Predict: The Next Generation Console Tech

[XpiderMX]

Alpha kit yes, beta no.

I think a lot of third party games started on the GTX 680.

And then, they downgrade the graphics params xD.

 

[AlphaWolf]

Obviously flops aren't a sole indicator, but why would they have a CPU in there that can do less floating points than their previous console?

Because it'll actually put a lot more of those flops to use.

 

[Proelite]

And then, they downgrade the graphics params xD.

Possibly. I think we'll see a lot of 720p 60fps games with 2xAA + FXAA next generation that look as good as their PC counterparts.

 

[nunomoreira]

Possibly. I think we'll see a lot of 720p 60fps games with 2xAA + FXAA next generation that look as good as their PC counterparts.

doubt they would ever go back to 720p
1080p should be the standard for the next consoles

 

[kalelovil]

Die size and TDP could both hinder that plan. A 7870 binned for 7970M is lower power than a normal 7870. They shave off 30W by reducing clocks and binning.

Die size shouldn't be a significant problem for a Pitcairn-level APU.

At 210mm2 it isn't much larger than Xenon, particularly considering Xenon had some of its core functionality outsourced to the EDRAM daughter die.

Two Jaguar quad-core CUs would probably add less than 50mm2 to this.
Even allowing for a margin of error, interconnect logic and miscellaneous parts such as a TrustZone, such a combination at 250-280mm2 would be quite manufacturable and cost-effective.

Moving up a notch, a hypothetical 24 CU 88xx-based GPU and 4 Jaguar CUs would likely still have around the same die size as Tahiti, but for the entire SoC.

Although I doubt we will see as many die shrinks as last generation, die size and thus cost will decrease over the lifetime of the consoles.

A 7870 binned for 7970M is lower power than a normal 7870. They shave off 30W by reducing clocks and binning.

That may have been true in April 2012, but by the time the next generation consoles launch AMD will have benefited from more than 18 months of refinements to their graphics technology and to 28nm process technology.

Admittedly where the chip will be manufactured and how the Common Platform Alliance's 28nm process compares to that of TSMC is a bit of a wildcard and could go either way.


By the time these consoles launch the HD 7850 will only be upper-midrange graphics technology.

 

[XpiderMX]

Possibly. I think we'll see a lot of 720p 60fps games with 2xAA + FXAA next generation that look as good as their PC counterparts.

2xAA+FXAA at 1080p is ok for me, but not 720p

 

[Bagel seed]

Whoo! Although I always preferred the GUS. And I built my own resistor DAC soundcard when I was poor.

Ahh, I only posted AWE so more people would get it. GUS was the best! I got the 1024k RAM upgrade for the MAX too. Such sweet memories of Playmidi and tracker/MOD stuff.

 

[ImaginaryInsider]

What exactly is he saying seemed possible? AMD guy's specs or Rangers specs with the 1+TF GPU?

It's right in the very quote you posted...

Somebody wanted a confirmation or dismissal on the possibility of 2.5 TF GPU, all Aegies said was based on the components, it's possible, but potentially too good to be true, and that he isn't sure about that aspect aka "so don't quote me".

He has shot down the 1.2 teraflop rumor based on what he's heard.

Take it all with a grain of salt, until we get document leaks newer than 2010.

 

[RedVi]

Thanks, isn't that a bit low then since Xenon apparently has 115 GFLOPS, is that wrong?
Or will x86 Jaguar cores with OoOE be that much more efficient than Xenon?

It just means Jaguar's peak FP performance isn't quite as good. Real world FP performance may be on par or better depending on on the front end, and integer performance will of course be multiple times better. Keep in mind also, with GPGPU, CPU FP performance should become a lot less important in the coming years - especially on a closed system using an APU/SoC.

 

[Heinrich4]

Die size shouldn't be a significant problem for a Pitcairn-level APU.

At 210mm2 it isn't much larger than Xenon, particularly considering Xenon had some of its core functionality outsourced to the EDRAM daughter die.

Two Jaguar quad-core CUs would probably add less than 50mm2 to this.
Even allowing for a margin of error, interconnect logic and miscellaneous parts such as a TrustZone, such a combination at 250-280mm2 would be quite manufacturable and cost-effective.

Moving up a notch, a hypothetical 24 CU 88xx-based GPU and 4 Jaguar CUs would likely still have around the same die size as Tahiti, but for the entire SoC.

Although I doubt we will see as many die shrinks as last generation, die size and thus cost will decrease over the lifetime of the consoles.




That may have been true in April 2012, but by the time the next generation consoles launch AMD will have benefited from more than 18 months of refinements to their graphics technology and to 28nm process technology.

Admittedly where the chip will be manufactured and how the Common Platform Alliance's 28nm process compares to that of TSMC is a bit of a wildcard and could go either way.


By the time these consoles launch the HD 7850 will only be upper-midrange graphics technology.

I fully agree, because as I said,in April 2012 was launched pitcairn 7970M 2.2 TFLOP (I'm based in wiki and this is wrong and 100 watts was correct) and refinements in 28nm process, clocks may see higher or significantly lower wattage ... see performance gains expected in 8850/Hainan under the same wattage and die slightly higher (7850 =212mm2=130watt=2.8billion trans.=1.76 Tflop -> 8850=230mm2=130watt=3.4billion trans.=2.99Tflop).

Maybe a 8850/Hainan "M" Like, could reach the same levels of efficiency gains in the same proportion of 7870 to 7970M.

 

[Rangers]

It's right in the very quote you posted...

Somebody wanted a confirmation or dismissal on the possibility of 2.5 TF GPU, all Aegies said was based on the components, it's possible, but potentially too good to be true, and that he isn't sure about that aspect aka "so don't quote me".

He has shot down the 1.2 teraflop rumor based on what he's heard.

Take it all with a grain of salt, until we get document leaks newer than 2010.

He's wrong.

He's a bit over his head.

 

[Rangers]

I can imagine what three blocks they're talking about, and at least two of them will help with graphics, although one is more general purpose. I have a special fondness for the third, since I spent many hours coming to terms with it's idiosyncracies.

So likely an audio chip of some sort and...

I have no idea at all on the others. Besides I guess the tesselation thing. Which doesnt make a ton of sense to me (I figure the built in tess capabilities in SI will be enough, given that super high tess factors dont seem to make much visual difference anyway, and Durango wont be powerful enough to take advantage of super super high end stuff anyway). But I'm just a lay-idiot anyway so could be wrong.

An audio chip seems odd too considering you've got 8 cores sitting around begging for stuff to do, but that's clearly one of them.

Edit: oh wait, I'm missing the ESRAM. Could well be block 2...

 

[XpiderMX]

He's wrong.

On which?

 

[bkilian]

Ahh, I only posted AWE so more people would get it. GUS was the best! I got the 1024k RAM upgrade for the MAX too. Such sweet memories of Playmidi and tracker/MOD stuff.

I actually recently wrote a S3M player that ran on a machine I no longer have access to, just for fun. Played Second Reality and Axel F pretty well. I used it for testing audio output on each code release

It just means Jaguar's peak FP performance isn't quite as good. Real world FP performance may be on par or better depending on on the front end, and integer performance will of course be multiple times better. Keep in mind also, with GPGPU, CPU FP performance should become a lot less important in the coming years - especially on a closed system using an APU/SoC.

Is that using 128 bit wide SIMD or 256 bit wide?

 

[tunafish]

Thanks, is that per core or entire CPU?

Entire CPU.

Xenon was 115 GFLOPS right?

Yes, but there is no hope in hell anyone has ever gotten anywhere near that on a real load.

The Jaguar cores are absurdly more efficient. When there was a lot of talk about jaguar cores possibly being in the next console, I went and bought myself a bobcat minilaptop to practise on the cpu. The more code I write for it, the more impressed I am of it.

This thing is crazy efficient. Without any optimization beyond what is done by the compilers, it gets near 1 IPC. Add the same level of effort we put into the consoles last time, and we are talking about something like 1.7 IPC. And that's x86 instructions that have memory read operands baked into them.

So while a xenon typically ran at something like 20% of it's capability (two threads at 0.2 IPC for a 2-wide core), this thing gets to 85% capability with one thread.

By these numbers a Bobcat thread would be something like 4 times faster than a Xenon thread. And that's for raw flops, which are the strongest point by far of the Xenon, and the weakest point of the bobcat. Simple integer stuff that's always needed like branching is just ridiculously faster on bobcat.

How does that compare to modern off the shelf CPU's i wonder?

In throughput, it's somewhere pretty close to a modern Intel dualcore.

 

[Rangers]

Rangers where are you getting the 10 CUs @ 1 Ghz rumour from?
The same source as the three custom GPU blocks?

I never posted that. Sheesh people, I'm now getting ripped for stuff I didn't even post.

Is 1Ghz is clocked a bit high for a console? I would think a few more CUs at a lower clock

Me too, but the 7770 @ 1ghz might actually be feasible. It's 80 watts TDP. The CPU's shouldn't add up to much, maybe 30 watts. You can probably then fit the whole console in 150 watts, 50 less than 360, which sounds about what I think MS would aim for (gotta be green and all that baloney ya know).

Theres passively cooled 7770's even, granted it's one giant heatsink http://www.tomshardware.com/news/sapphire-radeon-gpu-7770-heatsink,15927.html

I still think more CU's lower clock is more likely, but 1ghz is probably doable as is.

 

[bkilian]

An audio chip seems odd too considering you've got 8 cores sitting around begging for stuff to do, but that's clearly one of them.

Audio processing on the 360 would often use an entire core, it's not lightweight.

 

[Rangers]

Yes, but there is no hope in hell anyone has ever gotten anywhere near that on a real load.

The Jaguar cores are absurdly more efficient. When there was a lot of talk about jaguar cores possibly being in the next console, I went and bought myself a bobcat minilaptop to practise on the cpu. The more code I write for it, the more impressed I am of it.

This thing is crazy efficient. Without any optimization beyond what is done by the compilers, it gets near 1 IPC. Add the same level of effort we put into the consoles last time, and we are talking about something like 1.7 IPC. And that's x86 instructions that have memory read operands baked into them.

So while a xenon typically ran at something like 20% of it's capability (two threads at 0.2 IPC for a 2-wide core), this thing gets to 85% capability with one thread.

By these numbers a Bobcat thread would be something like 4 times faster than a Xenon thread. And that's for raw flops, which are the strongest point by far of the Xenon, and the weakest point of the bobcat. Simple integer stuff that's always needed like branching is just ridiculously faster on bobcat.

Very nice!

 

[Proelite]

Very nice!

Are you specialguy from gaf?

 

[(((interference)))]

Thanks for explaining that about the relative efficiencies of Jaguar vs Xenon.

It's right in the very quote you posted...

Somebody wanted a confirmation or dismissal on the possibility of 2.5 TF GPU, all Aegies said was based on the components, it's possible, but potentially too good to be true, and that he isn't sure about that aspect aka "so don't quote me".

He has shot down the 1.2 teraflop rumor based on what he's heard.

Take it all with a grain of salt, until we get document leaks newer than 2010.

Yeah, I don't think he has the info or technical nous to evaluate if it does, he obviously doesn't have a TF number for the GPU.
And since bkilian and lherre (who have first hand knowledge) both shot down the 2 or 2.5 TF GPU rumours we can be pretty sure it's not going to be that powerful in terms of pure FLOPS.