Some details of the Nehalem IGP from IDF Beijing 2009

-Additional Fixed Function units to improve Vertex throughput
-6 threads/EU(Execution Unit). GMA 4500 features 5/EU and GMA 3000 features 4/EU
-Hierarchial Depth Buffer
-Improved Extended Math performance and larger caches(it looks like GMA 3000/4500 share two Mathbox pipelines with 8 and 10 EUs respectively and this new generation will have three)
-15(?) EUs(10 in 4500 and 8 in 3000/3100)

New architecture is nice, but I think another big change overlooked is the potential of having a (very)fast memory connection. Current Core 2 systems only achieve 5.5-6GB of real world memory bandwidth. Core i7-based system with dual channel memory should achieve 13-14GB of measured bandwidth with DDR3-1333. The GPU won't be fully integrated into CPU but it will be MCM'ed(like how Core 2 Quad is) with CPU core in one die and GPU/PCI-Express/Memory Controller in another connected together by QPI link capable of transferring 25.6GB/s.

I think the biggest thing about Nehalem is solving the "memory problem". Nehalem servers have shown this throughly. Will IGPs be the next? And will it be enough to let Intel be competitive? It'll be interesting end of this year.

 

extra bandwidth is not going to make up for the lack of raw power. i'd consider it a success if it could reach hd3450/9300m grade performance.

 

I agree. Power mostly depends on the architecture and processor of the gpu.

 

The thing which I could call it the biggest flaw in the GMA 3000/4500 is that increasing settings don't affect performance much(or in another way, lowering settings don't always make your game run faster). Now you may call it a good thing because that looks like its not losing performance in higher settings. It's actually the other way around because there's a bottleneck there which doesn't allow it to render faster with lower settings!

Raw power is irrelevant when going from higher settings(say 1280x1024 High) to lower settings(800x600 Low) doesn't grant you better performance. Here's a good example. With Crysis 1024x768 DX9 Med settings the G45 is only 20-25% away from the AMD 780G and is competitive: http://www.hkepc.com/1510/page/5#view

Now say you want to play the game and thinking lowering settings will net you performance. With the 780G, lowering settings to Low will grant you 2-3x performance boost and you'll get by with 20-25 fps.

http://www.anandtech.com/printarticle.aspx?i=3430

With the G45 you lower the settings and you get mere 30-40% faster than with supposedly much higher settings. I have shown Crysis as its a readily available result. But I can tell you quite few games already do that.

That does not mean bandwidth is not important for the GMA. It's very important. When ATI's R300(Radeon 9700) and Nvidia's NV30(Geforce FX5800) gained 30% for doubling bandwidth and GMA series gets 50-60%, you know immediately that one of the bottleneck is still memory bandwidth. The thing is at least in desktops they crippled the G45 by using DDR2-800 memory when they could have used DDR3-1333.

Remember Intel's promise 10x performance by 2010?? Here's a rehash of the graph:
http://techgage.com/article/idf_07_sf_paul_otellini_keynote/2

130nm(base) GMA 950: 1x
90nm GMA X3000: 1.3x
65nm GMA X4500: 2.2x
-In end of this year, we'll be here--> 45nm Nehalem's GMA: 6x

Problem with marketing is you gotta read the fine print and in between. GMA X3000 seemed much more capable than GMA 950 but actual performance is not much better. It's more about bringing compatibility with full DX9/10 support and hardware geometry processing.

Same with the X4500. Intel claimed 3x performance over previous gen IGPs, but if you look carefully its comparing to the hardware VS-less/VS 3.0-less G33 in 3DMark06, which makes the comparison flawed because X4500 can run all of 3DMark06 while G33 can't. Performance-wise X4500 is only 1.7x better than X3000/X3100.

 

Dr Who, resident Intel rep at Xtremesystems, has hinted that Nehalem IGP will be much more powerful than anybody expects. I forget his reasoning, but it makes sense, and while he's still working for Intel, he's known for being pretty reasonable and doesn't exaggerate the teaser info.

 

Let's not get ahead of ourselves yet though. I'd be happy if the IGP is 30-40% behind the competition(currently its 1/4 to 1/3). G45 barely met the graph thanks to the delay(delay usually means they are having a hard time making it meet expectations, not because they want to make it better than they need to).

 

Things are starting to look promising: http://www.theinquirer.net/inquirer/news/992/1051992/intel-caught-graphics-shocker

 

That article offered no information...