Intel's new ultra mobile platform will sport a GPU. What should we expect?

F

Farid

Artist formely known as Vysez
Veteran
Supporter
Quick recap for thoses who don't know about this new Intel initiative:

  • Intel plans on releasing a new "ultra mobile" X86 platform. (Think internet tablets, PocketPC and eventually cellphones with Moorestown.)
  • The first generation (2008), Menlow, will feature a small in-order CPU, Silverthorne, and will have a chipset, Poulsbo, featuring the memory controller and, the relevant part for this forum, a GPU.
  • The next-gen (2009), Moorestown (video and marketing blab), will have the CPU, GPU and the memory controller on the same die, Lincroft, and its chipset, Langwell, will only handle I/O.
  • Menlow and Moorestown should both be 45nm product.

Diagram from PC-Watch (JP):
ubiq205_03l.gif


So, with the basic info I gathered on this platform sumed-up. The question is... what should we expect from the graphics cores of these platforms?

What 3D capabilities should we be looking for here, OGL ES 1.1, 2.0, OVG 1.0?Should we expect an Intel design or would PVR's SGX be a good bet?
 
Poulsbo Graphics:
3DMark03: 500
3DMark05: 150

http://www.theinquirer.net/gb/inquirer/news/2008/02/06/ever-wanted-know-silverthorne

"The FSB connects to a rather unique chipset called Poulsbo, and it is pretty feature packed for a 22x22x2.3mm package. It is a single chip NB and SB, has integrated 2D and 3D as well as a video engine, USB 2.0, PCIe, SDIO2.0, and high def audio. It will push pixels from LVDS 1366 x 768 x 24b all the way up to 1080i while consuming 600-800mW average. Yuck, average. In any case, if you were wondering what happened to the PowerVR core, go by a Silverthorne box and see for yourself.

It supports 400 and 533FSBs as well as DDR2 at the same speeds, up to 1G worth. The graphics is all new, not the Gen 3.5 of it's predecessor Little River. They expect this much lower power version to score a bit less than LR in 3DMark, 500 in '03, 150 in '05. You will probably hit the 2.25W TDP doing this."

DX9, OpenGL 1.5
 
CarstenS said:
A little bit more info:
http://extreme.pcgameshardware.de/m...-ip-intels-poulsbo-chipsatz-gma500-us15w.html

Run through your favourite online translator or enjoy the omnilingual pictures. ;)

Interestingly, the AF-tester produces quite a lacking trilinear filter analysis, but the aniso derived from it seems to be promising given it's basis.
Click to expand...

Thanks for the heads-up, Carsten!

I got a blast from the past seeing that you used 3DMark2001 and Village/templemark. Kyro, neva' forget!

Concerning the fillrate throughput tests. It would be cool if you could cross reference these results with some "meanier" solutions like the 9*00M* series from NV, as well as the previous generation of IGP from Intel. Just to get a frame of reference, you know.
 
I've tried to make them fillrate work on my laptop at home yesterday. Sadly enough, it only has a GMA900, which Intel doesn't bother to equip with proper drivers, so that it fails to emulate a hardware vertex shader and consequently also fails to run this test.

I've browsed through some old results from this fillrate test and found the shader fillrates roughly comparable to those of the ancient Radeon 9600 Pro :)

Code: 
FFP - Pure fillrate - 1167.590454M pixels/sec
FFP - Z pixel rate - 1494.546997M pixels/sec
FFP - Single texture - 1156.886841M pixels/sec
FFP - Dual texture - 577.300598M pixels/sec
FFP - Triple texture - 393.923584M pixels/sec
FFP - Quad texture - 312.159546M pixels/sec
PS 1.1 - Simple - 781.752075M pixels/sec
PS 1.4 - Simple - 781.783997M pixels/sec
PS 2.0 - Simple - 781.801758M pixels/sec
PS 2.0 PP - Simple - 781.760559M pixels/sec
PS 2.0 - Longer - 394.298035M pixels/sec
PS 2.0 PP - Longer - 394.300842M pixels/sec
PS 2.0 - Longer 4 Registers - 394.304077M pixels/sec
PS 2.0 PP - Longer 4 Registers - 394.297424M pixels/sec
PS 2.0 - Per Pixel Lighting - 99.006615M pixels/sec
PS 2.0 PP - Per Pixel Lighting - 99.006462M pixels/sec

Bandwidth-intensive stuff like Pixel, Z and single- texturing are miles off, though.

But as noted in the PC Games Hardware Extreme Thread, real games seem to be imposing quite a different challenge to the GMA500-IGP. Not even CS 1.6 on an empty map ran anywhere near flawless. :(
 
CarstenS said:
I've tried to make them fillrate work on my laptop at home yesterday. Sadly enough, it only has a GMA900, which Intel doesn't bother to equip with proper drivers, so that it fails to emulate a hardware vertex shader and consequently also fails to run this test.
Click to expand...

I don't think you should worry that much about that "minor detail" since unless something has changed there's no hardware vertex processing in the Poulsbo driver either.

Ignore the fact that PowerVR's own demos run like crap on it, just open the detailed 3dmark01 results and have a look at the results. Ironically in sw T&L mode the multitexturing fillrate as well as 8 light high poly count tests are higher than in hw T&L mode LOL.
 
Ailuros said:
I don't think you should worry that much about that "minor detail" since unless something has changed there's no hardware vertex processing in the Poulsbo driver either.
Click to expand...
Well, at least those tests could create an appropriate D3D-Device in Vista. :)

Ailuros said:
Ignore the fact that PowerVR's own demos run like crap on it, just open the detailed 3dmark01 results and have a look at the results. Ironically in sw T&L mode the multitexturing fillrate as well as 8 light high poly count tests are higher than in hw T&L mode LOL.
Click to expand...
So what? :)
GT1 Car Chase High is also higher (almost +50%) in S-TnL, as ist EMBM, albeit to a much smaller degree.
 
CarstenS said:
So what? :)
GT1 Car Chase High is also higher (almost +50%) in S-TnL, as ist EMBM, albeit to a much smaller degree.
Click to expand...

The lack of a proper driver implementation simply holds the core back from performing up to it's peak that's all. From what I heard Tungsten which writes the drivers for Poulsbo isn't using the onboard firmware and due to that the core is falling back to sw vertex processing.
 
*urgks*
If Intel really wants to compete with larrabee in the discrete desktop market, they better start getting their act together and investing a bit of their gazillions of dollars into some serious sw-engineering, aka driver team. :)
 
CarstenS said:
*urgks*
If Intel really wants to compete with larrabee in the discrete desktop market, they better start getting their act together and investing a bit of their gazillions of dollars into some serious sw-engineering, aka driver team. :)
Click to expand...

Larabee has afaik it's own driver team. In the case of IMG IP they use Tungsten drivers to write the drivers for them. That said the latter should have listened to IMG's protests about the issue a long time ago.
 
Blazkowicz said:
if quake3 only "runs at more than 30fps" I'm not much impressed as I would expect that of a nvidia TNT2.
Click to expand...

A KYRO2 ran Q3a times faster than a TNT2 and that with better IQ, but on the other hand those two were GPUs and not miniscule integrated parts. The presentation in question was rather about multi-tasking/-threading than anything else and I answered an obvious joke about printers and not GPUs.
 
Ailuros said:
If there's a printer available that can run as much as 4 H.264 streams simultaneously w/o a hiccup I'd appreciate a link.
Click to expand...
The printer technology produced by this company in Australia (run by my ex-boss) might stand a chance (well, maybe not of the decode, but the refresh rate).
 
You must log in or register to reply here.

Similar threads

J
Will OpenGL ES 3.0 prove to be a short-lived standard?
2 3
Replies
53
Views
17K
jedibeeftrix
J
J
Picking a GPU on par with the rest of my system?
2
Replies
22
Views
7K
almighty
A
A
Holy Wall of Text: My impressions post Playstation 4 (PS4) unveil
2 3 4
Replies
70
Views
19K
(((interference)))
I
A
  • Locked
News & Rumors: Xbox One (codename Durango)
613 614 615
Replies
12K
Views
2M
BRiT
BRiT
P
NVIDIA's Project Denver (ARM-based CPU)
2 3 4
Replies
74
Views
30K
jlippo
J
Back
Top