Sony's Next Generation Portable unveiling - PSP2 in disguise

[Deleted member 13524]

You have to factor one very important element while doing this type of comparisons. The Windows drivers for Intel integrated graphics are horrible, not just bad.

Yes, the SGX535's windows drivers seem to be cutting its potential performance in half, based on numbers from other platforms.

And I honestly believe the NGP may be faster than C-50 at 3D rendering (in an utopic environment where the OS's cpu/memory overhead is neglegible, driver optimization is ideal, resolution and quality settings are the same). At least if DX11-specific performance optimizations aren't being used.

What shocked me was NathansFortune's bold statement:

NGP is more powerful than any Fusion product currently on the market and draws less power at peak. x86 is wholly unsuited to embedded portable products.

Any Fusion product involves E-350. That's implying the NGP is "more powerfull than" a dual 1.6GHz Bobcat + 500MHz Robson + 8.5-12.8GB/s bandwidth UMA. Not to mention it's a system that'll have, at least, 2GB of total RAM memory.
Unless the A9's are clocked way higher than 1GHz and\or the SGX543MP4 is clocked way higher than 200MHz, I have a really hard time believing it.

 

[DavidC]

You have to factor one very important element while doing this type of comparisons. The Windows drivers for Intel integrated graphics are horrible, not just bad.

What makes you think drivers would make 3DMark06 have 2x difference? For their desktop/laptop graphics, nearly all the things that was suspected for having low performance due to drivers turned out to be hardware issues. The drivers did improve for the GMA 500, but 3Dmark scores never changed.

You can't compare two radically different GPU architectures by macro-level like shader count, TMUs, and clock.

 

[NathansFortune]

Yes, the SGX535's windows drivers seem to be cutting its potential performance in half, based on numbers from other platforms.

And I honestly believe the NGP may be faster than C-50 at 3D rendering (in an utopic environment where the OS's cpu/memory overhead is neglegible, driver optimization is ideal, resolution and quality settings are the same). At least if DX11-specific performance optimizations aren't being used.

What shocked me was NathansFortune's bold statement:


Any Fusion product involves E-350. That's implying the NGP is "more powerfull than" a dual 1.6GHz Bobcat + 500MHz Robson + 8.5-12.8GB/s bandwidth UMA. Not to mention it's a system that'll have, at least, 2GB of total RAM memory.
Unless the A9's are clocked way higher than 1GHz and\or the SGX543MP4 is clocked way higher than 200MHz, I have a really hard time believing it.

Sorry, I should have clarified. I was referring to 9W Fusion chips. The 18W chips are well out of range for embedded products and aren't really in the same category. The rumoured 5W chip will match the power ratings (nearly), but it will not come very close to the SoC in NGP in terms of compute power and graphical ability.

If AMD do start making ARM SoCs then they will be able to compete, but I don't see what they can offer that Nvidia don't already with Tegra...

 

[Deleted member 13524]

Sorry, I should have clarified. I was referring to 9W Fusion chips. The 18W chips are well out of range for embedded products and aren't really in the same category. The rumoured 5W chip will match the power ratings (nearly), but it will not come very close to the SoC in NGP in terms of compute power and graphical ability.

That makes a lot more sense, thanks.
BTW, 5W is close? So there must be a heatsink in there! Wow..

If AMD do start making ARM SoCs then they will be able to compete, but I don't see what they can offer that Nvidia don't already with Tegra...

IMO this is.. subjective.
Power consumption is highly dependable of the chip's fabric process, and so far Intel has shown an unbeatable execution time at process transitions.. AMD may pair with Intel with that when GF goes full throttle.
ARM SoCs have always been late(r) to the party in process transitions because of the large number of intermediaries between architecture/instruction-set design and mass production.
Even if the ARM architectures are evidently superior for lower power devices, I can imagine the smartphone market having x86 SoCs using newer processes and ARM SoCs using older processes, both sharing similar power envelopes.

Besides, the Bobcat cores are showing an unpredictibly great performance for low-power devices, and we don't know if/how AMD can change it to get even higher performance/power ratio.

 

[NathansFortune]

The fusion chip would still require a lot of extras that would take it up well above that level while the whole of the NGP setup consumes less power at peak (minus the screen of course).

 

[Deleted member 13524]

The fusion chip would still require a lot of extras that would take it up well above that level while the whole of the NGP setup consumes less power at peak (minus the screen of course).

I was mostly talking about 2H-2012->2014, where NGP won't be exactly state-of-the-art anymore.

 

[Ailuros]

At least if DX11-specific performance optimizations aren't being used.

Which would be (honest question)?

 

[Shifty Geezer]

There already is FIFA for iOS, priced just under $10 for iPhone and just over $10 for iPad. Not sure that it sold too well at those prices.

Would you want to play FIFA on iPad? I wouldn't. It already maxes out utilisation of a dual stick, dual shoulder-button controller. I think that's the sort of line hardcore gamers who want to game 'properly' and not have watered down designs to fit buttonless handhelds will follow, and where NGP will be able to hold its own for a couple of years regardless of improvements to handheld tech. Even if 6th generation iPod Touch has twice the performance of NGP, it won't displace NGP as the better platform for core gaming. It's highly implausible Apple will sully their clean design with the lumps and bumps of buttons and sticks that are so vital to core gaming.

 

[Npl]

That makes a lot more sense, thanks.
BTW, 5W is close? So there must be a heatsink in there! Wow..

Adding a heatsink wouldnt do any good, if there is no ample air and much less circulation.
Would be much better to attach a thermal pad or heatpipe to the backside, but I dunno where it should be attached to? Metallic backplate ?

I doubt anything above 2W total power is infeasible for a PSP2 form-factor. Probably launching with 28nm is planned because of that reason (diearea isnt much of a concern at 45nm, I red 35 mm² for the GPU and CPU should be < 15 mm²). ARM designs are pretty much the testbed for 28/32nm at many foundries (Clicky here) so might be there is enough confidence to get everything ready for 28nm

 

[Ailuros]

35mm2@45nm for the GPU? I would had estimated it around 25mm2 at most under 45nm (given that IMG claimed 32mm2@65nm/200MHz for a SGX543 MP4) and I have a hard time understanding that the "+"stuff after the MP, amounts for a total of ~10mm2.

 

[Npl]

35mm2@45nm for the GPU? I would had estimated it around 25mm2 at most under 45nm (given that IMG claimed 32mm2@65nm/200MHz for a SGX543 MP4) and I have a hard time understanding that the "+"stuff after the MP, amounts for a total of ~10mm2.

Dunno, wish I could find the post/article Im refering to, maybe its the whole SOC or some additional on-die parts (edram for framebuffer?) - we still dont know everything about it.

My point is that its a rather small diearea at 40/45nm even, while going 28nm might be necessary for reducing powerconsumption.

 

[Deleted member 13524]

Which would be (honest question)?

AFAIK, tesselation, new texture compression algorythms and SM5.0..

Tesselation is used in PC games to enhance the base models to look as higher-polygon ones, but it could just use lower polygon models from the start to achieve the same visuals as higher polygon models, increasing performance.
Of course, no one would do this in the PC, because then the base models would look worse to all the DX9, DX10 and DX10.1 owners, and that would be slapping some 90% of PC gamers (according to the latest Steam survey).
However, in a closed platform, I can see tesselation as a performance enhancer. X360 ended up using it rarely because it would make the ports to PC/PS3 rather difficult, unfortunately (although the tesselator in Xenos/R600/R700 isn't as capable as the DX11 one).

 

[Ailuros]

AFAIK, tesselation, new texture compression algorythms and SM5.0..

Tesselation is used in PC games to enhance the base models to look as higher-polygon ones, but it could just use lower polygon models from the start to achieve the same visuals as higher polygon models, increasing performance.

SGX supports tessellation but with an unknown level of programmability. I wouldn't suggest its on DX11 level but all that is besides the point. I don't think SONY would make use of any of it irrelevant of the level of programmability of the tessellation pipeline.

I know very well what tessellation is used for in PC games. Now try firing up a DX11 game on a low end Fusion APU and it'll crap out with single digit frame rates before it'll even come to generating added vertices under all the workload. Uhmm wait you get those even in 1024*768 with low details in DX9/10 games on a 500MHz E-350: http://www.anandtech.com/show/4134/the-brazos-review-amds-e350-supplants-ion-for-miniitx/5

I'm anything but against technical advancements rather the contrary, but for a handheld console today and for its life time I consider full DX11 compliance overkill because it'll simply capture too much die area in terms of required capabilities which then would inevitably would sacrifice performance due to power constraints. Would you need for instance 8096*8096 texture sizes on a handheld? Even a TMU supporting all the DX11 requirements can blow up in die area significantly compared to the NGP TMUs today.

Of course, no one would do this in the PC, because then the base models would look worse to all the DX9, DX10 and DX10.1 owners, and that would be slapping some 90% of PC gamers (according to the latest Steam survey).
However, in a closed platform, I can see tesselation as a performance enhancer. X360 ended up using it rarely because it would make the ports to PC/PS3 rather difficult, unfortunately (although the tesselator in Xenos/R600/R700 isn't as capable as the DX11 one).

Xenos and it's less flexible than DX11 tessellation pipeline (no hull/domain shader stages amongst others) would had been the next point I would had made. What makes you think SONY will opt for tessellation on a next generation hand-held when it serves more than anything to port their wealth of PS3 games down to hand-helds?

Here's what you'd find in the Intel GMA500 manual:

Procedural Geometry

• Allows generation of more primitives on output compared with input data
• Effective geometry compression
• High order surface support

...and here a few bits out of the VGX150 whitepaper (vector graphics core) which is the smallest graphics core for low end phones with vector graphics:

Dynamic tessellation of Béziers is handled in hardware and deferred
until the rasterisation phase to reduce bandwidth.

The Geometry Engine manages all geometry and tiling functions, and uses a custom instruction set targeting geometry processing (transformation, stroking, parameter generation), culling, clipping and tiling of poly-lines and Béziers. The Geometry Engine is a fully programmable multi-threaded processor based on Imagination’s META architecture for maximum flexibility.

I still don't think anyone would ever make any use of any such capability during its lifetime. I'd even dare to speculate that it got stuck in their architecture ever since the scratched vaporware Series5 GPU and it would had been more complicated to remove it than to leave it in there. ATI would tell you something similar for R600 too, due to the endless "flip-flops" before DX10 got final (tessellation? uhhmm yeah ok but let's remove that bit and oh the other bit too...but now it's a fixed function unit!...ok let's make it optional...well better remove it in the end and go for full programmability in DX11...)

As for texture compression many developers especially in the console space would love to have a 2bpp algorithm like PVR-TC instead of the typical 4bpp algorithms.

http://arstechnica.com/gaming/news/2010/11/post-8.ars/2

Yes. The HD version of Rage is 1.4GB installed, and all the world geometry is using 2-bit PowerVR texture compression. If we went to one of the other platforms that's not PowerVR-based, we'd be stuck with a 4-bit texture compression format, and that pushes the size over 2GB.

 

[Mobius1aic]

so where does PS3 level graphics start?

Even with an off screen capture, I can see that the textures are certainly more "pixely" than the first Uncharted title, and shadows don't seem as "crisp" (that's not necessarily a bad thing), so it's possible they are running fewer shadow samples that make them look softer at least in the off screen capture. The environment they show also is very limited and scope, so the geometry is very highly controlled with the exception of the end of the demo with the waterfall, which did look very good btw. I also enjoyed how the gyro was used as an aiming device too.

Oh and at 960 x 540 resolution (518,400 pixels), the NGP is rendering an image only 56.25% the size of a 1280 x 720 image which amounts to 921600 pixels. If the NGP could match the 720p output resolution, while pumping out the exact same visuals at the same framerate as a PS3, then it would be "equal" to a PS3. Running that MGS4 demo shown off at less than 20 FPS and at the 960 x 540 resolution, in contrary to fanboy antics, is not the same performance as the PS3. It's still very impressive for a handheld, but not on par with the PS3.

 

[Ailuros]

It's still very impressive for a handheld, but not on par with the PS3.

In terms of performance it definitely isn't (not just the GPU but the entire system aspects) and I don't have a clue how rumors concerning that one even made it to the media.

 

[GZ007]

I know very well what tessellation is used for in PC games. Now try firing up a DX11 game on a low end Fusion APU and it'll crap out with single digit frame rates before it'll even come to generating added vertices under all the workload. Uhmm wait you get those even in 1024*768 with low details in DX9/10 games on a 500MHz E-350: http://www.anandtech.com/show/4134/the-brazos-review-amds-e350-supplants-ion-for-miniitx/5

It doesnt make much sense to compare a PC DX11 game and a for example a ipad game. Fusion is heavily limited by CPU.
A ipad game directly ported to fusion 1to1 would end in 3 digit fps numbers. And In those cases it could use tesselation without problems.

Tesselation could be also much more usefull on mobile platforms where they need to minimize memory comunication and save power. At a given point of geometry complexity the invested die area could come back.

And using tesselation on a character that looks worst than in a PS1 game would have much bigger impact than in a PC game, where the geometry is limited mainly by the console ports and could be much higher without tesselation.

 

[Deleted member 13524]

I know very well what tessellation is used for in PC games. Now try firing up a DX11 game on a low end Fusion APU and it'll crap out with single digit frame rates before it'll even come to generating added vertices under all the workload. Uhmm wait you get those even in 1024*768 with low details in DX9/10 games on a 500MHz E-350: http://www.anandtech.com/show/4134/the-brazos-review-amds-e350-supplants-ion-for-miniitx/5

Are you really comparing coded-for-thousands-of-possible-system-combinations windows games to a fixed platform that can be optimized to the metal?
Would you like to run the very same benchmarks with a G71 @ 500MHz and i.e. an i7-970 (no CPU bottleneck) to see what happens? You may find some surprises there.

I'm anything but against technical advancements rather the contrary, but for a handheld console today and for its life time I consider full DX11 compliance overkill because it'll simply capture too much die area in terms of required capabilities which then would inevitably would sacrifice performance due to power constraints. Would you need for instance 8096*8096 texture sizes on a handheld? Even a TMU supporting all the DX11 requirements can blow up in die area significantly compared to the NGP TMUs today.

Xenos and it's less flexible than DX11 tessellation pipeline (no hull/domain shader stages amongst others) would had been the next point I would had made. What makes you think SONY will opt for tessellation on a next generation hand-held when it serves more than anything to port their wealth of PS3 games down to hand-helds?

Well, I never said a Fusion would be better for NGP, or that DX11 should be mandatory for a handheld console... Sure, some functions seem an exageration for a small screen, but I don't believe it would all be just a waste of transistors, nonetheless.

Here's what you'd find in the Intel GMA500 manual:

...and here a few bits out of the VGX150 whitepaper (vector graphics core) which is the smallest graphics core for low end phones with vector graphics:

IMG never made any marketing-related notifications about any tesselation ability, AFAIK.. Sure, the capability could be there, but IMG themselves must be wary of its performance and maybe that's why they don't make a big fuss about it.

I.E. the PICA200 has a tesselator, it's been marketed and I have little doubts it'll be used for 3DS games.

 

[Ailuros]

Are you really comparing coded-for-thousands-of-possible-system-combinations windows games to a fixed platform that can be optimized to the metal?
Would you like to run the very same benchmarks with a G71 @ 500MHz and i.e. an i7-970 (no CPU bottleneck) to see what happens? You may find some surprises there.

No I am not, but....

Well, I never said a Fusion would be better for NGP, or that DX11 should be mandatory for a handheld console... Sure, some functions seem an exageration for a small screen, but I don't believe it would all be just a waste of transistors, nonetheless.

When you're as die area (and in extension power) constricted as in a mobile/embedded platform, there will be inevitably critical design choices/dilemma for what exactly you dedicate each mm2 for. Sure would it be nice to have today the same performance/mm2/mW for a DX11 core vs. say a >DX9.0 core, but you can't have it all.

IMG never made any marketing-related notifications about any tesselation ability, AFAIK.. Sure, the capability could be there, but IMG themselves must be wary of its performance and maybe that's why they don't make a big fuss about it.

They don't make a lot of fuss about many aspects of their architectures these days on purpose. They used to make performance estimates with die area captured for different IP cores which has all vanished from their most recent whitepapers. Their 3 lines about procedural geometry in the SGX capabilities used by their partners are as sparse as other functionalities of the core itself are.

I.E. the PICA200 has a tesselator, it's been marketed and I have little doubts it'll be used for 3DS games.

I don't know the specifics of the PICA200, but it sounds more like what IMG already had in the optional vertex geometry processor for MBX http://www.imgtec.com/corporate/newsdetail.asp?NewsID=268

http://www.khronos.org/developers/library/2007_siggraph_bof_openvg/OpenVG%20BOF%205%20-%20Imagination%20Technologies.pdf

I wonder why SGX is so much more efficient with OpenVG, for which though in terms of power consumption it would be far more reasonable for OpenVG to implement a VGX150 along SGX. Single cycle 8xMSAA in the VGX and via multi-pass up to 24xMSAA, since some can't obviously get enough AA for OpenVG.

Tesselation could be also much more usefull on mobile platforms where they need to minimize memory comunication and save power. At a given point of geometry complexity the invested die area could come back.

And using tesselation on a character that looks worst than in a PS1 game would have much bigger impact than in a PC game, where the geometry is limited mainly by the console ports and could be much higher without tesselation.

Geometry related optimisations are all but new to the embedded space. CryTek's Polybump/viaDOT3 bm was an excellent example in that regard, which had been licensed by several vendors in the past, apart from HOS support mentioned above and any other forms of real tessellation. There are ways of getting there without having to opt for a rather expensive overall (due to a multitude of requirements) full DX11 pipeline at least for this past generation. Future embedded GPU generations are obviously an entire chapter of their own, but so are future way smaller manufacturing processes than those available today. There IHVs can feature creep their hearts out with programmable tessellation, double precision, atomic operations and what not.

 

[onQ]

Even with an off screen capture, I can see that the textures are certainly more "pixely" than the first Uncharted title, and shadows don't seem as "crisp" (that's not necessarily a bad thing), so it's possible they are running fewer shadow samples that make them look softer at least in the off screen capture. The environment they show also is very limited and scope, so the geometry is very highly controlled with the exception of the end of the demo with the waterfall, which did look very good btw. I also enjoyed how the gyro was used as an aiming device too.

Oh and at 960 x 540 resolution (518,400 pixels), the NGP is rendering an image only 56.25% the size of a 1280 x 720 image which amounts to 921600 pixels. If the NGP could match the 720p output resolution, while pumping out the exact same visuals at the same framerate as a PS3, then it would be "equal" to a PS3. Running that MGS4 demo shown off at less than 20 FPS and at the 960 x 540 resolution, in contrary to fanboy antics, is not the same performance as the PS3. It's still very impressive for a handheld, but not on par with the PS3.

I think people need to stop & think about what's being said instead of trying to prove the logic wrong,

people said that it was "PS3 level graphics on a handheld" why would it have to push out 720P on a handheld to be PS3 level graphics on a handheld?

I'm looking at a game on a handheld that look better than some games on the Xbox 360 & PS3 & you're telling me about "shadows that don't seem as crisp" ?

are you saying that all the games that was on the PS3 before Uncharted was released wasn't PS3 level games? in fact even today there is still PS3 games being made that's 960 x 540 & they are not made for a 5 inch handheld.

so again I ask where does PS3 level graphics start?

 

[Shifty Geezer]

I think people need to stop & think about what's being said instead of trying to prove the logic wrong...

There were mixed messages and huge ambiguities. No-one really knows what anyone meant. there's no point trying to justify them as legitimate claims as there is purpose to wielding such remarks as damned lies for which Sony should be punished. It's just the normal pre-release excitement that we hear.

Irrespective of what rumours and hope people may have had, we now have real specs to talk about, so we can forget all that. It's not like anyone has been bamboozled into buying an NGP on account of hearing it is as powerful as PS3!