Hacker News new | ask | show | jobs
by wbhart 2875 days ago
They've hired heavily on the GPU side and announced a partnership with AMD (presumably to use their infinity fabric).

This'll sound a bit strange at this point, but I have the feeling CPU's will become little more than an I/O controller in the long term. Much consumer computing will be done in the data centre. Data centres will be dominated with devices that look more like GPUs and TPUs than CPUs.

Naturally, GPUs will inherit more CPU like features before this happens.
1 comments
skybrian 2875 days ago
If it runs all of today's software with at most a recompile, that sounds like a CPU to me.
link
wbhart 2875 days ago
There's no way today's software can be scaled in terms of performance in the face of the "end of Moore's law", with a mere recompile. Today's software will be displaced by tomorrow's software. Of course I am talking fairly long term here. But I believe all major players are making very definite moves in this direction.

Major hardware companies have to have fairly long roadmaps, because they have so much inertia. And they have to reveal what they are up to, to their investors. I certainly don't believe these roadmaps reveal another couple of decades of die shrinks and IPC gains. They reveal a real shift of strategy, across the board. Consumers simply won't be able to afford the "CPUs" of the future.

link
skybrian 2875 days ago
I don't expect renting computers in the cloud to get more expensive. Do you? The end of Moore's law means slower improvements, not that things will actually get worse. If anything, it means computers in data centers need to be replaced less often, so on a per-hour basis, they're still cheaper.

For a scalable application, better performance basically means reducing expenses. If your cloud computing bills aren't high to begin with, it may not be worth the rewrite. Of course there might be new companies or new projects that can use TPU's for machine learning, etc.

But Moore's law is not the only way to scale better. Today's machine learning algorithms are ridiculously inefficient and that seems unlikely to remain true forever, given the amount of research being done. A series of algorithmic improvements might result in a 10-100x reduction in cost, or maybe even not needing TPU's anymore?

Who knows what the future will bring, but making straight-line predictions in a fast-moving field like machine learning seems unlikely to work out.

link