In the past I mentioned that Intel, Nvidia and AMD/ATI are trying to expand to each others application domains (CPU und GPU fusion, a demo with OGRE3D and Nvidia's expansions). One may think that it's about market share but if you consider that parallelism is the key for future performance increases, then it looks more like a natural evolution.
GPUs no longer separate Pixel- and Vertex-Shader processing units on hardware. With DX10 it got unified – pixel and vertex (+geometry/tesslation) shaders are executed by the same hardware units. Therefore if a scene uses few vertices but a lot of pixels the hardware is less idle as it adapts accordingly – which is equal to higher performance. Therefore GPUs become more programmable while CPUs become more parallel.
At Siggraph 2008 there have been a couple of talks about CPU/GPU architecture. Their course notes are now online: http://s08.idav.ucdavis.edu/
Interestingly Kayvon Fatahalian thinks that Intel's chips needs ~ 32 Larabee units @ 1GHZ to reach the processing power of a current GPU which he says is about 1 TFLOPS. But today a game needs GPU AND CPU processing power.
No matter how fast the CPU gets, CPU+GPU will always be even faster! Of course money matters and if the ratio moves heavily towards one favor, it might create a change. And that's maybe the point. INTEL's Larabee introduces a much higher flexibility, i.e. easy access to render-targets, so this could reduce costs. For this aspect also checkout the notes from the ID software guy where he mentions Voxels engines and other examples. On the other hand it brings even more choices and that doesn't make life always easier!
Until a shift happens (and if ever) we will probably see Larabee as add-on chip. Here's an interesting Intel statement from one of the sources of my earlier blog articles i mentioned at the beginning:
"Intel is not predicting the end of the discrete graphics business. Moore's Law has allowed Intel to innovate and integrate. As a result, we expect that we and others will integrate graphics and visual computing capabilities directly into our CPUs in the future much like floating point coprocessors and other multimedia functions have in the past. However, we don't expect that this integration will eliminate the market for higher-end discrete graphics cards and the value they provide."
Like this it sounds a bit like Intel's Next-Gen GMA chip. That's probably the short-term perspective. It is said that Larabee is available in 2009 with 24 to 32 cores. But if they release those at 2 GHZ per core it sounds more like a GPU replacement unless you consider the too high power consumption (sources: german , english).
If you want to read a bit more about NVIDIA's new GTX 200 architecture and it's performance, Tom's hardware has a detailed review.
And here a very detailed article about Larabee from early august.
Happy reading 😉