50% performance improvement texel for texel over the 6xxx series, nevermind Xenos.

 

[Kaotik]

It might be a "little" bit offtopic but any idea at what frequency might the Cortex-A8 usally idle that's inside phones? Sure, there is the problem of keeping the connection with network up but I can't see it being that big of a power hog considering we have dumbphones lasting for weeks on batteries much smaller than what smartphones use. According to some battery measurement device my n950 idles at around 3-4mW when I don't have any background tasks running, I'm in area with good coverage, the screen is completely turned off and aren't connected to mobile internet.

Not sure of Cortex-A8, but on Qualcomm Scorpion you can drop to around ~150MHz without problems on the basic phone funcions, though you have to do this by 3rd party over/underclock tools of course

 

[Rangers]

Pitcairn specs that look real:

HD7870 would look tasty in a console, 2.56TF fulfills Epic's requirement for Samaritan at 1080P. Is about 10X Xenos flops which I think is a respectable generational leap.

7850 would definitely seem doable in a console powerwise, but I think 7870 could be too.

 

[Rurouni]

On 8255, I set my minimum speed to 122mhz. If the monitoring software that I use can be trusted (I think it's accurate enough.. I hope), I use as low as 50ish mw on idle with the average around 150mw (again on idle). I have twitter, mail, ym, fb, and god knows what else silently using data connection in the background. On top of that, my GPS is always on (but not actively being used), I use wifi at home (which is where the idle numbers come from and that was when I was sleeping), cell network always active and mostly on 3g.
Btw, I also undervolt it so that at 122mhz it use 800(mv?) instead of the default 875.

 

[TheChefO]

Pitcairn specs that look real:

...

HD7870 would look tasty in a console, 2.56TF fulfills Epic's requirement for Samaritan at 1080P. Is about 10X Xenos flops which I think is a respectable generational leap.

7850 would definitely seem doable in a console powerwise, but I think 7870 could be too.

Tasty indeed.

It seems AMD cut quite a bit from the 7850 performance wise. I'd say somewhere between the 7850 and 7870 would be a good compromise. At least 2TFLOPS of GPGPU goodness.

I'd be curious to see if one of these would be better perf/watt than an underclocked Tahiti.

 

[bgassassin]

Look at what happened to MS with Valhalla, they went to a SOC chip but couldn't take advantage of any intercore efficiencies.

Wasn't that intentional to keep it from being "more powerful" than the previous 360s?

2.56TF fulfills Epic's requirement for Samaritan at 1080P.

Until Epic shows Samaritan running at that resolution on one AMD GPU, it doesn't fulfill it.

 

[Megadrive1988]

SemiAccurate has a new write-up

In the end, it looks like Sony is going to go for the take no prisoners option on the PS4. If you don’t push fab limits that hard, but do push advanced packaging to the limit, you could very well end up with a monster that is simply not manufacturable as a single die. It won’t be cheap, but it will undoubtedly punt a single chip, or a single chip with stacked DRAM, in to the weeds.
Once again, the end result comes down to the age old question of can they make it? On the surface, the answer is yes, but once manufacturing begins, things may not be quite so rosy. The talk from Sony about the PS4 that seemed like so much of a pipe dream last spring seems, well scary realistic right about now. We don’t expect the PS4 before late 2013 best case, 2014 seems much more likely, so things may change a lot before you can buy one.
So in the end, we close with a simple thought, the Playstation 4 is almost undoubtedly an x86 part with AMD graphics too. That is only the very beginning though. If Sony can back up the boasting with real silicon, and the packaging elves can make it in quantity, it should be a game changer, pun intended. Sony is aiming for the moon just like they did for the PS3. Let hope they come closer to the mark this time, game developers could sure use the power.S|A

more here:
http://semiaccurate.com/2012/03/02/sony-playstation-4-will-be-an-x86-cpu-with-an-amd-gpu/

 

[TheChefO]

SemiAccurate has a new write-up

... "Once again, the end result comes down to the age old question of can they make it?" ...

more here:
http://semiaccurate.com/2012/03/02/sony-playstation-4-will-be-an-x86-cpu-with-an-amd-gpu/

I'd say that is the question. Can they make it? Within a reasonable timeframe, in reasonable quantities and for a reasonable cost that is.

From what little I've seen and heard of TSV's in production, I'd say no.

Unless they really are aiming for 2014ish and if that's the case, will the performance advantage (assumed) make up for the year of lost sales?

Would it not make better sense to simply take a 250mm2 Pitcairn die and slap a 120mm2 cellx2 die next to it on the same package with a fat bus?

Just a guess but I'd assume the above would be cheaper and easier for projecting manufacturing units and as a side bonus, they wouldn't have to launch late.

 

[Rangers]

Until Epic shows Samaritan running at that resolution on one AMD GPU, it doesn't fulfill it.

So Epic saying so in their own slide is wrong?

Edit: ahh you're on the AMD vs Nvidia thing. Well, with GCN AMD should be much closer to Nvidia even if it was true.

 

[ultragpu]

While we're still on the Samaritan demo, I dug up an interview with Tim Sweeney from Games magazine and some interesting numbers have been spilled by the man himself.
http://www.gamestm.co.uk/interviews...xt-gen-consoles-and-the-future-of-videogames/

games™: How would the next generation of games consoles have to be built to run the sort of technology you’ve demoed with Samaritan?

Tim Sweeney: We only need what Moore’s law will readily provide. Compared to current-generation consoles, I’d much like to see roughly 8-12x more CPU performance, 10x higher GPU triangle and rasterizer throughput, and 20x more GPU computational (ALU) throughput. With that degree of leap, I’m confident we can ship next-generation games with graphics on par with the Samaritan real-time demo.

Not sure if Pitcairn is up for the task.
Here's a concept art for that huge mecho boss at the end.
http://4.bp.blogspot.com/-lkqg7fdty...U7wdJPo1vc/s1600/Bullramfinalheadversion2.jpg

 

[V3]

"In there" where exactly? For decent backwards-compability it will need around 25GB/s to RAM + similar amount to GPU, pretty much the only way to get it is to integrate it either inside or VERY close to the CPU/GPU.

Maybe its own RAM and acts as a co-processor to the main CPU.

 

[upnorthsox]

Pitcairn specs that look real:

HD7870 would look tasty in a console, 2.56TF fulfills Epic's requirement for Samaritan at 1080P. Is about 10X Xenos flops which I think is a respectable generational leap.

7850 would definitely seem doable in a console powerwise, but I think 7870 could be too.

If you look at the 7870 as the full chip and 7850 as the salvage chip, then I'd say something in the middle. 4 CU's disabled would be alot/excessive for the consoles, 2 CU's disabled (so a total of 18) would be a little more reasonable. Maybe bring the clock down 800mhz and you're still close to the 130w.

For either/both, that would be a reasonable chip and about what I expected.

 

[Brimstone]

The PS4 sounds like it will be a beast. What Charlie is saying seems to make sense. Take two CPU/GPU chips and combine them later down the road. If the speculation is ballpark accurate the only question now is how much memory will it have?

Its also intresting to note that Sony has shifted their stereoscopic 3d display strategy to Film Patterned Retarder (FPR). Sony, Panasonic, and LG are pushing this method now. The LG OLED coming out later this year will also use this tech.

Can't wait for 2014 to play PS4 games in stereoscopic 3d on a OLED screen. It's going to be awesome.

 

[bgassassin]

I'd say somewhere between the 7850 and 7870 would be a good compromise.

That's essentially what I've been proposing, though not clocked at 1GHz obviously.

So Epic saying so in their own slide is wrong?

Edit: ahh you're on the AMD vs Nvidia thing. Well, with GCN AMD should be much closer to Nvidia even if it was true.

How could you take that as Epic being wrong? Also it doesn't have to do with an "AMD sucks" viewpoint, it's simply looking at the information they've given us.

Epic gave us a formula to work with for Samaritan. It's very dependent on the resolution. Based on the supposed resolution (2560x1440) of the original you're looking at 4.4 TFLOPs. Three 580s is 4.5 TFLOPs. So now we see why they needed three 580s to begin with. As a comparison, a 6970 is 2.7 TFLOPs. I think anyone who reads these boards know two 6970s don't come anywhere close to three 580s.

Looking at their slides, the amount of FLOPs comes down by reducing the resolution. They would still need two 580s to run the demo at 1080p. At the same time a 7970 is ~3.8 TFLOPs. Considering it didn't blow away one 580 in benchmarks, there's no reason to believe one Pitcairn could handle Samaritan at 1080p.

I don't believe the gains of a closed environment are going to be enough based on how I see the console GPUs to be. Now if UE4 is designed to make that more efficient, then yeah I could see 1080p reached. But based on the info Epic themselves have given us, I don't see it as possible.

 

[jlippo]

I don't believe the gains of a closed environment are going to be enough based on how I see the console GPUs to be. Now if UE4 is designed to make that more efficient, then yeah I could see 1080p reached. But based on the info Epic themselves have given us, I don't see it as possible.

UE4 will most likely change how they render things quite radically, so yes things might change a lot..

IE.
If they decouple shading from resolution 1080p shouldn't be a problem..

 

[Acert93]

Stacking allows for lots of inter-die bandwidth, I suspect this is what Sony is pursuing, not larger total die area. Stack the CPU, GPU and some DRAM for super high bandwidth, - and have some more DRAM on a "slow" XDR or GDDR5 bus.

Cheers

Based on the technological limits angling toward FINFETS on a process node (power reduction), stacked memory / TEVs / SIs (power, bandwidth) and trying to angle for "affordable" performance corners that current tech provides does seem the reasonable approach over extremely large chips with large power draws, lower yields, and a lot of heat (not to mention limited production quantities early in the life).

Pitcairn specs that look real:

HD7870 would look tasty in a console, 2.56TF fulfills Epic's requirement for Samaritan at 1080P. Is about 10X Xenos flops which I think is a respectable generational leap.

7850 would definitely seem doable in a console powerwise, but I think 7870 could be too.

Also consider the benefit of a 7870 class chip, slightly lower clocks (but ~ same functional units), and then a lower-powered memory solution (such as the one AMD was showing off with SI and stacked memory) and you have a pretty solid GPU that fits well into the console ecosystem.

Btw, imagine how much power, heat, space, noise, and cost could be saved by forgoing an optical drive...