Complex Calculations?

[AustinB]

I'm genuinely interested why huge company's(Facebook.etc) need many GPU powered server (Intel Phi, nVidia Tesla) to do these "complex calculations"?

I just get really confused  thinking about it.

 

What do these equations look like?

What do they do?

How long are they?

How complex are they?

Why would you need one?

 

Sorry if this is the wrong section if not thanks.

[Builder]

I'm genuinely interested why huge company's(Facebook.etc) need many GPU powered server (Intel Phi, nVidia Tesla) to do these "complex calculations"?

I just get really confused  thinking about it.

 

What do these equations look like?

What do they do?

How long are they?

How complex are they?

Why would you need one?

 

Sorry if this is the wrong section if not thanks.

It's hard to explain. It has to do with statistics and data analysis, which can be broken down into small tasks that need to be done again and again and again.

 

The Xeon Phi and Tesla products allow massive parallelism, so all of these smaller tasks can be done at once, greatly speeding up number crunching.

 

Basically, they have massive amounts of data, so they need massive amounts of processing power. The analysis they're doing can be done by hand with smaller numbers, but they do it by the billions. The difficulty does not scale linearly.

[AustinB]

It's hard to explain. It has to do with statistics and data analysis, which can be broken down into small tasks that need to be done again and again and again.

 

The Xeon Phi and Tesla products allow massive parallelism, so all of these smaller tasks can be done at once, greatly speeding up number crunching.

 

Basically, they have massive amounts of data, so they need massive amounts of processing power. The analysis they're doing can be done by hand with smaller numbers, but they do it by the billions. The difficulty does not scale linearly.

 

Like Multi-threading at a huge scale?

[Builder]

Like Multi-threading at a huge scale?

Precisely.

[Shaqo_Wyn]

I'm genuinely interested why huge company's(Facebook.etc) need many GPU powered server (Intel Phi, nVidia Tesla) to do these "complex calculations"?

I just get really confused  thinking about it.

 

What do these equations look like?

What do they do?

How long are they?

How complex are they?

Why would you need one?

 

Sorry if this is the wrong section if not thanks.

Probably comes in handy when they're trying to find patterns in all personal information we willingly give em. 

[AustinB]

Probably comes in handy when they're trying to find patterns in all personal information we willingly give em. 

Our Patterns to give us suggestions, basically use the information to sell for marketing and experiments...

 

 

Precisely.

 

 

How long would these equations be?

[TideRower]

How long would these equations be?

What do you know about statistics?

Even the arithmetic mean can be a long equation if you have lots of data.

 

I don't really know what they do, but I don't think that even with their amount of data they'd need much processing power just for statistics. I have never used more than maybe 100 numbers as data, but I also just used a single core and it was really fast all the times. But maybe they just need to do it all the times in a matter of seconds.

They might use it for simulations, i.e. differential equations, which can require a lot of processing power.

[LukeTim]

Our Patterns to give us suggestions, basically use the information to sell for marketing and experiments...

Yes, exactly. As I understand it, Facebook in particular does colossal amounts of graph analysis, because that's really what everyone's network of friends boils down to... one huge graph. And graph analysis on the scale of a billion nodes can take a lot of power to crunch through in a reasonable amount of time to be useful to you as a user... or to facebook as advertisers. Luckily, though, this analysis is not too difficult to parallelize. I had a friend who studied this at University, using an on-chip network of about 18 ARM11 CPUs to perform graph analysis in a very parallel manner.

 

There's also other things, though, that take some real power to do... like Facebook's facial recognition system which is based on neural networks. In a similar fashion to the graph/network analysis, neural networks can be very computationally intensive when they grow in complexity, but they are also highly parallelizeable.