Article [engadget] An algorithm could make CPUs a cheap way to train AI

[Hitman928]

An algorithm could make CPUs a cheap way to train AI

AI is the backbone of technologies such as Alexa and Siri -- digital assistants that rely on deep machine learning to do their thing. But for the makers of these products -- and others that rely on AI -- getting them "trained" is an expensive and often time-consuming process. Now, scientists...

www.engadget.com

Researchers say a new algorithm could make CPUs competitive again for neutral network training.

 

[eek2121]

An algorithm could make CPUs a cheap way to train AI

AI is the backbone of technologies such as Alexa and Siri -- digital assistants that rely on deep machine learning to do their thing. But for the makers of these products -- and others that rely on AI -- getting them "trained" is an expensive and often time-consuming process. Now, scientists...

www.engadget.com

Researchers say a new algorithm could make CPUs competitive again for neutral network training.

I hear Python does a good job at this as well.

 

[moinmoin]

Researchers say CPUs can be used universally and algorithms can be programmed to do anything you want, news at eleven!

 

[cytg111]

Skynet... its Skynet

 

[thecoolnessrune]

Top-of-the-line GPU Deep Learning Platforms cost $100 grand because they can sell them for that. Just like with the Xeon Platinum CPUs that recently took a price haircut, if they could sell those for $50K a piece they would. If this somehow "takes off" and CPUs become relevant for deep learning again, expect capable CPUs to start increasing in price along with it.

 

[Hitman928]

Researchers say CPUs can be used universally and algorithms can be programmed to do anything you want, news at eleven!

The point isn't that CPUs can be used to train NN, they've been used since the beginning. The point is that currently, GPUs and specialized hardware are dominating the deep learning scene because they can do it so much faster than a CPU. This research suggests that this new algorithm would allow CPUs to be competitive (if not faster) than GPUs for at least certain types of NN training. This would be a huge deal not only for big time AI companies and hardware vendors, but would make training NNs much more accessible to everyone.

 

[maddie]

The point isn't that CPUs can be used to train NN, they've been used since the beginning. The point is that currently, GPUs and specialized hardware are dominating the deep learning scene because they can do it so much faster than a CPU. This research suggests that this new algorithm would allow CPUs to be competitive (if not faster) than GPUs for at least certain types of NN training. This would be a huge deal not only for big time AI companies and hardware vendors, but would make training NNs much more accessible to everyone.

Exactly. I'm going to look more into this as it could be useful for some stuff I'm exploring.

 

[soresu]

Suddenly those telling me bfloat16 and matmul on CPU are pointless seem somewhat shortsighted.

 

[Hitman928]

Suddenly those telling me bfloat16 and matmul on CPU are pointless seem somewhat shortsighted.

If this research pans out, then yes, expect more DL extensions to be added to CPUs. Maybe they'll have special SKUs for DL models as well, we'll see.

 

[UsandThem]

Skynet... its Skynet

You must have not watched Dark Fate.

Spoiler

Skynet never happened or existed.

 

[moinmoin]

This research suggests that this new algorithm would allow CPUs to be competitive (if not faster) than GPUs for at least certain types of NN training.

Until those algorithms are adapted to FPGAs which blow both CPUs and GPUs out of the water again.

Honestly this area of research has insanely fast cycles. This is a neat snapshot. But beyond including specialized hardware there's simply no way CPUs will be competitive longer than until the next refreshes of GPUs and FPGAs even for certain types of NN training.

 

[soresu]

Until those algorithms are adapted to FPGAs which blow both CPUs and GPUs out of the water again.

Honestly this area of research has insanely fast cycles. This is a neat snapshot. But beyond including specialized hardware there's simply no way CPUs will be competitive longer than until the next refreshes of GPUs and FPGAs even for certain types of NN training.

Again with none of those options being as efficient as custom designed NN ASIC's - Google are even using ML to optimise logic/wire placement in chip design now, I imagine gen 4 of their TPU will be a nice efficiency improvement when combined with whatever high level changes they have planned.

Link to the article about ML logic/wire placement here.

 

[soresu]

An interesting part about the linked article in my post above is the implication that not only can ML placement be better than a humans, it can also be done in hours/days instead of weeks, meaning it could well be responsible for drastically streamlining chip development in an increasingly expensive time of leading edge node shrinks.

 

[ThatBuzzkiller]

https://arxiv.org/pdf/1903.03129.pdf

This is the paper in question the article was talking about. Big limitation of SLIDE algorithm so far is that the current implementation doesn't support convolutional layers which means it can't be used to train CNNs (convolutional neural nets) or any other types of neural networks with convolutional layers ...