PowerVR MBX Futuremark 3DMarkMobile06 Certified Results

[Kristof]

Official certified Futuremark 3DMarkMobile06 results are now available for the Texas Instruments OMAP2420 processor incorporating PowerVR MBX with VGP for both Linux and Symbian OS. Full details:

http://www.imgtec.com/News/Release/index.asp?ID=609


Certified Linux Configuration

Texas Instruments OMAP2420 Software Development Platform with PowerVR MBX graphics core and VGP co-processor.
OS: Linux Montavista (2.4 Kernel) running on ROM.
CPU clock 266MHz
Bus speed 133MHz
Graphics core clock 66MHz
Used optimizations compared to the 3DMarkMobile06 default version:

• PVRTC texture compression used
• glMultiDrawElements used

Game Test 1 Frames per second 37.30 fps
Game Test 2 Frames per second31.79 fps
Fillrate with 1 texture units Texels per second 36.09 Mpps
Fillrate with 2 texture units Texels per second 36.03 Mpps
Triangle count without lighting Triangles per second 2.06 Mtps
Triangle count with lighting Triangles per second 1.03 Mtps

Certified Symbian Configuration

Texas Instruments OMAP2420 Software Development Platform with PowerVR MBX graphics core and VGP co-processor.
OS: Symbian 9.1 2005 week 42 textshell running on RAM
CPU clock 300MHz
Bus speed 100MHz
Graphics core clock 50MHz
Used optimizations compared to the 3DMarkMobile06 default version:

• PVRTC texture compression used
• glMultiDrawElements used

Game Test 1 Frames per second: 37.76 fps
Game Test 2 Frames per second: 28.68 fps
Fillrate with 1 texture units Texels per second 33.73 Mpps
Fillrate with 2 texture units Texels per second 33.56 Mpps
Triangle count without lighting Triangles per second 2.36 Mtps
Triangle count with lighting Triangles per second 0.79 Mtps

Link to FutureMark results page: http://www.futuremark.com/bdp/certified/

 

[mboeller]

Are this MBX-lite results?

The 0.54 - 0.66 Pixel/clock ratio seems to be a little bit odd. Have you implemented only a half pipeline in the MBX? Also the results with the 2 texture-units seems a little bit strange. Do you really mean texture-units or texture layers?

What was the resolution of the benchmark? 640x480 or 320x240?



Manfred

 

[Lazy8s]

Regular MBX, 240x320 (AA could be on at that level.)

Achieving 30-fps with a low clockspeed means that even heavily power restricted OMAP2420 handhelds like phones will have home console framerates along with the image quality, and the OMAP2 902i DoCoMos perform at this level of graphics. I wonder what kind of graphics demo Texas Instruments will have running on the P902i at 3GSM.

 

[Ailuros]

I doubt any AA would be enabled in such a synthetic benchmark.

 

[amk]

Link to videos of the benchmark on a theoretical platform. They reused the animation from/for the Nokia robot demo.

The 0.54 - 0.66 Pixel/clock ratio seems to be a little bit odd. Have you implemented only a half pipeline in the MBX?

I'd previously come to the conclusion that MBX could do 1 texel per clock, MBX-Lite 0.5 texels per clock and MBX-Pro 2.

Perhaps trilinear filtering was on. That is a substantial performance hit on MBX (see this post), as it was on PVR2. The videos and screenshots certainly looked trilinear filtered to me, so perhaps it was on by default. The disclaimer on the video refers to FPS and resolution optimised for playback, not quality settings.

Also the results with the 2 texture-units seems a little bit strange. Do you really mean texture-units or texture layers?

From 3DMarkMobile06 page, one of the feature tests is "Pixel Processing – Simple fill rate test, with single and multi-texturing, benchmarks raw pixel processing performance. Texels per second are measured." Maybe IMG mean "virtual" texture units, visible to the API and simulated by MBX?

 

[Lazy8s]

Interesting that the ratio of triangle rate with and without lighting differs so much between the Symbian and Linux configurations.

 

[Ailuros]

Interesting that the ratio of triangle rate with and without lighting differs so much between the Symbian and Linux configurations.

At least with lighting the results seem normal to me; a 30% difference for 32% higher clock frequency.

 

[Megadrive1988]

MBX is a pretty old technology now, is it not?

 

[mboeller]

I'd previously come to the conclusion that MBX could do 1 texel per clock, MBX-Lite 0.5 texels per clock and MBX-Pro 2.

so have they tested an MBX-lite or an MBX?

Perhaps trilinear filtering was on. That is a substantial performance hit on MBX (see this post), as it was on PVR2. The videos and screenshots certainly looked trilinear filtered to me, so perhaps it was on by default. The disclaimer on the video refers to FPS and resolution optimised for playback, not quality settings.

Trilinear filtering could explain the, at first sight strange MTexel/sec rate.

From 3DMarkMobile06 page, one of the feature tests is "Pixel Processing – Simple fill rate test, with single and multi-texturing, benchmarks raw pixel processing performance. Texels per second are measured." Maybe IMG mean "virtual" texture units, visible to the API and simulated by MBX?

I ment this sentence:

Fillrate with 1 texture units Texels per second 33.73 Mpps
Fillrate with 2 texture units Texels per second 33.56 Mpps

I still think something is wrong here. Why should someone mention the Mtexels/sec for 2 Texture-Units?

 

[Simon F]

s
I ment this sentence:

Fillrate with 1 texture units Texels per second 33.73 Mpps
Fillrate with 2 texture units Texels per second 33.56 Mpps

I still think something is wrong here. Why should someone mention the Mtexels/sec for 2 Texture-Units?

Presumably it should read "texture layers"

 

[Lazy8s]

MBX is around two years old. It was complete in silicon in Fall 2003 and was released in consumer product in Fall 2004. While PowerVR would now be transitioning into the R&D for handheld products two generations after MBX, the current generation MBX is still a high performance solution for the present market. The immediate follow-up to MBX, the handheld SGX line, is currently in silicon at the prototype stage.

The OMAP2420 benchmarked here has always been confirmed to be the standard MBX, not MBX Lite.

 

[Loewe]

so have they tested an MBX-lite or an MBX?

I don't understand your question. It is clear, they tested MBX.

Texas Instruments OMAP2420 Software Development Platform with PowerVR MBX graphics core and VGP co-processor.

Trilinear filtering could explain the, at first sight strange MTexel/sec rate.

Thats rigth, but I have tested a MBXLite (Axim X50v) without a texture and I get near 35 MPisel/sec. The MBX Lite runs with 75 MHz, the result is that what I expect for a PowerVR core!

I ment this sentence:

Fillrate with 1 texture units Texels per second 33.73 Mpps
Fillrate with 2 texture units Texels per second 33.56 Mpps

I still think something is wrong here. Why should someone mention the Mtexels/sec for 2 Texture-Units?

Simon said it, MBX is a core with one unit but with the possibility to use two texture layer.

 

[Megadrive1988]

MBX is around two years old. It was complete in silicon in Fall 2003 and was released in consumer product in Fall 2004. While PowerVR would now be transitioning into the R&D for handheld products two generations after MBX, the current generation MBX is still a high performance solution for the present market. The immediate follow-up to MBX, the handheld SGX line, is currently in silicon at the prototype stage.

The OMAP2420 benchmarked here has always been confirmed to be the standard MBX, not MBX Lite.

okay I asked because when I deja'ed PowerVR MBX I saw posts from 2002 and a few from 2001. I have not been keeping up with the release of devices with MBX, and it was months ago that the SGX family was announced.

 

[Lazy8s]

From the time when PowerVR does its job and starts offering a new schematic of a processor architecture, over a year of production time between the makers of the chip and the end device is typically needed before the result can reach the market. With MBX, PowerVR announced the design at the start of 2001 and had been hinting at its development even more than a year earlier.

 

[Guden Oden]

This business model, and that it takes a year plus for their stuff to reach market is the reason powervr is the perpetually non-successful 3D company that it is.

 

[Lazy8s]

Being included in chip families which sell hundreds of millions of units combined yearly tells indisputably of a success story.

The reality is that the billion-unit handheld market offers undeniable potential and that even companies with the infrastructure to manufacture products based on their own processor designs still aren't beating IP companies to market in the sector to any critical degree. Graphics processor architectures are just general designs anyway, so PowerVR is always available to get back into the PC market should a partner emerge.

 

[Ailuros]

This business model, and that it takes a year plus for their stuff to reach market is the reason powervr is the perpetually non-successful 3D company that it is.

The PDA/mobile market is an entire beast of it's own. Here timing gets dictated by the customers in that case being, large semiconductor manufacturers. Or do you seriously think that firms like Intel, TI and others lack in resources to integrate a 3D chip in their SoCs?

Which competing sollution was exactly faster integrated that you come to such a conclusion? The PDA/mobile market is ideal for IP licensing, much unlike the PC graphics market.

 

[Lazy8s]

With only about 50-MHz of clockspeed needed to run Proxycon quality graphics properly at 30 frames per second, the 3x clockrate of Aurora's MBX+VGP would result in real-world performance that compares favorably to even the home consoles of its generation.

Sega Sammy designed the Aurora platform with the possibility of sublicensing the SH3707 to partner companies who would want to build handheld gaming systems and PDAs. Apple's gaming iPod and Microsoft's Origami would do well to use it.

 

[Teasy]

Is 3DMarkMobile06 available to normal people or just developers? I was interested to see what my Axim X50v (which has a MBX GPU inside) would get in the test. But looking at the Futuremark site it only seems to mention "3DMarkMobile06 Developers Edition".

 

[Lazy8s]

Just developers for now.

Futuremark gave a recent presentation on benchmarking handhelds where they showed a chart which seemed to indicate about a 22 frame per second performance for the Axim on a 15k polygon per frame workload, which is maybe not so coincidentally in the range of their 3DMarkMobile06 game-type tests. Although the chart was labeled for illustration purposes only and the workload wasn't identified to be one of the 3DMarkMobile06 tests, the chart may have reflected their findings from testing with the new benchmark software.