Info on embedded GPU ?

[bonfo]

Hi,
i'm working on my final thesis and its objective is to analyse embedded GPU looking on their energy consumption. For this thesis i have the possibility to work on the PowerVR MBXLite GPU (embedded in the Freescale iMX31 SoC).

It's a while that i'm surfing the net, but i can't find some information . So i'm trying to write here

- I'd like to know a little bit more on the history of the PowerVR products and some information on why their Tile-Based-Rendering lost on the GPU for PC but it's really good for embedded system;
- Some information on other embeded GPU and where i can find some information on them;
- Really good points or bad points on the MBX rendering pipeline;
- Good benchmarking system for GPU Power Consumption.

Of course i don't want you to make the work.... I only ask for some good links or key information

Thanks to all

 

[Captain Chickenpants]

You should be able to get a factsheet from the imgtec web site(www.imgtec.com for power stuff. If you look in the article section here on beyond 3d their are explanations of the PowerVR architecture. Additionally if you register at the pvrdev web-site there should be documents covering information about our architecture.

If you do a search of the forums here you will find quite a few discussions about PVR and the desktop market.

I don't really know about where to find specific info on other embedded GPU's but a search of the forums here is probably a good start.

Power consumption is a tricky subject as it is dependant on the process used, clock frequency. Are you interested in power management strategies e.g turning off parts of the chip not in use.
I think to get this kind of information you will have to request the information from the companies in question.

CC

 

[Lazy8s]

http://en.wikipedia.org/wiki/Powervr outlines some of the history of PowerVR. Their processors have always been competitively fast in benchmarks, so the product strategy of the chip and graphics card companies who produced their IP, like NEC for Series 2 and STMicroelectronics for Series 3, is more likely at fault for PowerVR's failure in the desktop PC market than the technology.

Standardization for compatibility on the PC shapes PC APIs and applications, so the unconventionality of PowerVR's TBDR architecture is less of a liability in an embedded system than on a standardized platform like the PC.

 

[jkemp]

Standardization for compatibility on the PC shapes PC APIs and applications, so the unconventionality of PowerVR's TBDR architecture is less of a liability in an embedded system than on a standardized platform like the PC.

Well we at Nokia are all about standardized platform. For us, MBX was from that perspective no different than any other solution available - OpenGL ES does pretty good job in hiding the implementation architecture from the application itself.

Strong and weak points of TBDR are obvious (google around, I won't repeat 'em here) but as a guideline, performance on embedded is all about memory bandwidth. To start with, you typically have much less theoretical bandwidth available because of the used memory and bus technologies. On top of that, in integrated environment all the functional blocks (cpu, modem, wlan, audio, video, graphics, gps, mmcio, camera, etc) fighting for the same memory bandwidth. If you have a smart architecture/algorithm to reduce the bandwidth requirements your design has good chances of succeeding.

 

[Lazy8s]

Standardization isn't the difference between PowerVR's performance in a PC or an embedded system -- my mistake.

PowerVR's larger marketshare in the embedded sector is what makes software more likely to be tuned for it there than on the PC. Developers would be more careful to avoid redundant practices for TBDRs such as Z-prepass.

 

[bonfo]

Thanks for the hints

So you think that the absence of PowerVR in the PC market is indipendent form their technology and depends more on their strategically choices ??

About power consumption...
..from the point where i am i'd like to find a system to track the behaviour of the MBX. So track clocks, times, voltage and current.
To do that i need a benchmarking system that plays 3d/2d scenes with a complete range of possible workload combination and that can repeat the scene.


I've already checked for other embedded GPU, but seems the only other one used is the NVIDIA GoForce 5500, and, for what i've seen, is not a SoC GPU.

 

[Lazy8s]

Some other lines of portable 3D processors:

Z3D -- Not sure whether they were cores for SoCs or discrete chips. Developed by Mitsubishi who later transferred their semiconductor business into a new company, Renesas. Z3D was used in some early phones.

GShark -- Not sure whether they're cores or chips. Developed by Takumi.

Glamo -- Not sure whether cores or chips. Developed by SMedia.

Acceleon -- Cores for SoCs. Developed by BitBoys who were bought by ATi, and are now the business of AMD/ATi.

TxG (T4G/T5G) -- Not sure whether chips or cores. Developed by Toshiba and used in several phones from Toshiba and Sharp.

F3D -- Not sure whether chips or cores. Developed by VirtualDigm in F3D1000, and developed by UChips in F3D2000 and beyond. Both companies' processors were used in phones.

Imageon -- They come as discrete chips and also SoC cores. They're from AMD/ATi and are now being integrated as cores into Qualcomm's MSM chipsets.

Mali -- Cores for SoCs. Developed by Falanx who were bought by ARM and will be offered in a line of next generation cores later this year. They use a hybrid tile based and immediate mode rendering approach and therefore make an interesting comparison to PowerVR.

Pica -- Cores for SoCs. Developed by DMP (Digital Media Professionals).

MBX Lite was produced into a discrete chip also, the 2700G, formerly from Intel and now owned by Marvell.

The Mobile Data Visualization lab's website has a listing for some portable devices with hardware 3D:

http://mobile.sdsc.edu/index.html

Imagination Technologies' business model of licensing PowerVR IP cores instead of selling the chips themselves makes them dependent upon other companies for even getting to market, so a lack of partner wanting to compete in the PC sector would be a more immediate obstacle than the quality of their technology for returning to the desktop.

 

[bonfo]

Some other lines of portable 3D processors: ...

what a wonderful list !!! Thanks

Imagination Technologies' business model of licensing PowerVR IP cores instead of selling the chips themselves makes them dependent upon other companies for even getting to market, so a lack of partner wanting to compete in the PC sector would be a more immediate obstacle than the quality of their technology for returning to the desktop.

Yes, of course.
But i'd like to know why they leave the desktop market. Or better, i'd like to know if there is a technical reason that put them out of the Desktop market.

Could be the Tiled Based Rendering still a good technology fo the desktop market? Does Nvidia or Ati use it ??

 

[Arun]

Could be the Tiled Based Rendering still a good technology fo the desktop market? Does Nvidia or Ati use it ??

Do not confused tiled based rendering and tile based deferred rendering. The former is used by nearly every desktop GPU out there; the latter isn't used by a single one. All TBR means is that you render things in 8x8 or 16x16 tiles when you can, so that you can burst things to memory with higher efficiency - a TBDR's benefits (and complications) are an order of magnitude higher than that, in theory.


Uttar

 

[bonfo]

ehm...
...ok. I think i've confused the TBR and TBDR.

Thanks for clarifying me that they are different, but now i have to understand what the word "deferred" implies.
It means "only" that the HSR is applied before than usually and so only the visible pixelsi are rendered?

It's easy to understand how much this improve performance, but for me is not so easy to understand why is so difficult to implment? What are the differences between a not-deferred one?

Thanks

 

[dharold]

And of course if you have any particular questions you'd like us to answer, forward them to me.

David

Head of PR
Imagination Technologies
david.harold @ imgtec . com

 

[bonfo]

And of course if you have any particular questions you'd like us to answer, forward them to me.

David

Head of PR
Imagination Technologies
david.harold @ imgtec . com

Thanks, really.
But i have already send some questions... and the answers as well