What are the Pros & Cons of having a PowerVR GPU in a Next Gen Console?

[Nesh]

Or you and the widest majority had too high hopes in general for next generation consoles.

I think people miss the good old days when a reveal of a new console came with tech that exceeded the performance of the current time or at least came close to meeting our expectations. Hardware prediction was less predictable before. Considering that we are expecting the consoles to be released in 2013 xmas, knowing that its based on old tech will sure disappoint

The original PS blew people away, the DC did it again. PS2 looked like a dream machine before its release. MGS2 and GT3 were beyond anything we have seen on any platform even though PCs were more powerful.

I guess we are hitting diminishing returns
The very few biggest PC GPU manufacturers lead the GPU industry more than they ever did a few decades ago with the most efficient solutions at any given time. Hence console manufacturers can no longer come up with their own practical state of the art exotic solutions anymore that may exceed what is available. They simply cant compete with the expertise and knowledge the GPU companies have now. Their timetable follows the path of those GPU companies who have already invested on the best technology they can come up with. And they have to do it if they want a developer friendly environment that can output good results fast.
I only hope they surprise us in a positive way and the current "specs" they use are simply placeholders :smile:

Oh the good old times when it wasnt just Intel/AMD/IBM and ATI/NVIDIA. We had also Hitachi's, Yamaha's, Toshiba's, Sony's, Lokheed Martin's etc making chips for our consoles and arcade machines

 

[function]

If profitability / affordability / sanity is restored to console making, and as PC users move to ultrabooks, it might give the arcades some breathing room again. And that could mean a Sega are reborn.

Dreamcast 3 on PVR series 8 with raytracing hardware. Yes.

 

[TheChefO]

If profitability / affordability / sanity is restored to console making...

FYI, Playstation released way back in 1995 sold for $299. At this price, Sony was losing $100/console.

By the time ps2 came out, the new Playstation division was making more money than any other at the $40Billion multinational corporation.

That's the business model.

Of course, this is assuming that such a profitable model is acceptable enough, or if Apple's yearly $500 product refresh and >30% profit margins are now the only viable way forward...


Having said that, I'd love to see a return of PVR in consoles. DC was a killer box for the time and the concepts behind PVR are truly innovative and proved their worth by providing more bang/buck in the real world ... though with each passing year, those advantages seem to dwindle.

 

[Shifty Geezer]

Having said that, I'd love to see a return of PVR in consoles. DC was a killer box for the time and the concepts behind PVR are truly innovative and proved their worth by providing more bang/buck in the real world ... though with each passing year, those advantages seem to dwindle.

Really? Deferred rendering is in vogue and TBDR is seeing a resurgance in mobiles. I'd say the advantages are becoming more competitive.

 

[TheWretched]

AMDs new time demo (Leo) is an attempt to show that forward renderers are still great. If I understood it correctly, it uses the compute shaders for light culling before the render pass is started.

 

[3dcgi]

AMDs new time demo (Leo) is an attempt to show that forward renderers are still great. If I understood it correctly, it uses the compute shaders for light culling before the render pass is started.

In this case deferred rending is different than TBDR as Imagination Tech uses. A deferred shading engine can run on non-deferred hardware and a forward renderer runs fine on a TBDR.

 

[Rys]

I'd say the advantages are becoming more competitive.

Agreed, since the biggest thing in the system that isn't scaling up quickly is precisely what PowerVR is designed to consume less of.

 

[steampoweredgod]

What surprises me is that mobile solutions can approach console graphics so close, yet from what I hear 2.5W are used by ipad2.

A next gen console could easily have 2XX~W, that's about 100 fold increase. If performance increased likewise by 100 fold over that[mobile], the looks possible would not only go above current consoles but would blow them out of the water.

 

[Shifty Geezer]

What surprises me is that mobile solutions can approach console graphics so close, yet from what I hear 2.5W are used by ipad2.

A next gen console could easily have 2XX~W, that's about 100 fold increase.

Power consumption increases exponentially with performance. 2x the performance might cost 4x the power, so 100x the power at 200W would net you maybe 7x the performance - that sort of thing. So if iPad2 is equivalent to PS360 at 2 watts, next-gen at 200W would be the natural 10x increase per generation. Sort of thing.

 

[Ailuros]

Power consumption increases exponentially with performance. 2x the performance might cost 4x the power, so 100x the power at 200W would net you maybe 7x the performance - that sort of thing. So if iPad2 is equivalent to PS360 at 2 watts, next-gen at 200W would be the natural 10x increase per generation. Sort of thing.

It would help to know the die area and power consumption of the Rogue in the future Novathor A9600 to get a closer estimate. It being a tablet/smart-phone design (smaller manufacturing processes and what not encounted) it doesn't sound like something that would burn a whole lot of Watts in the end. Judging by it's announced raw specifications you'd get compared to the MP2 in the iPad2 about 12x times the GFLOPs, less than 3x times the texel fill-rate and over 4x times the triangle rate. Besides the much higher featureset/capabilities it should be on estimate at least 4-5x times faster than a MP2, if you'd scale that kind of performance by another 5x fold I could be wrong but it doesn't sound like it would end up anywhere near of 200W for the graphics block by itself.

Definitely not a 100x fold increase in performance, but it should be at least 20x if die area between such a Rogue and power consumption is not too far apart from a MP2 in A5.

 

[steampoweredgod]

Power consumption increases exponentially with performance. 2x the performance might cost 4x the power, so 100x the power at 200W would net you maybe 7x the performance - that sort of thing. So if iPad2 is equivalent to PS360 at 2 watts, next-gen at 200W would be the natural 10x increase per generation. Sort of thing.

If one had infinite silicon budget, what would stop one from taking 100x ipad 2 soc chips, and dividing the screen in a hundred chunks and dedicating a chip to each? the ipad 2 has close to 720p resolution, and graphics tasks are ridiculously parallelizable. Each chunk of the screen would effectively get 100x more resources, and dealing with such low resolution(chunk res) massive supersampling could be applied by each chip, dealing perfect image quality.

 

[hoho]

the ipad 2 has close to 720p resolution, and graphics tasks are ridiculously parallelizable.

How would you handle triangles? Will each chip get to chew through all of them to figure out what needs to end up on the tile that GPU is supposed to render or will you have something else process them before having the GPU to actually shade the thing? What about memory bandwidth?

 

[steampoweredgod]

How would you handle triangles? Will each chip get to chew through all of them to figure out what needs to end up on the tile that GPU is supposed to render or will you have something else process them before having the GPU to actually shade the thing? What about memory bandwidth?

Assuming infinite budget, we can simply duplicate the ram too, so bandwidth would also go up too. As to how to allocate such, well renderfarms have been around since a very longtime when systems were even weaker than the ipad2, iirc, so it is possible... how to do it? you'd have to ask someone with knowledge of such.

People who've developed for 360 titles using tiling may've some ideas regarding issues that may arise.

 

[hoho]

Assuming infinite budget, we can simply duplicate the ram too, so bandwidth would also go up too.

Ok but what about offscreen rendering and synchronizing the results? I'd love to see the crossbar switch that synchronizes everything in that monstrosity

As to how to allocate such, well renderfarms have been around since a very longtime when systems were even weaker than the ipad2, iirc, so it is possible... how to do it?

Render farms do it very easily: each node takes a frame and renders it from start to finish. Would you like to see a 100 frame lag with that "GPU"?

 

[steampoweredgod]

Render farms do it very easily: each node takes a frame and renders it from start to finish. Would you like to see a 100 frame lag with that "GPU"?

I assume there must be some cases, even if special, when one frame might depend on the output of the previous frames. IF this does occur, then in some circumstances alternate approaches might be taken.

I'd love to see the crossbar switch that synchronizes everything in that monstrosity

Do gpus have crossbar like switches for their 100s of cores? Or are some graphics tasks less demanding of synchronization. I would assume as graphics are said to be ridiculously parallelizable, that since they gain vast performance as the resources scale in parallel, synchronization issues cannot be that big or else it would impede performance gains from ever greater parallelization.

EDIT:

Worse case scenario I can still envision cases wherein this could be simplified for realtime rendering. Games with fixed stylish camera angles, and cinematic controlled camera movements(Think oldschool resident evil, final fantasies, onimusha), would be able to easily tell what will fall on each part of the screen, as the camera and its movement is predetermined in advance.

 

[RudeCurve]

I think you guys are confused. The way PowerVR technology works is everything is done on chip so you don't need massive external bandwidth to "connect separate nodes" like in a render farm. All "nodes" in a PowerVR GPU communicates at full chip speed.

 

[hoho]

I think you guys are confused.

Considering how little I know about PowerVR tech you are quite definitely right. I just tried to apply what I know about regular AMD-NV GPUs to them

The way PowerVR technology works is everything is done on chip so you don't need massive external bandwidth to "connect separate nodes" like in a render farm. All "nodes" in a PowerVR GPU communicates at full chip speed.

So how many of those "cores" or "nodes" can you add to the full GPU before you hit some nasty bottleneck? How much work will be duplicated in those nodes (triangle setup?)?

 

[Simon F]

I think you guys are confused. The way PowerVR technology works is everything is done on chip so you don't need massive external bandwidth to "connect separate nodes" like in a render farm. All "nodes" in a PowerVR GPU communicates at full chip speed.

Interestingly enough,

• the very first (non-FPGA) PowerVR system used multiple chips (1+ ISP(s), 1 TSP)
• we did experiment with a PC system with a pair of PCX1/2 boards, and
• there was the Naomi-II system that had the T&L "Elan" chip driving multiple CLX2 renderers

but, yes the vast majority are single chip systems.

 

[upnorthsox]

Interestingly enough,

• the very first (non-FPGA) PowerVR system used multiple chips (1+ ISP(s), 1 TSP)
• we did experiment with a PC system with a pair of PCX1/2 boards, and
• there was the Naomi-II system that had the T&L "Elan" chip driving multiple CLX2 renderers

but, yes the vast majority are single chip systems.

So then a split chip (say 2 quads per split with a total of 16 cores) should be a real possibility then. What are the downsides?

 

[Rodéric]

You have some documentation for PowerVR:

http://www.imgtec.com/powervr/insid...cture guide for developers.1.0.8.External.pdf