Intel Nehalem, Larrabee CPU/GPU Information

http://www.hothardware.com/Articles/Intel_Showcases_Dunnington_Nehalem_and_Larrabee_Processors/

Figured I'd put it in the gaming section since that's where most people would be interested...

Should be interesting, though. Will integrated graphics actually get good with the launch of these?

Edit: Other links...
http://www.pcper.com/article.php?aid=534
http://www.dailytech.com/Intel Discusses GPU Hybrid CPUs/article11088.htm

 

Intel claimed awhile back that they would have improved integrated graphics 10x by 2010 (compared to the performance of the X3100).

So yeah, lets hope for the sake of their gaming clout that integrated graphics will actually start to work a little better.

But I don't think Larrabee is technically integrated graphics. It is more of a GPU integrated into the CPU isn't it?

 

10x... so that means the 3D06 mark should be 5000. Right.

 

So now Intel abandoned the X3100? Yea, great job.

 

Nope, Larrabee is Intel's answer to Nvidia/ATI discrete graphics. Putting Larrabee as integrated graphics is akin to putting Geforce 8800GTX integrated, which is not possible without horrendous costs really.

Intel's current X3x00 architecture will continue with GPU integration on Nehalem. It's using the similar fundamental architecture with updates. Here's how it will go in the future...

GM45/GM47: GMA X4500/X4500HD IGP, basically X3100 with higher core clocks, more execution units and more bandwidth, and dedicated hardware video unit. 10 EUs, 640MHz for GM47 and 533MHz for GM45, and DDR3-1066 support.

Auburndale: Laptop version of Nehalem with IGP on CPU. 12 EUs, 2 more fixed function units

 

Nahalem isn't IGP on CPU so much as IGP on package/socket beside CPU.

I'd wait for clarification of the flexability and X86 support in Larrabee spoken of by intel before deciding even if we need a traditional IGP. From most of their early information the push is closer towards a more flexible single big chip solution than in parting out resources that can sometimes be useless in certain apps.

Just have to wait and see as always.

 

.

Nehalem, the versions with IGP, is MCM version. It'll be as close as you can get to having the GPU core designed as part of CPU core. The two dies will be connected with QPI bus, and will be able to communicate with each other very fast.

The versions with IGP, is also based on the X3x00 architecture. It has 12 EUs(X3100 has 8, X4500 has 10), has more fixed function units, and support for OGL2.1.

 

Which is still to separate dies on a single package, which is a more production than a single chip, and while the QPI/CSI bus is fast it's not as fast as part of the single chip, and not as flexible either.

From a production standpoint it's little benefit to have the VPU functionality as a daughter die in an MCM configuration, since IGPs are not throughput bound but resource bound, so if you want to share those resources it makes much more sense to move it to the chip than through another I/O layer. eDRAM is fine as an MCM configuration on the Xenos, but RAM has less of a latency issue than having to wait for a clock cycle or two or retrieve information across the chip for data calls from local buffers, that will add up to inefficiency very quickly, heck we already talk about the inefficient arbiters and scheduling of the unified shaders in the R6xx/G8x designs, this will introduce even worse situations if they seek to share any resources whatsoever.

It wouldn't matter much if this was what everyone were doing, but since the competition is moving towards unified packages, then there's no advantage to only doing a partial step that's more costly from a production standpoint.

 

In other words, Intel's next gen IGP STILL is gonna suck....

Perhaps AMD will have more luck with its PUMA release...

 

Flexible?? I think its the other way around. Its more flexible to have it on MCM since its easier to switch one part of the MCM with another device.

The connections won't be enormously slower than fully integrating because the cores will be close together physically. And since when was graphics more latency bound than bandwidth??

Probably. We still don't have the X4500 yet, which is what's going to try to compete with the 780G etc. Part of the problem with having a bad IGP for Intel is that they don't seem to want to dedicate too much resources to creating a better IGP, they strictly follow the "IGP is for low cost" scheme. The G45 will be sold at exact same price as the P45 chipset(previous Intel IGPs costed $4-5 USD more than the equivalent non IGP chipset).

If Intel can ideally make a really good driver, X4500 should be competitive(Read: Not faster). HL2 runs well, but Portal doesn't, which is based on the same graphics engine, which puts the performance difference almost squarely at bad drivers.

 

Any ideas when laptops with X4500HD or 780M will become available? And who will offers these?

 

The X4500 will be available after June, that is when Montevina will be shown off at a computer show in Taipei. As for the 780M, that should also be available in Q2 of this year.

 

I am curious. Would a single dual or quad core Intel Core 2 Duo be able to provide appropriate power to drive 3D graphics comparable to say, and 8600m GT? As in the Core 2 Duo acted as a GPU instead of a CPU?

 

No, CPUs are too general purpose to run games fast at all. It's designed to do everything at relatively fast speeds. Unless your GPU is running vertex shaders off CPU, at the level of performance ALL IGP offers, a CPU does not matter at all. The current only IGP to run vertex functions on CPU is the X3000/X3100. It's pretty slow as you may know.

If CPU truly did every 3D functions like it did before 3DFX cards, then you probably won't even be able to run 3-4 year old games properly.