
Originally Posted by
Charlie Demerjian
Let's look at this long term, say five or so years, the design cycle of a modern CPU. As we've noted earlier, the X86 CPU is about to take a radical turn, and the designs you will see at the turn of the decade won't resemble anything you see now. What do we mean by that? Mini-cores and Larrabee.
[...]
Kevet and Keifer were a mini-core and a CPU made of 32 of those cores respectively aimed at server workloads. It was four times what Niagara was reaching for, but also five years later. Intel is going for the swarm of CPUs on a slab approach to high performance CPUs, and more importantly, is going to upgrade the chips on a much swifter cycle than we've been used to.
With 32 small and simple cores, you can design each core much more quickly than a normal CPU, much more quickly. Design complexity, verification and other headaches make things almost a geometrically increasing design problem. A small core cut and pasted 32 times can mean smaller teams doing more real work instead of busy work, and more teams tweaking things for niches.
[...]
Now, if you add in GPU functionality to the cores, not a GPU on the die, but integrated into the x86 pipeline, you have something that can, on a command, eat a GPU for lunch. A very smart game developer told me that with one quarter of the raw power, a CPU can do the same real work as a GPU due to a variety of effects, memory scatter-gather being near the top of that list. The take home message is that a GPU is the king of graphics in todays world, but with the hard left turn Sun and Intel are taking, it will be the third nipple of the chip industry in no time.
Basically, GPUs are a dead end, and Intel is going to ram that home very soon. AMD knows this, ATI knows this, and most likely Nvidia knows this. AMD has to compete, if it doesn't, Intel will leave it in the dust, and the company will die. AMD can develop the talent internally to make that GPU functionality, hunt down all the patents, licensing, and all the minutia, and still start out a year behind Intel. That is if all goes perfectly, and the projects are started tomorrow.
The other option is to buy a team of engineers that produce world-class products, are battle tested, and have a track record of producing product on the same yearly beat Intel is aiming for. There are two of these in existence, ATI and Nvidia. Nvidia is too expensive, and has a culture that would mix with AMD like sand and Vaseline. That leaves ATI, undervalued and just as good.
Bookmarks