Yeah, like tons of vector power, but little scalar power.
The vector processors can also do scalar processing. According to the people actually using them they are very, very fast.
It would be similar to re-introducing math coprocessors (like what Ageia is trying to do, and will fail
Er, maths co-processors were a great success, so good they integrated the functionality into every CPU available today.
Intel and especially AMD are also reintroducing the idea of co-processors, spurred on by of all things Cell (a "wake up call" according to the AMD CTO).
instead of making a dual-core x86 chips, you would have one puny x86 core and a lot of dedicated SSE cores. I can only imagine how Windows would handle that kind of chip without a major rewrite.
A bit slower of some stuff, a hell of a lot faster on computer intensive stuff.
Both Intel and AMD are talking about building exactly that sort of processor.
It's a good idea, but the technical makeup of Cell is terribly imbalanced. Note that Sony originally thought they could use multiple Cell processors to replace the GPU. When real-world benchmarks weren't very good, they went with a more traditional layout, and the fanboy hype of distributed computing (and those idiotic schematics) pretty much died.
The distributed computing idea gets a lot of interest and is still under development. As for real world benchmarks GPUs will be better at some things (e.g. texture shading) and worse at others, one Cell developer got a Cell to outperform a GPU 5 times over - and that was *without* optimising the algorithm.
If you plan to write your own algorithms, you need to know custom VMX coding to use them, though. Older algorithms don't benefit, either, so you have to re-write a lot of stuff. Hence, the high costs.
Only for performance sensitive code (around 5%). Also there is a lot of effort going into compilers so you have to do only minimal coding changes.
What you may not be aware of is the changes going multi-core will make to conventional processors, you are going to need to rewrite everything no matter what processor you are using.
What may seem like odd design decisions in Cell today will be seen as major advantages in a few years.
Build a 8-16 core CISC (or RISC) processor and you'll find the need for cache coherence (keeping caches in sync) will send cache latency through the roof and cripple existing code. look up the design of Suns's rock processor, it's a bizarre design (16 cores, shared L1s, 4 shared 512K L2s) for a very good reason.