On-package GPU and Graphics Turbo

Arrandale and Clarkdale are two-die packages. There's the 32nm CPU die and next to it is a 45nm DirectX 10 GPU die (no DX11 support until possibly Larrabee).

This isn't Larrabee (yet), it's a direct descendent of the graphics in G45. While G45 was built on a 65nm process, the 'dale graphics is built on a 45nm process.

The smaller transistors enable much higher performance. While G45 had 10 shader cores, the 'dale GPU increases that to 12. A number of performance limiting issues have now been resolved, so we should see much more competitive performance from Intel's graphics.

The memory controller has been moved off of the CPU die and is on the GPU die instead. It's still on-package so you get decently low latencies, but it shouldn't technically be as low as on Lynnfield. This is a temporary problem that fixes itself once the CPU/GPU are on the same die with Sandy Bridge.

Sandy Bridge brings on-die graphics

I've already explained turbo mode quite a bit so I won't rehash it here. The technology basically allows you to run your CPU at the fastest possible frequency regardless of how many cores are active. Westmere has this.

Arrandale will support graphics turbo modes, while Clarkdale won't. Clarkdale graphics is already running as fast as possible regardless of TDP.

If the GPU demand is higher than the CPU demand, the CPU will allocate more of its TDP to the GPU and vice versa.

AES-NI: Encryption/Decryption Acceleration Quad Core Performance From Two Cores?


View All Comments

  • strikeback03 - Friday, September 25, 2009 - link

    Remember, there was initially supposed to be a 45nm low end, it was axed and 32nm pulled up. So I wouldn't be surprised to see a 32nm Sandy Bridge, but probably a while after the higher-end versions launch, same as we saw with Nehalem. Reply
  • sunefred - Friday, September 25, 2009 - link

    Thats a perfectly sized and nice looking case that the Clarkdale is housed in. Is that comercially available yet? Reply
  • MadAd - Saturday, September 26, 2009 - link

    yeah third on that, what itx case is it? Reply
  • MadAd - Saturday, September 26, 2009 - link

    nvr mind, its one of the travla series

  • cdalgard - Friday, September 25, 2009 - link

    does anyone know what case is used for the mini-itx system that is shown in this article? Reply
  • KompuKare - Friday, September 25, 2009 - link

    "Why not give all chipsets video out support? Intel is big on differentiation. The P55 chipset is a bit more overclocking friendly, a set of optimizations that you won't see in H57. Both types of motherboards will take Clarkdale processors."

    I do wish reviewers would be a bit harder on Intel with their artificial market segmentation.

    My personal pet hate is the game of which CPUs have Intel-VT and which do not. I see a lot of people being disappointed trying to run XP-mode in Windows 7 not to mention trying to run VirtualBox and VMWare properly.
  • gstrickler - Thursday, September 24, 2009 - link

    50%-60% faster than the lousy IGP in G45 when running 3DMark Vantage. The G45 had the X4500HD, so a 50%-60% improvement still puts it at about half the performance of the Nvidia 9400M (G) chipset IGP. Fast enough for daily tasks and maybe games from 4+ years ago. That makes it adequate for HTPC and the casual desktop PC or business PC, but it's still way behind the IGP in Nvidia or AMD chipsets.

    The whole idea of putting an IGP on the CPU just sounds like a bad idea to me. It's great for embedded SoC systems, but for a GP CPU, I think it's the wrong place for it. Don't want the Intel IGP? Too bad, since they moved the memory controller off the CPU die and onto the die with the IGP, you get it whether you want it or not. Intel is going to force their less lousy IGP into your machine. Hopefully the IGP is fully power gated so that if you're not using it, it at least doesn't cost you power.

    I'll wait for some independent tests before I bless or condemn it. 3.33GHz Clarksdale with Turbo Boost and HT enabled is estimated to be almost as fast as a 2.66GHz C2Q on SPEC*Int_rate2006. Let's see, that's 2 cores running at perhaps 3.66GHz + about 33% (for the HT threads) = 7.33GHz * 1.33 = 9.75GHz vs 10.66GHz (4 x 2.66GHz). Allowing a little extra for the IMC and other CPU improvements. Sounds about right on performance, and the power consumption should be down quite a bit.

    Aside from the IGP, it sounds pretty good if the price is right.
  • Inkie - Friday, September 25, 2009 - link

    So, what you find wrong with the IGP is that it isn't very good for gaming? It's just a question of degree. No IGP out there is very good, from anyone. Serious (3D graphics) gamers buy graphics cards anyway. Reply
  • gstrickler - Friday, September 25, 2009 - link

    No, what I find wrong with the IGP is that it still sucks compared to the IGPs from Nvidia and ATI/AMD. With OpenCL (and maybe DirectCompute), the OS and some applications are starting to use the GPU to accelerate non graphics functions. Given that, putting in an IGP that is significantly lower performance is a performance impediment, even for non gamers.

    While I have no doubt that Intel will support OpenCL and/or DirectCompute on Larrabee, I don't remember seeing any indication of whether they will support it on this GPU. Even if they do, the performance is likely to be far lower than the performance you would get from an IGP from Nvidia or ATI/AMD.

    Like I said, it's "less lousy" than Intel's previous IGPs, and it should actually be adequate for most uses up to low end gaming, but it doesn't have the power to provide many of the benefits that an IGP can provide and there is no reason to build it into the CPU. It should be in the chipset, and like previous generations, you should be able to choose a chipset with it or without it.
  • Inkie - Saturday, October 03, 2009 - link

    Anybody serious about OpenCL or DX 'Compute Shaders' will also buy a discrete GPU, for now. Reply

Log in

Don't have an account? Sign up now