Google's Tensor Processing Unit could advance Moore's Law 7 years into the future

[Energycore]

4 minutes ago, Master Disaster said:

But this is a sciencey conversation?

 

Anyway these things look pretty cool, until they get self aware and kill us all skynet style. If anyone is going to be responsible for the machine uprising it will be Google.

Perhaps. We don't have to use it strictly because we're not researchers though.

[captain cactus]

Coming to the consumer market in 2023.

 

When it's 7 years later and we'd get this anyways.

[vanished]

1 hour ago, laminutederire said:

 

Actually an order of magnitude, with the physician sense of the term is actually correct for anything between x9 and something like x50. the order or magnitude just refers to an approximation of the gap you have. it doesn't have to be strictly x10.

 

8 hours ago, LukeTim said:

25x is still on the scale of one order of magnitude... Remember, it's logarithmic. Two orders of magnitude would be 100x or more.

 

Yes, that's all correct but I'm thinking of it this way:  marketers typically want to push right to the limits when making their product sound good.  Since about 10x is the minimum improvement they'd need to get away with calling it an "order of magnitude", I would just be surprised if they'd lump a 25x gain into that same term.  I would think they'd say something like "an amazing 25x improvement" or something like that, if they could

[niofalpha]

17 hours ago, Ryan_Vickers said:

well it's close enough to the 11.3x factor of 7 years' worth of doubling every 2 years.  I just could have sworn the time span was 18 months, not 24.

It was revised in the 70's to be every 2 years.

[laminutederire]

31 minutes ago, Ryan_Vickers said:

 

 

Yes, that's all correct but I'm thinking of it this way:  marketers typically want to push right to the limits when making their product sound good.  Since about 10x is the minimum improvement they'd need to get away with calling it an "order of magnitude", I would just be surprised if they'd lump a 25x gain into that same term.  I would think they'd say something like "an amazing 25x improvement" or something like that, if they could

Yeah but I think they actually don't care Because they know who they're talking to, and those people are going to invest millions of euros of hardware, they're not the kind which is going on the 1080 hype train They won't invest in a product just because it's well advertised.

That piece of hardware is interesting though

[laminutederire]

That will be a huge loss for Nvidia though!

[CtW]

8 hours ago, laminutederire said:

That will be a huge loss for Nvidia though!

That depends on how quickly they can get these to market (for enterprise users) as well as if they can integrate them into existing machines. If these require you to purchase an entire new system and use new software, while nvidia products can just be put into normal systems, then this will have a hard time gaining market share.

[laminutederire]

13 minutes ago, CtW said:

That depends on how quickly they can get these to market (for enterprise users) as well as if they can integrate them into existing machines. If these require you to purchase an entire new system and use new software, while nvidia products can just be put into normal systems, then this will have a hard time gaining market share.

Yeah, but people in machine learning kinda hate Nvidia because of their proprietary shit, and teir lack of response to their demands. They won't necessarily mind purchasing those if they aren't two expensive and can be plugged in their computer as an extra hardware. Instead of having a sli setup, you could imagine having a graphics card for gaming or whatever and having one of these.

As for large setups, they have better performance per watt, and that's something data centers crave for since those can get quite expensive to run and tough to cool down...

(plus I don't like Nvidia so I wouldn't mind )

[vanished]

On 5/19/2016 at 2:33 PM, niofalpha said:

It was revised in the 70's to be every 2 years.

Ah that explains it.  I guess I'm just old... although I wasn't even born in the 70s

[Tech_Dreamer]

kinda good , but at the same time this has a lot of potential to be used for something bad if turned into a weapon, break strong encryption & blink of an eye facial recognition anywhere in the world in real time , pretty much tech like this would be used for things like that at the end

[EmeraldKiwi]

“TPU is tailored to machine learning applications, allowing the chip to be more tolerant of reduced computational precision,..."

 

Wait a minute. This part is slightly confusing. It doesn't make sense to me that you would want to be less accurate with a processor.

 

Does it mean (in normal people words) that maybe instead of calculating things to 50 decimal places, it calculates to 30? I really don't know what I'm trying to say, but it is confusing that they would want to be tolerant of "reduced computational precision". 

[MrDynamicMan]

TL:DR asic chips are good at doing what they are good at doing. Big freakin woop.

 

This has nothing to do with moores law, as their is no node shrink.

[christianled59]

I could see these running in multiple modules (cores) on the same pcb and replacing traditional cpus.