They tried but failed, due to the inherent heterogeneous overhead and programming complications. With the Kepler architecture they're focusing on graphics again, which is where the money is for them, and they've taken a serious step back from consumer GPGPU. AVX-512 instead is homogeneous, which opens up a whole new world of possibilities.
I could list a bunch of things, but that would just be the tip of the iceberg and not do it justice. Don't look for one specific killer app. There are three different kinds of parallelism for increasing performance beyond clock speed scaling: ILP, TLP, and DLP. ILP is pretty much maxed out, TLP you get from multiple cores, and DLP is most efficiently extracted using vector instructions. So although the seed had already been planted with AVX, AVX-512 will really add another dimension to CPU performance as a whole, and not target one specific killer app.
Think about how superscalar execution and multi-core have transformed computing. There is no single killer app for them, but you sure don't want to go back to single-issue or single-core. The same thing will happen with wide vectors. A lot of applications will benefit. Some more, some less, but you'll soon be wondering how we ever lived without them.
Of course it will take many years. The same was true about multi-core in 2005, and there are still software companies that only recently started looking at it. TSX should help, but won't be ubiquitous for many years either (especially since indeed Intel segmented support for it). But we still think of multi-core and transactional memory as revolutionary.
So you shouldn't look at the slowness of the market to determine whether something is revolutionary or not. AVX-512 can execute loops up to 16 times faster than legacy 32-bit code. That's not going to leave things unchanged. That's a revolution.
Facebook
Twitter