Larrabee crap

Yesterday, Patrick Gelsinger, who is one of the lead engineers at Intel, gave us a hint on upcoming Intel processors and architectures. For the most part it was of course quite exciting to learn more about the direction that Intel is taking for the future of their processor designs. Indeed, it seems they are completely embracing the SIMD philosophy that they were already pursuing with their Streaming SIMD Extensions (SSE4 for short). See the relevant articles on c|net here.

This is something which of course I think it the right way because vector processing is a well-established technique for scientific and visual computing. I am glad Intel agree.

The only problem is of course with the implementation. Intel being Intel, they are implementing their great new idea the Intel way: that is, very stupidly by stacking a new set of instructions-this time called AVX-on top of their already horrendously messy x86 architecture. This is very much like building a faster and faster car each time by adding one more engine-and extra wheels-to the new iteration, complete with extra pedals and clutches inside the cabin. The bottom line for the programmer is, your get more power with every new model but the driving experience is really starting to resemble a circus exercise.

Moreover, since nVidia are becoming all too powerful with their incredible GPU's and since AMD have also bought ATI, with the intention of developing a "Fusion" chip with both the CPU and a GPU on the same die, Intel are now designing their own GPGPU (that stands for General-Purpose GPU), code-named Larrabee. The idea is excellent in principle but once again they are completely ruining it (not to use a less PC word to describe it) by building it upon the decades-old x86 architecture. Supposedly because programmers already know how to use the x86 instructions and it's too much work for them to learn something new. Poor things!

I say this is totally idiotic. For one thing, these two new architectures, AVX and Larrabee, won't be widely available for another 2-3 years at least so there's plenty of time for people to learn to use it. But more importantly, now that Intel are readying 6- and 8-core CPU's anyway, they could just as well make chips with a couple of x86 cores and an extra say, 4 Itanium cores, all on the same die. And that's what I find totally baffling about it all: Intel have already spent all this money developing this incredibly powerful Itanium chip, which is quite a neat design overall, but now they're totally giving up on it. Honestly, I think keeping the x86 architecture on life support like this is inhumane. I'm not generally in favor of euthanasia but in some exceptional cases it can be a good thing. Just let x86 rest in peace, finally!

Sayonara!