WinHEC 2008

I attended WinHEC this last week; here are some of the things I found cool/interesting (I'm not a big MS fan but they do some cool stuff anyway):

1. Hyper-V.

Hyper-V Server 2008 is a hypervisor/virtualization engine which is free for download from MS. This runs pretty much any OS supported by x64 architecture. It can host both 32 and 64-bit OSes but the host node itself must be 64 bit. It supports up to 24 logical processors and up to 4 SMP cores per host. In the upcoming Server 2008 R2 (not the free download) it will support live migration of VMs between nodes with no network drops and no downtime. They demoed that live to HUGE effect.

2. Windows 7

The biggest things I paid attention to in Windows 7 were the new device management interface and the XPS printing subsystem. It is kind of hard to describe the new "device experience" system in words but it basically replaces Printers&Faxes with anew interface which integrates print, fax, and ALL attached devices. All device functionality should in theory be accessed through this new unified interface. A lot of the talks at WinHEC were geared toward pitching this to the device manufacturers. If they buy into it, it should be really, really cool.

As far as XPS, I'm a print driver geek so much of what I paid attention to might not be terribly interesting to you, but the gist is that they have integrated a software XPS rasterizer into the print filter subsystem which is supposed to make it brain-dead simple for driver authors to integrate XPS into their drivers. Why are they pushing this? The new Direct2D has quick paths to XPS but not to GDI. So for apps which want to render using Direct2D as well as print this should reduce the amount of code that has to be written. And of course, XPS is kool-aid to MS and of course they want everybody to move over.

3. Heterogeneous multi-core

Again, much more was told than I can recount here but the general movement of Intel, AMD, and MS in synchrony is toward heterogeneous multi-core systems -- imagine a single chip with 4 Nehalem cores plus 16 Atom cores, all individually power-gated, sleepable, with the OS shuffling tasks between cores depending on speed and power requirements.

HetMC also included talk about integrating GPUs directly with CPUs, so you might see a chip with 4 fast cores, 16 slower but more power-efficient cores, and 1024 tiny GPU cores all on a single die, again with the OS automatically slicing tasks.

Intel's general model for cache in such a system would be individual L1/L2 cache per-core but with a globally shared L3 cache with sizes in the 32-64 megabyte range.

Now all this heterogeneity will require new tools to let programmers deal with it without too much trouble, so MS is incorporating parallelization libraries into VS2010 which from what I saw, don't suck too bad.

It's weird how being at a conference surrounded by Microsoft fanatics rendered me more sympathetic to MS while I was there. Now that I'm home again I'm getting some clearer thoughts about exactly what I saw, but I'm still impressed with a lot of things.

One of the things I think happen with 3D development is that much work is still done on the CPU and this then passes to the GPU what this processor can handle which is basically rendering based calculations. I'm not anywhere close to a 3D programmer, but this is my perception. So, I always retained this notion that 3D is still highly inefficient as far as processor cycles are concerned.

So my question is, if I'm more or less right above, will this GPU integration into CPUs reduce processor cycles and further tighten the processing and rendering in 3D? Is that the idea?

Quote:

---

Originally Posted by Mario F.

So my question is, if I'm more or less right above, will this GPU integration into CPUs reduce processor cycles and further tighten the processing and rendering in 3D? Is that the idea?

---

Not really 3D itself, since the idea behind putting the GPUs on the video card was to reduce bus pressure. For graphical applications we'll still be using the GPUs on the graphics card directly. For GPUs integrated onto main processors the idea is more physics, AI, and general purpose computing applications. For example look at the nVidia Cuda program which provides methods of running generic code on video card GPUs. The difference will be that compute-oriented (rather than graphics-oriented) work will happen on the on-die GPUs instead of the video GPUs.

Right now the only bummer with using on-graphics GPU capacity is the traffic across the video bus (whatever technology that might be). The theory is that on-board GPUs will reduce that traffic and start driving a movement toward using the GPUs for general-purpose computation.

One thing I forgot to mention was that the VS2010 pre-beta (I got a copy!) has fairly complete suppot for C++0x -- they used the lambda syntax in their native-code demo for the parallelization framework.