Saturday, December 05, 2009

barking up the wrong tree, good money after bad, and the Intel's Larrabee

A few years back, just before Apple's official preannouncement of the iPhone, Intel sold off their XScale assets, ostensibly to concentrate on their x86-64 architecture (although they continue to move the Itanium forward at a modest pace, and hold the rights to other ISAs they've never brought to market).

But concentrate they have, convincing most naysayers they could indeed squeeze a good deal more performance out of their flagship processor line, with even greater improvements in performance/watt.

Larrabee, intended to be a GPGPU, and originally to have begun shipping about now, in 24-core and 32-core versions, was to have been the next stage in this x86-everywhere strategy, by combining many, relatively simple x86-64 cores with a single, very wide vector processing unit.

The problem with this strategy is that the x86-64 architecture isn't optimized for such use, and the advantage it would have enjoyed a few years ago, due to instruction compatibility with the chips powering nearly every laptop and desktop computer, has been rendered largely moot by developments in software (OpenGL, CUDA, OpenCL, and Apple's Grand Central Dispatch, recently turned open source).

On the other hand, Intel's investment hasn't necessarily been wasted. They can take what they've learned in developing Larrabee and turn it into improved x86-64 processors and improved compilers to generate code for them. The 512-bit wide SIMD unit may find its way into dual, quad, and eight-core chips, making them radically faster for some purposes, in many cases obviating the need for dedicated GPU hardware.

x86-64 has a lot of life left in it, and folding the Larrabee project back into the mainstream processor line would help extend that life span far into the future.

No comments: