Most money is made on the low/medium range hardware. The top of the line makes less money. Its essencially for entusiasts and also for publicity ("we have the fastest GPU, so we have the best GPUs!")

Take this theoretical approach. Basically, every CPU could do what a GPU does now thru a SW renderer. Well, it won't be fast, but everything what a GPU can do, a CPU can do, too. Even full shader support (e.g. check out http://www.transgaming.com/... if you read thru a couple of SW render books, this will pop up at some point).

Anyhow, now, imagine, a CPU, which addresses a couple of the bottleneck SW rendering implies and add some specific HW to support the renderer. For example better SIMD performance (super SSE) and some HW to support the rasterizer (texture unit), further more, split the algorithm to do tile based rendering. With about 32 cores, you would end up at a 256x144 pixel tile @ 720p resolution. This is pretty small and very efficient to render. (Just to give you an idea. Tile based rendering doesn't necessarily mean, brute force splitting areas into pieces).

This is basically what Larrabee does. Kind of interesting how history repeats itself. It started in SW, moved far into discrete electronics (fixed function set), came back to shaders and ends up fully programmable again. The next NVidia & ATI chips won't be any different. Shaders will take over. Its just at its core, Larrabee could function as a CPU as well. Oh, yeah, not to forget the CELL, which I think inspired Larrabee a lot.

Because a CPU-core doing graphics would be many tens of times worse at it than even the worst graphics card you can currently buy. CPUs are not built for the very specific computations that are needed in graphics rendering, they are multi-purpose calculators.

That's maybe the controversy around Larrabee. See, the CELL has a discrete core for general purpose calculations, and a set of specialized streaming cores (which work great on vectors, too).

On one side of the spectrum, there is the classic CPU, that is, only general purpose operations (like add, mul, and mem moves) and all i86 have some for of SSE (not every cpu has that, e.g. PPC, or MIPS or ARM does not as standard - PowerPC has VMX, now a days - but that only as a side node). This ISA is very slow at a lot of data. Needs to issue a command + data all the time. Oh, yeah, and good at branching and loops.

Then you have GpGPUs. Not a lot of general purpose ops here. Usually no branching or looping. Definitively not suited for large calculations and iteration, but very good at processing large amounts of data.

The CELL is in the middle and combines one general purpose core, adds these "GPU" like stream processors (which they are not really) but adds a bit more op codes to make them more flexible. Does not reach GPU performance, due to the more flexible ISA. The CELL (design) has its issues, like no shared mem access, is one of them (if you'd say that's an issue).

Now comes along Larrabee. Uses a lot what the CELL had envisioned (and other, possibly), solves the mem access "issue", but adds even more things from GPUs (texture unit, more instructions specialized for pixel/vertex processing ?) and makes all cores coherent with some sort of general purpose ops (x86 compatible).

I am not sure where this fits into here. Its like a merged GPU & CPU, but heavily inspired by the CELL (just because it was there before - intel sure had similar ideas back then, when CELL started).

It is a massive multicore CPU with extreme extended SSE and a little bit more pixel supporting HW.

Technically, Larrabee wouldn't need a CPU, but well, its not 100% compatible with the x86, so, still a need to have an x86 "control" cpu to run the OS on.

Actually, current processors are much more powerful than 64 of the first Pentiums. Not only do they have multiple cores, but their cores also clock around 50 times as fast and they have almost 250 times as much cache (as a single Pentium I). These are just raw technical improvements, they have also improved in many other ways.

If someone was dumb enough to pair a cell cpu with a cell gpu it would be equivalent to going back in time to pentium 1 computers with DOS operating system. don't think that will ever happen.

the cell processor is a scam.

*sigh*...
Wether its a scam or not, my point is.
GPU's with more "cores"(or whatever they are called in the Cell i forgot it).
Now its possible to connect 4 Videocards and make them work as one.
Imagine the power of multithreaded GPu's combined with 1 other or more.

I don't think there will be a "Cell CPU" and a "Cell GPU". There could, however, be a next iteration of CELL which looks similar to Larrabee.

There was an article somewhere, where the CELL architect explained why they choose the EIB instead of a xross bar and SRAM instead of caches - basically what Larrabee is. Basically to keep the size down and avoid cache coherency problems but maintaining extreme memory bandwidth (the EIB can theoretically transport ~300GB/s data - IBM reach 192GB/s or so).

So, I wouldn't expect a next generation CELL all of a sudden use cache mem and x-bar interconnects, but maybe optimized and more SPUs, any number of PPU cores (2+) and possibly some more pixel processing HW on die (like moving parts of the RSX onto the CELL). Larrabee has a hardware texture unit to support the shaders. Such a cell would not require a graphics chip but could possibly be competitive to today's GPUs.

That also shows, Larrabee would be a showcase implementation for a console where they can break with legacy support (don't need to be compatible to any OS out right now). I just can't imagine, Sony would break with STI (their multi billion $ investment) and go with Larrabee instead of investing into the next CELL version. MS might, though, except they join the "CELL" aliance. I doubt they'll develop something like that on their own. Possibly, with AMD/ATI as a partnership ?

A Larrabee Wii2 could possibly emulate a Wii, though.