Overwritten.net
Community => General Discussion => Topic started by: WindAndConfusion on Tuesday, December 02, 2008, 05:53:48 PM
-
http://arstechnica.com/news.ars/post/20081202-windows-7-will-sport-a-dx10-software-gpu.html
So maybe you saw this already: Windows 7 will add support for "software accelerated" DX10 graphics on machines that don't even have GPUs. The performance is crap (as you'd expect), but it is significantly less crap than you might have guessed. In particular I noticed this:
(FPS-A is average FPS, FPS-m is minimum, and FPS-M is maximum.)
CPU
Time FPS-A FPS-m frame-m FPS-M frame-M
Penryn 4 Core @ 3.0GHz
351.35 5.69 2.49 1967 10.95 980
Intel DX10 Integrated
386.94 5.17 1.74 1974 16.22 995
You may notice that an Intel CPU (in fact several Intel CPUs) generally outperformed Intel's best GPU.
-
And that's why all these companies like Epic are pissed at Intel.
-
Tweak Guides had a post about this last week. Apparently it's called Windows Advanced Rasterization Platform (WARP).
Here's the In-Depth guide (http://msdn.microsoft.com/en-us/library/dd285359.aspx) from MSDN.
-
Not sure, but this could be very forward thinking. As the silicon miniaturization reaches some absolute limits, the only way to increase CPU performance becomes multiplying the cores. A software-rasterizer standard may seem silly now, a throwback to Quake-1 technology. But as more CPU cores become available to dedicate themselves to the graphics load, and the engine is adapted accordingly, it may end up rivaling the GPUs. The bus bottleneck would have to be solved, though. Video cards go straight out to the display. Just a thought.
-
Alright, I know you already know this, but I can't help saying it anyway:
Not sure, but this could be very forward thinking. As the silicon miniaturization reaches some absolute limits, the only way to increase CPU performance becomes multiplying the cores. A software-rasterizer standard may seem silly now, a throwback to Quake-1 technology. But as more CPU cores become available to dedicate themselves to the graphics load, and the engine is adapted accordingly, it may end up rivaling the GPUs.
Won't happen for two reasons.
The fastest system RAM you can expect to buy these days is PC2 8500; with two channels that gives you "17" (~16.6) GB/sec theoretical peak bandwidth, which is still less than the Radeon 9700 (that came out in 2002!). Currently the slowest ATI cards have 8GB/sec bandwidth, and the fastest have 115(!!).
The best CPUs today can run four concurrent threads (or eight if you count HyperThreading, which you shouldn't in this case), whereas the cheapest GPU ATI currently manufactures can run 80 (albeit at a lower clockspeed, but also with a shorter pipeline).
The bus bottleneck would have to be solved, though. Video cards go straight out to the display. Just a thought.
Out of all the things holding back software GPUs, I think this is the least of them. 1080p60 with 32-bit color is only about 470 MB/sec, whereas HyperTransport is about 50 times that fast and PCIe-x16 is about 8 times that fast.