Basically, in exchange for enormous power savings and higher speed, the chip doesn't guarantee each of its billions of individual results will be correct -- only like 90% of them will, and even then they'll be mostly correct. This can be very useful for ray tracing, sound processing, voice recognition, and other real-world applications that can normally handle a certain amount of noise in the signal. It's not likely you can build an entire CPU around this technology, of course; a CPU that ran the wrong subroutine would be nonviable. The technology would be more of a specialized GPU for smartphones.

So let me put it this way: a current microprocessor is like the kid who studied for the test or months, takes the time to read each question, and marks the correct answer on his scantron. The inexact processor is the kid who woke up 15 minutes before the test and just marked "C" all the way down the scantron. Sure, the latter got it done faster and didn't waste any time doing any of that stupid "reading" or "studying"...

Now... if you run the information through the processor 5 times, will you be able to eliminate the errors while still having it run 3x as quick and efficient? I'm not a big fan of 7% data corruption. 1 in 14 is not an acceptable error rate when it comes time to electronically file my taxes (as one example).

Oh right, please sign me up. I've always wanted a computer that I can't diagnose because I'm not sure if there's an error with the software or if the computer just plain farked up.

I wish my boss would let me work faster, use less energy but make a few mistakes. I've already got the use less energy and make mistakes parts down pretty well.

Lord of Allusions: Was it invented by Michael Cera?

There's always money in exact computing.

doglover: LincolnLogolas: So let me put it this way: a current microprocessor is like the kid who studied for the test or months, takes the time to read each question, and marks the correct answer on his scantron. The inexact processor is the kid who woke up 15 minutes before the test and just marked "C" all the way down the scantron. Sure, the latter got it done faster and didn't waste any time doing any of that stupid "reading" or "studying"...

Basically.

But that's the feature. You want a chip like this for certain applications where you'd rather have longer battery life than HD graphics in 1080p, So many modern programs are run perfectly out the nth when they could just be fudged.

If you want a 100 on the scantron test, sure go with the perfect kid. But if your goal is to get the best possible football team, you'd like your starting quarterback to just mark it all C and get back to bed so he's fresh for the big game.

So in other words this is the perfect chip for government employees?

So it's a Pentium 4?

You can only buy it at a general store.

BigJT: Now... if you run the information through the processor 5 times, will you be able to eliminate the errors while still having it run 3x as quick and efficient? I'm not a big fan of 7% data corruption. 1 in 14 is not an acceptable error rate when it comes time to electronically file my taxes (as one example).

No- I haven't read the paper but it sounds from TFA like what they're doing is just removing some of the least-used hardware from the chips. This makes it incapable of handling certain cases that inevitably arise, so in these cases they just produce junk values.

Think, for example, about the overflow detection and correction hardware in a regular adder. The vast majority of the time you're adding extremely small numbers together, and you never come anywhere close to the physical limit of the hardware, but you still have to produce and power that overflow detection hardware. You just cut it out and say that you're OK with your chip returning junk values in the case of overflow. It's easy to see how this kind of thing could save a ton of processing time and money for hardware, to say nothing of the crazier things like branch prediction.

The trick is they're only targeting these chips at applications that can tolerate a lot of noise in the first place, i.e. video processing. You'd be surprised at just how bad a video image can get before people start getting really upset. In fact, sometimes people specifically want to add random noise to images specifically so they look better than their exact counterparts, such as in dithering or randomized anti-aliasing.

The great thing about this technology is how angry and confused it makes certain people*.

*that think they understand computers a lot better than they actually do.

Fubini: No- I haven't read the paper but it sounds from TFA like what they're doing is just removing some of the least-used hardware from the chips. This makes it incapable of handling certain cases that inevitably arise, so in these cases they just produce junk values.


Wonder if they could just shut those parts of the chip down for energy savings, rather than deleting them entirely?

Cthulhu_is_my_homeboy: Fubini: No- I haven't read the paper but it sounds from TFA like what they're doing is just removing some of the least-used hardware from the chips. This makes it incapable of handling certain cases that inevitably arise, so in these cases they just produce junk values.


Wonder if they could just shut those parts of the chip down for energy savings, rather than deleting them entirely?

You can, and there are new chips that take this approach. The problem is that you then need a substantial amount of hardware to control that ability to switch on or off certain parts of a chip. The result is that we have things like multi-core chips where you can selectively do that (called dynamic frequency and voltage scaling) per core or per collection of cores.

The amount of hardware they're looking at removing in this kind of setup would be a handful of transistors here and there, so you can't justify the power control hardware that allows you to selectively enable and disable parts of chips. There are also concerns about the amount of power it takes to put circuits to sleep or wake them back up, and the additional latency introduced.

Cutting out a handful of transistors here and there doesn't sound like a lot, but they're inevitably going to be targeting massively parallel approaches with this, so saving a handful of transistors per core makes a lot of sense when you're working with hundreds of cores on a chip.

I think this could work, but I'm just guessing.

I'm looking at the article on a 37 inch, 1080p monitor, with 20-20 vision, and until I blow up the images I can't see more than the tiniest differences in the quality of the first two images. It becomes much more apparent when I enlarge it, and the 3rd image is craptastic either way, but on a phone screen even version 3 would probably be fine. It would depend very much on which data it allows errors on.

Like Terrified Asexual Forcemeat said, some applications can tolerate some noise. Maybe on audio it might even add some ambiance. I can see sound engineers swearing it warms the sound like an old tube amp. :)

Terrified Asexual Forcemeat: Basically, in exchange for enormous power savings and higher speed, the chip doesn't guarantee each of its billions of individual results will be correct -- only like 90% of them will, and even then they'll be mostly correct. This can be very useful for ray tracing, sound processing, voice recognition, and other real-world applications that can normally handle a certain amount of noise in the signal. It's not likely you can build an entire CPU around this technology, of course; a CPU that ran the wrong subroutine would be nonviable. The technology would be more of a specialized GPU for smartphones.

Just think of the fun with GPSs.

"Stay on current road for 1.2 miles, then bear left. or right."

>10 PRINT "HELLO WORLD!";
>20 GOTO 12

The addledoo CPU?

Good enough for me.

Fifteen times more efficient? Seems like they could run the same logic multiple times, get an average that would be acceptable, and still out perform the 'exact' chips.

Just wait until certain people hear about getting a neural network out of a local minima/maxima.

Hit nearly any key to continue.

The design improves power and resource efficiency by allowing for occasional errors.

Um, errors are bad. M'kay?

xaveth: Hit nearly any key to continue.

CSB, I have to use a command-line tool at work that displays a printout of results and then says "Press Any Key To Continue". You can push any key all day long and nothing will happen until you hit enter or space.

What better way to demostrate it than on a photo of a retarded construction worker. Derp, derp.

This is step one in turning machines into people.

Fubini: You'd be surprised at just how bad a video image can get before people start getting really upset. In fact, sometimes people specifically want to add random noise to images specifically so they look better than their exact counterparts, such as in dithering or randomized anti-aliasing.

Isn't that one of the reasons people didn't like the 48 fps preview of "The Hobbit", because it looked "too smooth"?

Also, I think the addition of noise can allow for better detection of threshold signals, but I might be yakking out my ass.

LincolnLogolas: So let me put it this way: a current microprocessor is like the kid who studied for the test or months, takes the time to read each question, and marks the correct answer on his scantron. The inexact processor is the kid who woke up 15 minutes before the test and just marked "C" all the way down the scantron. Sure, the latter got it done faster and didn't waste any time doing any of that stupid "reading" or "studying"...

No, it's more like the difference between:

- The commenter who read TFA and knows what it says
- The commenter who skimmed TFA and has a good idea what it says
- The commenter who DNRTFA and said something ignorant

Guess which one you are.

BigJT: Now... if you run the information through the processor 5 times, will you be able to eliminate the errors while still having it run 3x as quick and efficient? I'm not a big fan of 7% data corruption. 1 in 14 is not an acceptable error rate when it comes time to electronically file my taxes (as one example).

Did *anyone* RTFA? The chip discriminates acceptable error rates based on the type of calculation being done. It munges your YouTube videos, not your taxes.

theorellior: Fubini: You'd be surprised at just how bad a video image can get before people start getting really upset. In fact, sometimes people specifically want to add random noise to images specifically so they look better than their exact counterparts, such as in dithering or randomized anti-aliasing.

Isn't that one of the reasons people didn't like the 48 fps preview of "The Hobbit", because it looked "too smooth"?

Also, I think the addition of noise can allow for better detection of threshold signals, but I might be yakking out my ass.

fark'em. I say standardize on 60 FPS. It's my gaming frame rate threshold.

The All-Powerful Atheismo: Lord of Allusions: Was it invented by Michael Cera?

There's always money in exact computing.

There's always money in all kinds of high precision applications, but not everyone needs them all the time. Your doctor uses some Star Trek thing to take your temperature; you use a $2 thermometer from Walmart and it's good enough. Your doctor needs that precision all the time; you don't.

Hand Banana: The design improves power and resource efficiency by allowing for occasional errors.

Um, errors are bad. M'kay?

I feel sure someone close to you has said this at some point.

assjuice: What better way to demostrate it than on a photo of a retarded construction worker. Derp, derp.

There's a special irony in your comment. And I do mean special.

Sylvia_Bandersnatch: BigJT: Now... if you run the information through the processor 5 times, will you be able to eliminate the errors while still having it run 3x as quick and efficient? I'm not a big fan of 7% data corruption. 1 in 14 is not an acceptable error rate when it comes time to electronically file my taxes (as one example).

Did *anyone* RTFA? The chip discriminates acceptable error rates based on the type of calculation being done. It munges your YouTube videos, not your taxes.

Exactly!!!

I can see this tech becoming a part of future multi-core processing, as a core for 'things that can use it', but with other 'specialist' cores working congruently as 'detail refiners', on 'things that would crash with such inaccuracy'

/would love more power efficiency on day2day tasks, with the power to 'Shoop da Woop' or game.

HoratioGates:

I'm looking at the article on a 37 inch, 1080p monitor, with 20-20 vision, and until I blow up the images I can't see more than the tiniest differences in the quality of the first two images.

Seriously, I'm looking at them on a cheap Samsung 18.5" monitor and the differences are obvious. The second image looks like somebody did a bad Photoshop smudge job. Look at the cheeks. Look at the white dots on the coat. Look at the broken tooth. Look at the headband inside the helmet. These are not tiny differences.

And I wear glasses that are three years old (and I REALLY need them replaced), but my problem tends to be long range vision.

/ I guess I've looked at too many Photoshops in my time. You can tell by the pixels!

This tech thread is as good a place as any to post this: WTF does my mom's laptop keep refreshing whatever webpage it is on (I obviously have it on FARK right now)? Every minute or two, it refreshes. If I have the curser on something, then the webpage changes from the FARK page to that other page.

This is farking annoying as hell. I don't notice it on any computer except hers. HTF do I make it stop?

/ using firefox

Terrified Asexual Forcemeat: Basically, in exchange for enormous power savings and higher speed, the chip doesn't guarantee each of its billions of individual results will be correct -- only like 90% of them will, and even then they'll be mostly correct. This can be very useful for ray tracing, sound processing, voice recognition, and other real-world applications that can normally handle a certain amount of noise in the signal. It's not likely you can build an entire CPU around this technology, of course; a CPU that ran the wrong subroutine would be nonviable. The technology would be more of a specialized GPU for smartphones.


Could you restate that as if you were talking to a small stupid child? Or an MBA?