Predict: The Next Generation Console Tech

[almighty]

Theoretically speaking,
Pixel fill is 8x RSX/Xenos (32ROPs @ 1GHz vs 8ROPs @500MHz).
16x Xenos tex fill, ~10.7x RSX tex fill ( 128@1GHz vs 16/24TMUs @500 MHz)
Z-only fill is 32x Xenos/RSX (256Z-only/clk@1GHz vs 16Z-only/clk @500MHz
Triangle setup rate is 8x (4 setup engines @ 1GHz vs 1 @500MHz)

And due the GTX 680 being bandwidth starved its likely less in reality.

Although triangle set up rate will be null due to tessellation

 

[pjbliverpool]

How so when it doesn't offer 10x more performance in every area?

It's fill and texel rate are not 10x that of the current consoles, it also doesn't offer 10x the performance in other areas.

Your failing to account for efficiency gains that 8 years of GPU development brings. A 14x shader thoughput improvement on paper is going to be more like 28x in reality. 16x texturing throughput more like 24x in reality etc...

So even though you may only be able to draw 10x the actual pixels on screen with a 680 compared with Xenos, pixel for pixel you have 25-30x the texture and shading budget. That to me counts as a 25-30x performance increase even if you can only draw 10x the pixels or setup and tesselate 10x the geometry.

And I don't buy the "680 is memory starved so won't hit those real world figures" argument. Nvidia would not build a core which was severely limited by it's memory interface. They'd either give it a bigger interface or cut down the core, to do anything else would be ridiculous, Why spend money building a bigger faster core when 50% of that performance will never see the light of day because of bandwidth limitations? Better to save money and build a smaller core. Besides, the same argument can be made for AMD's 7xxx GPU's. Or do you believe they are also severely performance constrained?

Anyway, to get back to the point of the thread and my original comment, just because a next generation consoles GPU may only have 5x the flops of Xenos on paper, doesn't mean you won't see a much bigger real world improvement. Even if it only has say 4x the fill rate, if it can shade the same number of pixels 10 times faster then that's a 10x performance increase in my book and that's what you'll see on screen.

 

[SedentaryJourney]

I know that in 2013 the European Commision will change the stand by power consumption law.

Currently house hold appliances have to have a standby power consumption of 2w or less, that is due to drop to 1w in 2013.

That could affect things in the new consoles

That shouldn't be an issue. Standby power is already under 1w.

 

[almighty]

And I don't buy the "680 is memory starved so won't hit those real world figures" argument. Nvidia would not build a core which was severely limited by it's memory interface..

Then I suggest you go and read reviews on the GTX 670/680 as most of them talked about the cards being bandwidth limited in certain games and situations.

Crysis and Metro 2033 being stand out titles that the cards choke on due to bandwidth issues.

And they performance hit they take in some games when AA is enabled is atrocious.

Meanwhile on the memory front we were only able to push the memory a further 400MHz to 6.4GHz. This is noticeably less than what we’ve been able to push reference-based GTX 670/680 cards (which have reached nearly 7GHz), and it’s not immediately clear why we’re seeing this difference. With both a custom PCB and additional memory chips it’s possible that either one could be holding back memory overclocking, or more likely it’s a combination of the two. Regardless of the reason this does put the GTX 680 Classified in a bit of a bind, since the GTX 680 is no stranger to memory bandwidth starvation.

It's a well known fact that Keplar, well the top end Keplar cards are bandwidth starved in quite a few games and situations. That's one of the reasons that I went with the 7970's that I have as I like high AA levels and I didn't want to have to turn it off or reduce it due to bandwidth limitations.

 

[AlNom]

And due the GTX 680 being bandwidth starved its likely less in reality.

Pixel rates... sure. Texture rates... are a different story as there are texture cache improvements/additions (especially since 7-8 years ago), which do mitigate external bandwidth requirements to an extent.

 

[almighty]

Pixel rates... sure. Texture rates... different story since you're ignoring texture cache improvements.

True, makes no difference anyway any the consoles aren't getting a GTX 680 :smile:

 

[bitsandbytes]

Isn't it likely that next gen games won't be pushing more pixels than current gen and therefore a big real world "gain"?

Looking at the PS2 to PS3 we went from God of War 2 pushing ~230k pixels to Uncharted pushing ~921k pixels.

At least it should help?

 

[almighty]

Isn't it likely that next gen games won't be pushing more pixels than current gen and therefore a big real world "gain"?

Looking at the PS2 to PS3 we went from God of War 2 pushing ~230k pixels to Uncharted pushing ~921k pixels.

At least it should help?

Entirely depends if developers stay with ~720p or aim for ~1080p

 

[pjbliverpool]

It's a well known fact that Keplar, well the top end Keplar cards are bandwidth starved in quite a few games and situations. That's one of the reasons that I went with the 7970's that I have as I like high AA levels and I didn't want to have to turn it off or reduce it due to bandwidth limitations.

The reviews are talking about a few percent worth of performance loss due to bandwidth limiations, not the "mutilples of xenos performance" we are discussing here. So what if you may only see 25x Xenos shader performance rather than 28x because the 680 could do with a little more bandwidth? The original point still stands.

And anyway, to get away from all this ridiculous AMD vs NV bull crap, forget I even mentioned the 680, take my original point and swap in the 7970 which I'm sure you don't have the same reservations about and is a better comparison point for the next generation consoles anyway.

The 7970 had 17.5x Xenos's shader throughput and near 15x its texturing throughput (both on paper) but in reality those numbers are both going to be closer to a 30x real world improvement. Hence if the next gen consoles get a GPU that is roughly 1/3 of the 7970 - ~1.2 TFLOPS or 5x Xenos on paper - the actual performance increase where it most matters will be more like 10x.

 

[almighty]

The 7970 had 17.5x Xenos's shader throughput and near 15x its texturing throughput (both on paper) but in reality those numbers are both going to be closer to a 30x real world improvement.

Either provide numbers and proof to back that 30x real world or stop pulling them out your rear end.

Modern GPU's are more efficient but to claim such numbers is extremely laughable.

 

[Bagel seed]

All this multiplier talk compared against super old consoles is a little misleading and hard to imagine. Why not just look at what desktop gpus are capable of now and start from there?

What I know is mainstream cards of today (1.5-2TF) can run all the latest games at decent settings and 1080p res. If new consoles end up with equivalents to those we'd probably see the same games runnable at high settings, and maybe engines like UE4 at med-high? That *feels* like the jump we had in 2005. Maybe even a bit more.

 

[arijoytunir]

hd 6770 is 1.36 TF and hd 7770 is 1.28 TF . The Flop count has reduced but the performance in the same price segment has increased from 10-30% . The epic samaritan demo ran on "3" gtx 580s but later it could run only on a single gtx 680. A gtx 660 matches a single gtx 580 performance today for half the price. So the R&D of the sony and microsoft can come up with gtx 680 like power in hd 8850 size and price . The flop count is not as important as some thinks .

 

[Inuhanyou]

I've heard everyone who purportedly knows something saying that, "flop counts are deceiving".

I wonder what other things people are not taking into account that will actually make the nextbox GPU competitive in the performance arena.

I say this coming from a background with 0 experience in HW configurations or architectural efficiencies per GPU generations

 

[almighty]

hd 6770 is 1.36 TF and hd 7770 is 1.28 TF . The Flop count has reduced but the performance in the same price segment has increased from 10-30% . The epic samaritan demo ran on "3" gtx 580s but later it could run only on a single gtx 680. A gtx 660 matches a single gtx 580 performance today for half the price. So the R&D of the sony and microsoft can come up with gtx 680 like power in hd 8850 size and price . The flop count is not as important as some thinks .

Epic greatly reduced the settings to get it running on a single 680.... lower resolution and no msaa to name but a few

 

[Inuhanyou]

^ Where did you hear the lower resolution talk? Although i did hear about FXAA.

On the other hand, i'd be fine with Agni's level, which is a much lower hardware baseline to shoot for.

 

[Ruskie]

I've heard everyone who purportedly knows something saying that, "flop counts are deceiving".

I wonder what other things people are not taking into account that will actually make the nextbox GPU competitive in the performance arena.

I say this coming from a background with 0 experience in HW configurations or architectural efficiencies per GPU generations

Well, there are pixel fillrate, texutre fillrate and bandwith which can (depending on workload, resolution and other things) hinder high theoretical FLOPS of GPU, but in nutshell. If Durango GPU is ~1TF it will be very limited shader wise, as it either has small number of CUs or its clocks are very very low.

I guess they would go for something like HD7770 in the rumored case and that would certainly mean new generation in ~5 years, not 8, because that card is already too slow for next gen engines.

 

[Bagel seed]

Ok so in theory console efficiency + new architecture efficiency + bandwidth + 1.5TF should give us that satisfactory leap we hope for...

But still, we're back to that squeezing blood out a stone situation again, just to keep up with mid to high end pc cards. I don't want to see that reversion to corridor gameplay to boost graphics like we had this gen.

 

[pjbliverpool]

Either provide numbers and proof to back that 30x real world or stop pulling them out your rear end.

Modern GPU's are more efficient but to claim such numbers is extremely laughable.

It's well know that architectures improve efficiency over time. There have been 7 architecture changes from Xenos to GCN. You think a doubling of efficiency over that many generations is laughable?

What low expectations you have. The SIMD setup of GCN is far more efficient than the old Vec4+1 of Xenos and I've heard around 2x quoted a few times on these forums. No I don't keep a record of every interesting quote I've ever come across so I can't just link back to them. Go and do some research of your own if you're that bothered.

For example look here:

http://www.anandtech.com/show/5261/amd-radeon-hd-7970-review/26

This shows ROP efficiency between a single generation (6xxx -> 7xxx) increasing by over 50%.

Anyway, we're straying too far from the thread topic here so assuming you still disagree, let's just agree to do so and leave it at that.

 

[almighty]

^ Where did you hear the lower resolution talk? Although i did hear about FXAA.

On the other hand, i'd be fine with Agni's level, which is a much lower hardware baseline to shoot for.

Agni's was still running on a single GTX 680 so I would expect a lower baseline for next generation consoles.

 

[Inuhanyou]

Agni's was still running on a single GTX 680 so I would expect a lower baseline for next generation consoles.

That's only a sliver of the equation.


Agni took a majority of its code(character models environments) directly from Visual works with no optimization according to the devs themselves, they just lifted them directly and made a tech demo out of them.

It was the power of the 680 that allowed this, however they always stated quite specifically, that when actually optimized(aka models and environmental assets built from the ground up for the hardware) they could replicate the same visual detail on much weaker hardware.

They don't even use global illumination for Luminous engine, it should not be too taxing on the hardware at all.

(Also it used 8XMSAA and FXAA together, so it obviously can be reduced significantly for consoles when optimizing)