Sacrificing Accuracy For Speed and Efficiency In Processors

Skudd writes "Modern computing has always been reliant on accuracy and correct answers. Now, a professor at Rice University in Houston posits that some future applications could be revolutionized by 'probabilistic computing.' Quoting: 'This afternoon, Krishna Palem, speaking at a computer science meeting in San Francisco, will announce results of the first real-world test of his probabilistic computer chip: The chip, which thrives on random errors, ran seven times faster than today's best technology while using just 1/30th the electricity. ... The high density of transistors on existing chips also leads to a lot of background "noise." To compensate, engineers increase the voltage applied to computer circuits to overpower the noise and ensure precise calculations. Palem began wondering how much a slight reduction in the quality of calculations might improve speed and save energy. He soon realized that some information was more valuable than other information. For example, in calculating a bank balance of $13,000.81, getting the "13" correct is much more important than the "81." Producing an answer of $13,000.57 is much closer to being correct than $57,000.81. While Palem's technology may not have a future in calculating missions to Mars, it probably has one in such applications as streaming music and video on mobile devices, he said.'

Comments

Bank balance

[johnny cashed]

And $81,000.31 is a much more correct answer!

Re:Bank balance

[CastrTroy]

I agree. The whole problem with the example given in the summary is that your bank balance should never be wrong. There is no room for error in calculating bank balances. I also don't want to hear skips and pops in my music because they though it would be more energy efficient to use a processor that produced errors. I already get 26 hours of charge out of my MP3 player. I'd rather have them focus on getting more space for cheaper so I can carry lossless audio on my portable mp3 player.

Re:Bank balance

[BadAnalogyGuy]

If you are listening to music on a portable media device, it's safe to say that you aren't going to be able to hear the difference between the lossy format and the lossless format.

It's like drinking from a well. Connoisseurs may claim to be able to taste the difference between it and tap water, but that's just the extra tang from all the bull shit.

Re:Bank balance

[phoenix321]

High accuracy is required for encoding music and video, though.

Maybe we could have a selective accuracy, where programmers can set their needs via registers or direct it to different CPU cores. Accurate cores for banking transactions, AV-stream encoding and 2D GUI operations while inaccurate cores are used for AV-stream decoding, and computer game 3d drawing and AI decisions.

There's a whole lot we are calculating now without the need for more than 3 significant digits - and a whole bunch where we intentionally use random numbers, sometimes even with strong hardware entropy gathering.

These are all cases where we could just scrap the accuracy for faster processing or longer battery times. No one cares about single bit errors in portable audio decoders or in high fps 3d gaming.

Re:Bank balance

[Hojima]

There are countless applications for a computer that don't depend on accuracy, but do depend on speed. For example: gaming, stock analysis, scientific/mathematical research etc. Just about every use for the computer can benefit from this. Bear in mind these applications can take the hit of inaccuracy, if not benefit from it depending on the situation. Yes there are some instances were accuracy is crucial, but that's why they will continue to make both of the processors. It's what they call a free market, and there will be always be a new niche to fill.

Re:Bank balance

[memco]

Ok, so not only do I have to give up efficiency in the chip itself, but now my efficiency suffers because I now have to determine which chips are useful for which applications. I don't want to have to start thinking about whether or not I plan to use my new laptop for anything requiring accuracy greater than such and such a percentage. I suppose this might be effective for niche markets, but it seems messy if you try to make it part of all computing platforms.

Re:Bank balance

[shaitand]

'It's like drinking from a well. Connoisseurs may claim to be able to taste the difference between it and tap water, but that's just the extra tang from all the bull shit.'

Probably not the best example. Humans have an amazing ability to taste very minute differences in water. My TDS meter tells me that tap water here is extremely pure to begin with, but I can pick the same that has undergone carbon and ro filtering versus straight tap water in a blind taste test with 100% accuracy. I'm certainly no connoisseur.

Actually, I'm from rural Illinois, and all the water be it tap or properly maintained well is fairly sweet there with minimal filtering. Actually the streams there are a bit muddy tasting but the water itself is sweet as it flows. It definitely beats this Florida swap water. I tasted unfiltered Florida well water once (most Florida wells have filters built in) and I vomited. The tap water here won't make you sick and it isn't that nasty but it still tastes funky.

That said, I doubt I could tell the difference between tap, well, Illinois, or Florida water that has had that additional filtering (Carbon and Reverse Osmosis, any of those machines for $0.39/gallon at the grocery store will do). My TDS meter shows a difference in purity even from one dispensing machine to the next, but I can't taste that difference. Whatever minerals survive that process are probably pretty much the same anywhere and taste good. That filtered water tastes better than any of the unfiltered waters.

Re:Bank balance

[ZombieWomble]

I don't think any of these things particularly benefit from this type of processor. All of the situations you describe involve some degree of randomness, but this CPU doesn't sound like a source of useful random values at all.

The randomness in these processes occurs in particular places, in particular quantities - this processor presumably produces some characteristic amount of randomness in each calculation, but the odds on it being a meaningful amount for whatever arbitrary calculation you're doing is vanishingly small - and given it's apparently treated to give different amounts of randomness in different bits, it's almost certainly non-uniform as well.

In almost all simulations you want to make use of extremely well-controlled random numbers - something which adding some noise as part of your floating-point calculations is not.

Reminds me of...

[rob1980]

Q: Why didn't Intel call the Pentium the 586?
A: Because they added 486 and 100 on the first Pentium and got 585.999983605.

Re:Reminds me of...

[machine321]

My computer's not slow, it's just being careful.

Accuracy with financial calculations.

[onion2k]

Accuracy with financial calculations is extremely important. Hasn't this guy ever watched Superman 3?

Re:Accuracy with financial calculations.

[Firethorn]

quite correct the thousands is far more important than the cents, however 13,810.00 is really close to 13,000.81 right. it has all the same numbers in a similar order.

Not by the fuzzy logic the guy's using. He's going for scientific accuracy. IE 13,000.81 (+-.001%). It's just our brains that compare symbols that would consider those numbers 'close'.

In which case a $810 error in a $13k account is a big friggen error, and would violate the standards of the chip he's working on. Now, I don't know HOW he's making sure high order bits are done more accurately than the low order ones, but that's what the article mentions him doing.

wll,

[greenguy]

i scrfcd accrc 4 spd a lng tm ago

Re:wll,

[hackstraw]

This whole thing is old and silly.

Seymour Cray is known for saying "Do you want fast or accurate?" because back then there was no IEEE 754 spec (which is not infinitely precise) for floating point numbers at the time and machines were pretty primitive then and his machine did Newtonian approximations of many numeric calculations that were accurate to a point, just like John Carmack did (in software) with Doom's inverse square root.

The moral of the story is that in 2009 and beyond its probably best to have hardware continue to be accurate. This is why we have digital 1s and 0s instead of some other base of computation.

Now, in software, feel free to make things as sloppy as you want. If your bank (not mine) wants to round 13,000.83 to some other value, then by all means go for it. But I think that most of us are OK with accurate hardware.

Re:wll,

[vux984]

i scrfcd accrc 4 spd a lng tm ago

and it was going so well too... until you got thirsty and told your friend ..

"hy! I wnt sm ck!"

Top Ten Slogans

[AmigaHeretic]

TOP TEN SLOGANS FOR THIS NEW PROCESSOR:

9.9999973251 - It's a FLAW, Dammit, not a Bug

8.9999163362 - It's the new math

7.9999414610 - Nearly 300 Correct Opcodes

6.9999831538 - "You Don't Need to Know What's Inside" (tm)

5.9999835137 - Redefining the PC -- and Mathematics As Well

4.9999999021 - We Fixed It, Really

3.9998245917 - Division Considered Harmful

2.9991523619 - Why Do You Think They Call It *Floating* Point?

1.9999103517 - We're Looking for a Few Good Flaws

0.9999999998 - "The Errata Inside" (tm)

Re:Top Ten Slogans

[gbjbaanb]

TOP TEN SLOGANS:

runs Excel just as well as always :-)

It seems like when you need a precise calculation

[rolfwind]

like financial things, you compare the end product with the end product of the same calculation run either through the chip again or another chip (or increasingly likely another core).

Still would be faster too.

Not completely correct

[chthon]

If you read the chapter about the history about the IEEE FP standard in Microprocessors : a quantitative approach, then you will see that in the past accuracy was already sacrificed for speed in supercomputers.

gfx

[RiotingPacifist]

can't this be used in gfx cards, i mean with anti-aliasing and high resolutions it doesn't really matter so much if 1/2 a pixel is #ffffff or #f8f4f0 , hell you can probably even get a pixel entirely wrong for one frame and nobody will care (as long as it doesn't happen too often).

Re:Hmmm

[BadAnalogyGuy]

You get a one-way ticket to pound-me-in-the-ass prison.

Watch out for your cornhole, bud.

Sacrifices are expected

[LiquidCoooled]

I have spent the last 9 months coding up a dynamic scalable UI for the nokia tablets.

I have had to make huge compromises to accuracy to obtain the desired performance.

I had the choice of using the full featured (but slow) widget sets and graphical primatives which existed already, or find a way to make it work as I expected it to.

The results have left people breathless :)

take a look here: http://www.youtube.com/watch?v=iMXp0Dg_UaY

Re:Sacrifices are expected

[ljw1004]

That was a very confusing video. What I learned from it: you haven't done some stuff, Zoom Fish!, widgets, Zoom Fish!, behind schedule, zoom, Fish!, widget framework, Fish!

I guess it's a system that lets you zoom in on fish?

Nothing new here...

[chill]

Isn't that essentially what JPEG, MPEG and every other lossy codec or transform does?

Re:Nothing new here...

[chill]

There have been dedicated MPEG encoding and decoding chips for many years. DXR3 comes to mind.

I think the only new twist is applying the idea to general calculations as a whole as opposed to a specific function or set of functions in software. An interesting idea. Maybe we'll end up with double-precision, single-precision and ballpark floats.

Games?

[ndogg]

I could definitely foresee this being used in game systems, especially for graphics.

As long as it mostly looks right, that's all that really matters.

Re:Hmmm

[PitViper401]

But what about the conjugal visits?!

Re:uhhh....

[Dogtanian]

NEWS FLASH: Binary consists of 1 and 0.

NEWS FLASH: People use computers for calculations with more than single-digit binary results.

"probablistic computing" is another way of saying "sloppy engineering".

No- insisting on excessive precision where an "almost certainly right to within +/- x%" solution would be more than good enough and much simpler to obtain is known as overengineering.

I suspect that the financial examples chosen didn't illustrate the point as well as intended (financial companies generally don't like *any* inaccuracy), but that doesn't change the general principle.

Would you prefer a routing algorithm that gobbled up power and took ages to run for a guaranteed shortest route or one that was far more efficient and 99.9% certain to give a route that was within 3% of the shortest possible distance?

The first thing that comes to mind...

[Lordfly]

...is gaming applications.

Programmers spend a lot of time coming up with algorithms that simulate randomness for AI or cloud generation or landscapes or whatever... if the processor was just wonky to begin with it'd make certain things a lot more natural looking.

It's interesting that AI in games is always touted as being "ultra-realistic", but always ends up being insanely easy to trip up. Having something "close enough" would add just enough realism/randomness to situations to perhaps make games and environment more dynamic.

I wouldn't want these things processing my bank balance, though, unless it rounded up.

Re:DSP's?

[Firethorn]

I think the point would be using a DSP/math coprocessor that uses 1/30th the power in exchange for a .001% loss in accuracy for non-essential tasks like music decoding.

I mean, combined with the lousy earbuds most people use, who'd notice? Especially if it makes their MP3 player last 3 times as long as ones that use more traditional and technically accurate DSP/decoder?

Re:Use in MP3 Players

[phoenix321]

If you bought a popular artist recently, your music is autotuned already.

Anyway, this means that less than 0.1 percent of your decoded audio samples will be less 0.1 percent off. This is probably very acceptable outside concert halls and living rooms if it delivers large bonuses in battery saving or calculation speed.

For example, we could use a much beefier compression algorithm than MP3 or current algorithms even longer on even smaller devices.

Obligatory NASA research

[DirePickle]

From NASA:

Computer scientist Arthur Boran was ecstatic. A few minutes earlier, he had programmed a basic mathematical problem into his prototypical Akron I computer. His request was simply, "Give me the sum of every odd number between zero and ten." The computer's quick answer, 157, was unexpected, to say the least. With growing excitement, Boran requested an explanation of the computer's reasoning. The printout read as follows: THE TERM "ODD NUMBER" IS AMBIGUOUS. I THEREFORE CHOOSE TO INTERPRET IT AS MEANING "A NUMBER THAT IS FUNNY LOOKING." USING MY AESTHETIC JUDGEMENT, I PICKED THE NUMBERS 3, 8, AND 147, ADDED THEM UP, AND GOT 157.

A few moments later there was an addendum: I GUESS I MEANT 158.

Followed shortly thereafter by: 147 IS MORE THAN 10, ISN'T IT? SORRY.

Bad example...

[Firethorn]

It looks like you got sucked into the bad example land. Later on in the article it mentions that it's intended for stuff where accuracy isn't paramount, but where it's not really necessary. Multimedia applications over space/bank calculations.

I mean, there's 1.764 Million pixels in my screen that I'm typing my post on at the moment. Does it really introduce much error if I round it to 1.8M? I'm also running at 60Hz. Do you think that I'd really notice if there's a .01% chance that instead of getting white(255) I get white(254)? That'd be an average of 176 pixels a refresh, assuming an all-white screen. Thing is, those pixels wouldn't be the same every time. Then again, logically each pixel would tend towards red/blue/yellow depending on the error. But only slightly. In a HD movie, are you really going to notice?

Re:A Little Bit of History Repeating

[osu-neko]

Sorry, I missed the part back in the 80's where using fuzzy logic caused my processor to consume 1/30th the power.

Article badly misrepresents the idea

[Jimmy_B]

The author of the linked article has completely misunderstood what this research is about. It is NOT about tolerating errors in the output of computations; that would be completely infeasible. It's about tolerating errors in intermediate values, by using redundancy. For example, three adders made out of unreliable transistors plus a control unit to have them vote, may be smaller and use less power than one adder made out of reliable transistors. However, you can't make everything out of unreliable transistors. In particular, the control unit, and the parts that compare results to each other, have to work reliably and can't be duplicated. That is what is meant by "some information was more valuable than other information", not the low-order bits of a numeric computation.