• Some users have recently had their accounts hijacked. It seems that the now defunct EVGA forums might have compromised your password there and seems many are using the same PW here. We would suggest you UPDATE YOUR PASSWORD and TURN ON 2FA for your account here to further secure it. None of the compromised accounts had 2FA turned on.
    Once you have enabled 2FA, your account will be updated soon to show a badge, letting other members know that you use 2FA to protect your account. This should be beneficial for everyone that uses FSFT.

Allowing Errors Makes Chip More Powerful, Efficient

H

HardOCP News

[H] News
Joined
Dec 31, 1969
Messages
0
How do you make a processor that is fifteen times as efficient than today's processors? By allowing for occasional errors.

Researchers have unveiled an “inexact” computer chip that challenges the industry’s 50-year pursuit of accuracy. The design improves power and resource efficiency by allowing for occasional errors. Prototypes unveiled this week at the ACM International Conference on Computing Frontiers in Cagliari, Italy, are at least 15 times more efficient than today’s technology.
Click to expand...
 
and this is how the 3 laws get violated and our robot overlords start to annihilate us!
 
This is about as good of an idea as allowing a high school drop out to operate a nuclear reactor.
 
How is this any different from using GPUs for traditionally CPU-bound purposes? GPUs use the same type of fuzzy math calculations that involve errors, but aren't important for their normal applications. That's why they're a lot better at things like cryptography.
 
Well, wouldnt quantom computing be similar too? The answers isnt always accurate but it is extremely fast?
 
Sounds like the chips used in our Shuttle program.
 
Sir, I'll take that slower accurate chip over there in the bargain bin.
 
saitei said:
How is this any different from using GPUs for traditionally CPU-bound purposes? GPUs use the same type of fuzzy math calculations that involve errors, but aren't important for their normal applications. That's why they're a lot better at things like cryptography.
Click to expand...

In the GPU's they are letting the errors go by. In the CPU data integrity is obviously needed. You can still get performance gains while maintaining integrity. For example, you can lower the voltage swing and account for the additional faults that arise by using a more aggressive error correction scheme(Say going from SECDET to a quadruple scheme or higher). That does add some additional overhead, but it only a very little in the overall consumption. You can achieve a pretty dramatic reduction in the voltage swing with some really high error correction schemes.
 
The article states that these woul typically be for special purpose/embedded application

The big problem I see with these is that if the processor has inexact calculations it might mke systems more vulnerable to hacks
 
Can't they just rebadge the Pentium 5 chip with the FDIV bug and call it "new". All the R&D has already been done and it includes this new "occasional error" feature. ;)
 
Errors and failures are not the same. The RAM in your PC is upsetting all the time. The data on your HDD is getitng corrupted all the time. However, there is correction algorithms to fix the eror and ths the net result is no failures.

Put it tihs way. Lets you need to muliply 2 number together. The chance of failing is 0.001%. If you do it twice...it would be 0.001% of 0.001%. If you do it 3x..0.001% of 0.001% of 0.001%. You can see where this goes. Now, if you craft the CPU bascially do a bunch of multiplies instead of 1 just. Put in a bit of code to check all the multplies and pick the one with the most matches.

If the total energy to do the many vs the one is less...the answer is obvious. Remember, this is occassional errors...not constant errors.
 
1+1 can equal 3 once in a while if it will save me a few bucks on my energy bill. In all seriousness, I wonder how these errors would manifest themselves during normal operation.
 
Anything that makes smartphone batteries last longer will make a lot of people happy.
 
Accept the possibility that you won't know where the electron is inside of a chip and you don't have to worry about Moore's Law.

This is an interesting approach. You simply provision the chip for mistakes and through compensation you don't have to worry about growing electrical leakage or excessive heat. However what kind of compensation engine would it take to allow the chip to function in a state we're all use to? A state without errors?
 
I think it's a possible approach. If it's 15 times more powerful/faster/efficient can't you just use some of that extra power to do redundent checks and still have extra power left over?
 
I like this:
The inexact hardware is also a key component of ISAID’s I-slate educational tablet. The low-cost I-slate is designed for Indian classrooms with no electricity and too few teachers. Officials in India’s Mahabubnagar District announced plans in March to adopt 50,000 I-slates into middle and high school classrooms over the next three years.
Click to expand...

Yo India... it's okay if your kids are sort of smart... but not quite.
 
What dirty hippie thought up this "energy efficient" nonsense? I'd rather have this than t¥is.
awsiemieniec said:
Can't they just rebadge the Pentium 5 chip with the FDIV bug and call it "new". All the R&D has already been done and it includes this new "occasional error" feature. ;)
Click to expand...
I spy a probable fellow gen X'er... ;)

Serpent said:
I think it's a possible approach. If it's 15 times more powerful/faster/efficient can't you just use some of that extra power to do redundent checks and still have extra power left over?
Click to expand...
No. That's what we have now.
 
Trepidati0n said:
Errors and failures are not the same. The RAM in your PC is upsetting all the time. The data on your HDD is getitng corrupted all the time. However, there is correction algorithms to fix the eror and ths the net result is no failures.

Put it tihs way. Lets you need to muliply 2 number together. The chance of failing is 0.001%. If you do it twice...it would be 0.001% of 0.001%. If you do it 3x..0.001% of 0.001% of 0.001%. You can see where this goes. Now, if you craft the CPU bascially do a bunch of multiplies instead of 1 just. Put in a bit of code to check all the multplies and pick the one with the most matches.

If the total energy to do the many vs the one is less...the answer is obvious. Remember, this is occassional errors...not constant errors.
Click to expand...

+1

This is worth rereading by a bunch of people in this thread. Allowing errors to occur in CPU cycles is not the same as the CPU outputting errors.
 
The article mentioned certain application like hearing aid which such occasional error is not going to impact its functionality.

So this study does have its benefits.
 
Dreaz said:
That's pretty much the end-game for technology tbh.
Click to expand...

I would prefer machines to remain machines, and biologics to remain biologics. Call me old fashioned that way, but I don't want my calculator psychoanalyzing me and then trying to lock me in a deep freeze "for my own good". I want it to do math and give me the correct result.
 
Looking forward to see such CPUs implemented in nuclear power plants or as brain behind strategic nuclear arsenal.

I bet that Daleks also running such chips :D

Seriously tho, what is the point of this? Error correction is paramount. That is the primary principle.
 
I think it's so funny that most people here seem to be against this. How many of you overclock and/or run your ram out of spec, don't run ECC ram....?
 
I fail to see how this is "revolutionary" as I didn't read the white paper.

But this technology has been around since the 50's before math coprocessors were common. Even the first versions of FORTRAN IV, V, 66, 77, 90 all had the compiler option to select "Fast Or Accurate floating point evaluation" Various C compilers also had this option. They Trade higher speed & less memory resources for less accurate results.
 
Trepidati0n said:
Errors and failures are not the same. The RAM in your PC is upsetting all the time. The data on your HDD is getitng corrupted all the time. However, there is correction algorithms to fix the eror and ths the net result is no failures.

Put it tihs way. Lets you need to muliply 2 number together. The chance of failing is 0.001%. If you do it twice...it would be 0.001% of 0.001%. If you do it 3x..0.001% of 0.001% of 0.001%. You can see where this goes. Now, if you craft the CPU bascially do a bunch of multiplies instead of 1 just. Put in a bit of code to check all the multplies and pick the one with the most matches.

If the total energy to do the many vs the one is less...the answer is obvious. Remember, this is occassional errors...not constant errors.
Click to expand...

You need to take numeral methods and statistics again. You just got a "F" from me. Calculating computer error bounds from floating point math operations does NOT work that way.
 
DigitalGriffin said:
You need to take numeral methods and statistics again. You just got a "F" from me. Calculating computer error bounds from floating point math operations does NOT work that way.
Click to expand...

With these type of errors, relative error (the percentage it is off) can be small. However absolute error can be enormous especially when doing divide with small digits! There's a whole field of computer science dedicated to numerical methods and quantifying bounds of error. This is especially important for things like Newtonian row reduction on large matrices where float point operations must be applied over and over and over and over again on the same numbers.

Error grows with each iteration. Hence the error bounds (called degree of certainty) grow.
 
At my last job we had to do a real time analysis of very high data rate yet highly noisy signal data. (S:N was low) We had to perform similarity analysis by measuring the angle between two vectors (one a set of pre defined Mexican hats curves) in real time. The data rate was so intense we were pegging 4 core CPU's at 90->100%. So I offloaded it all onto the video card's GPU (early early beta Open CL and PhysX stuff as well) The origal vectors were noisy, so we using single precision floating point available in the GPU cores was not an issue.
 
These chips aren't replacing x86 processors, even though that's what the article title would leave us to believe. The article is kind of a cock-tease in that these processors will only be used on embedded devices for application-specific processes, of which already have highly efficient processors.

Now they present cheaply produced processors that have 0.25% errors on them... that's great.
They can keep it.
 
You must log in or register to reply here.
Back
Top