EVGA SR2-It is here and OH MY GOD is it a thing of beauty!

[nn_step]

You have to forgive nn_step, brilliant software guy but when it comes to hardware he's all AMD and they never used FBDimms..

In correct. I only buy the best parts for my research [Right now Intel completely ignored several significant advancement in virtualization that AMD supports] Also You can't buy an Intel setup with hardware support for multiple address spaces and virtualized I/O Memory. [Something REQUIRED for super reliable computing]

Ah... lol

Speaking of ram... any idea when they're gonna sell affordable non-ECC 4GB DDR3 sticks?

Working with 12GB of ram is worst than being claustrophobic in a room the size of a coffin...

Question, why would anyone even think about buying more than 4GB that Isn't ECC? Now unregistered ECC is one thing but losing the very protection and sanity check that any serious system requires just seems illogical to me.

I actually don't care about Pi much... It's more about the algorithms that are involved.
I do a couple other things that require a ton of ram... (Though they're only on the order of 8 - 32GB of ram.)

Actually the algorithms are only part of the equation. Given enough Hard drive bandwidth and a smart enough compiler with hints; not much system ram is required.
[TOP SECRET] Lets just say I have something new in the mix [/TOP SECRET]

[poke349]

Question, why would anyone even think about buying more than 4GB that Isn't ECC? Now unregistered ECC is one thing but losing the very protection and sanity check that any serious system requires just seems illogical to me.

1. When the motherboard doesn't support ECC ram.
2. Ram is surprisingly reliable... Like I said, not a single ECC event in one year of torturing 64GB of ram.
3. I don't do any extremely long-running stuff on my small machines. So if a memory error occurs, software error-detection will catch it and I can simply run it again.

Actually the algorithms are only part of the equation. Given enough Hard drive bandwidth and a smart enough compiler with hints; not much system ram is required.
[TOP SECRET] Lets just say I have something new in the mix [/TOP SECRET]

Highlighted the keyword go ya.

[nn_step]

1. When the motherboard doesn't support ECC ram.
2. Ram is surprisingly reliable... Like I said, not a single ECC event in one year of torturing 64GB of ram.
3. I don't do any extremely long-running stuff on my small machines. So if a memory error occurs, software error-detection will catch it and I can simply run it again.




Highlighted the keyword go ya.

1. It costs $1.50 [or less] in parts to ensure support for ECC be it AMD or Intel or Via. What self respecting Motherboard manufacturing company would skimp on $1.50 in parts that ensures it a place in the professional market?
Not to mention ECC Memory is only 9/8ths the cost of Non-ECC memory [or 12.5% more expensive] which is worth the added cost.

2. Actually Errors are alot more common than you think. [Most software only record errors that can't be corrected automatically] Now recent tests give error rates ranging from 10^-10 to 10^-13 error/bit·h, roughly one bit error, per hour, per gigabyte of memory.
Thus even with the BEST memory on Earth, you are guaranteed at minimum of 1 bit corruption Every 10 hours when talking more than 100GB RAM

3. Actually only about half of all errors will have any Real effect on a system [Even when talking about No Error detection or correction at all] And with the SMALL added price, ECC Ensures that the error rate moves to the lovely
10^-18 or to one error, per century, per 10 gigabytes of memory. [That can't be instantly corrected]

[D_A]

1. It costs $1.50 [or less] in parts to ensure support for ECC be it AMD or Intel or Via. What self respecting Motherboard manufacturing company would skimp on $1.50 in parts that ensures it a place in the professional market?

$1.50 a board?? Consider how many hundreds of thousands of boards you're looking at AND that they're aimed at home users who can deal with the odd wrong character in their email and that $1.50 suddenly costs them a LOT of money. It's the same reason car makers use non-galvanised screws inside cars when it would "only" cost them 0.1 of a cent per screw to use galvanised ones.

[poke349]

1. It costs $1.50 [or less] in parts to ensure support for ECC be it AMD or Intel or Via. What self respecting Motherboard manufacturing company would skimp on $1.50 in parts that ensures it a place in the professional market?
Not to mention ECC Memory is only 9/8ths the cost of Non-ECC memory [or 12.5% more expensive] which is worth the added cost.

2. Actually Errors are alot more common than you think. [Most software only record errors that can't be corrected automatically] Now recent tests give error rates ranging from 10^-10 to 10^-13 error/bit·h, roughly one bit error, per hour, per gigabyte of memory.
Thus even with the BEST memory on Earth, you are guaranteed at minimum of 1 bit corruption Every 10 hours when talking more than 100GB RAM

3. Actually only about half of all errors will have any Real effect on a system [Even when talking about No Error detection or correction at all] And with the SMALL added price, ECC Ensures that the error rate moves to the lovely
10^-18 or to one error, per century, per 10 gigabytes of memory. [That can't be instantly corrected]

My Asus Rampage II GENE doesn't support ECC memory. If it did, I'd be running a W3520 + 6 x 4GB right now since ECC DDR3 4GB sticks is a lot cheaper than non-ECC DDR3 4GB sticks.

If memory errors are THAT common, then can you explain why my 12GB of ram didn't encounter a single error in 7 months of continuous torture? (overclocked as well)
Based on your statistics I should have had plenty of errors by now.

If an error were to occur one bit for every 10^13 bits transfered. Then with 4GB/sec usage (which is a very conservative estimate for Core i7), I should encounter an error once every 8*10^13 / (4*10^9) seconds which is about one error every 5 minutes.

Certainly that can't be right, or all the WCG people would be in trouble...

With 10^-13 error rate @ 4GB/sec, the probability of encountering an error in each hour is:

1 - (1 - 10^-13)^(3600 seconds * 8 bits * 4*10^9) = 0.9999900704956 = 99.999007%

I've been running it for 7 months now. But I'll give you the benefit of the doubt and say it's only sustained 3 months of memory intensive usage.
Then the probability of not encountering an error is:

(1 - 0.9999900704956)^(24 hours * 30 days * 3 months) = 2.30970653^-10807

which is about one in 4.3295543 * 10^10806.

I find that hard to believe.
If I made an error in the math somewhere, feel free to point it out.
Perhaps I'm misinterpreting the 10^-13?


On the software side, I use my own error-detection. (Mainly for catching CPU errors, but they work just as well for memory errors.)
Bit flips are obvious to catch because they propagate and produce non-sense data. (Although this is somewhat application specific.)

[nn_step]

My Asus Rampage II GENE doesn't support ECC memory. If it did, I'd be running a W3520 + 6 x 4GB right now since ECC DDR3 4GB sticks is a lot cheaper than non-ECC DDR3 4GB sticks.

If memory errors are THAT common, then can you explain why my 12GB of ram didn't encounter a single error in 7 months of continuous torture? (overclocked as well)
Based on your statistics I should have had plenty of errors by now.

If an error were to occur one bit for every 10^13 bits transfered. Then with 4GB/sec usage (which is a very conservative estimate for Core i7), I should encounter an error once every 8*10^13 / (4*10^9) seconds which is about one error every 5 minutes.

Certainly that can't be right, or all the WCG people would be in trouble...

With 10^-13 error rate @ 4GB/sec, the probability of encountering an error in each hour is:

1 - (1 - 10^-13)^(3600 seconds * 8 bits * 4*10^9) = 0.9999900704956 = 99.999007%

I've been running it for 7 months now. But I'll give you the benefit of the doubt and say it's only sustained 3 months of memory intensive usage.
Then the probability of not encountering an error is:

(1 - 0.9999900704956)^(24 hours * 30 days * 3 months) = 2.30970653^-10807

which is about one in 4.3295543 * 10^10806.

I find that hard to believe.
If I made an error in the math somewhere, feel free to point it out.
Perhaps I'm misinterpreting the 10^-13?


On the software side, I use my own error-detection. (Mainly for catching CPU errors, but they work just as well for memory errors.)
Bit flips are obvious to catch because they propagate and produce non-sense data. (Although this is somewhat application specific.)

http://www.cs.toronto.edu/~bianca/pa...gmetrics09.pdf

your error is in that it is measured in Bit hours [thus the bandwidth feeding the CPU isn't the determining factor]

Also I repeat Errors automatically corrected By ECC will not be reported. [Because reporting a non-issue is worse than worthless]

[tool_462]

Dave, keep your dang thread clean and remotely on topic!

[artemm]

Does this board even require ECC?

[Movieman]

Dave, keep your dang thread clean and remotely on topic!

This is the WCG section where the friendships are more important than any silly rule like keeping on topic.
Besides, if I tried to do that in this thread I'd be the biggest hypocrite of all time!

Does this board even require ECC?

Either way, gaming memory or ECC..
ECC support is in the bios if you want to use it.

[STEvil]

I think the big question should be: why isnt all memory ECC now?

[Movieman]

I think the big question should be: why isnt all memory ECC now?

speed and people wanting that speed.
The benchers and gamers don't want to be limited to 1333 memory at CAS9..
Not an issue for what we do crunching but they'd rebel at the "slow stuff"

[poke349]

Here's my calculations based on bit-hour:

Lemme know if I messed up again:

Assume I average 8GB of ram usage:

Probability of not encountering an error in one hour:
(1 - 10^-13)^(8 bits * 8 * 2^30 bytes) =
0.9931516101651806070663949027917378

Probability of not encountering an error in one day:
(1 - 10^-13)^(8 bits * 8 * 2^30 bytes * 24 hours) =
0.847955819523137202583397727187477

Probability of not encountering an error in one month:
(1 - 10^-13)^(8 bits * 8 * 2^30 bytes * 24 hours * 30 days) =
0.007098993078000191887076602721466

Probability of not encountering an error in 3 months:
(1 - 10^-13)^(8 bits * 8 * 2^30 bytes * 24 hours * 30 days * 3 months) =
3.5775874478084587076749241202893770 * 10^-7

or about one in 2.795 * 10^6.

k, it makes more sense now. One in 2.795 * 10^6 is still hard to believe. But that's with 10^-13 error rate.

If 10^-13 is what my non-ECC ram is rated for, then it's perfectly reasonable to believe that they were intentionally overbuilt to have a lower error rate than that.

Using 10^-14 as an error rate, the probability of not encountering an error in 3 months of sustained 8GB usage is:
(1 - 10^-14)^(8 bits * 8 * 2^30 bytes * 24 hours * 30 days * 3 months) =
0.226651723475289943371528045846591
or about 1 in 4.4...

Now we're talking.


I guess I'll try to get back on topic?

[TJ TRICHEESE]

I think evga is overstocked on six pin pcie connectors, first 4 way sli 285's, now this.

putting unnessary power connections every where will make it look crap when all cables are in

what form factor does evga call this, maby this: extra extra extra super retarded never fit in any case large-ATX or EEESRNFIACL-ATX

[Movieman]

I think evga is overstocked on six pin pcie connectors, first 4 way sli 285's, now this.

putting unnessary power connections every where will make it look crap when all cables are in

what form factor does evga call this, maby this: extra extra extra super retarded never fit in any case large-ATX or EEESRNFIACL-ATX

Things always appear different based on the perspective of the person looking.
What I see is that they put the extra 6 pin power connectors as close as possible to the areas that would benefit from that added power and only needed if one was using multiple high end video cards.
You may well have a point on the looks when all the cables are in place but I see function over form and in the end, it is function that matters.
The board is 13.58"x15" and isn't to any form factor.
Some are calling it WATX.
Reality is a XL-EATX MB tray will work with some minor mods.
If you look at the board as I have and closely you find that they needed all of this space. The board is packed and that answers why the size.
People miss that dualies require more than the space to slap in another cpu socket and 6 more dimm slots. the cpu's not only communicate with the board as in a single cpu setup but also with each other.
last, it's not for everyone. Expensive and very expensive to build a full out system with all the goodies but it is for those people who want what it offers:
A monster benching system, a overclockable dual 1366 system(me)
A huge encoding or rendering system and finally yes, THE system for the guy that wants the biggest E-PEEN in the world..

[nn_step]

speed and people wanting that speed.
The benchers and gamers don't want to be limited to 1333 memory at CAS9..
Not an issue for what we do crunching but they'd rebel at the "slow stuff"

But the slow down can be significantly minimized should a vendor such as G.SKILL or OCZ just start binning and selling ECC Sticks

Here's my calculations based on bit-hour:

Lemme know if I messed up again:

Assume I average 8GB of ram usage:

Probability of not encountering an error in one hour:
(1 - 10^-13)^(8 bits * 8 * 2^30 bytes) =
0.9931516101651806070663949027917378

Probability of not encountering an error in one day:
(1 - 10^-13)^(8 bits * 8 * 2^30 bytes * 24 hours) =
0.847955819523137202583397727187477

Probability of not encountering an error in one month:
(1 - 10^-13)^(8 bits * 8 * 2^30 bytes * 24 hours * 30 days) =
0.007098993078000191887076602721466

Probability of not encountering an error in 3 months:
(1 - 10^-13)^(8 bits * 8 * 2^30 bytes * 24 hours * 30 days * 3 months) =
3.5775874478084587076749241202893770 * 10^-7

or about one in 2.795 * 10^6.

k, it makes more sense now. One in 2.795 * 10^6 is still hard to believe. But that's with 10^-13 error rate.

If 10^-13 is what my non-ECC ram is rated for, then it's perfectly reasonable to believe that they were intentionally overbuilt to have a lower error rate than that.

Using 10^-14 as an error rate, the probability of not encountering an error in 3 months of sustained 8GB usage is:
(1 - 10^-14)^(8 bits * 8 * 2^30 bytes * 24 hours * 30 days * 3 months) =
0.226651723475289943371528045846591
or about 1 in 4.4...

Now we're talking.


I guess I'll try to get back on topic?

Given those assumptions [Which may or might be off by a factor of ten]. Your calculations are correct.
Also that is assuming the firmware tracking Error rates, reports single bit errors and tracks them [Which I find highly unlikely]

[TektoNQuaD]

Looks Like A very nice mobo

[Bobsama]

Dave; getting this rig up Wednesday or Thursday?

Oh, and I just cooked my rig for the past ~week. When I got back into the room, it was 82F or so in here.

[Movieman]

Dave; getting this rig up Wednesday or Thursday?

Oh, and I just cooked my rig for the past ~week. When I got back into the room, it was 82F or so in here.

All depends on when the parts arrive but weds or thursday is a good guess except for one part that won't be here till friday.

[kpy_here]

Here's a mini-preview of the board, from our Romanian Fellas
http://lab501.ro/stiri/cebit-2010-ev...pentru-desktop

[Molotox]

MM, you are gonna have to change avatar! I wet my pants everytime I see this board! I wish I could get a student loan for a rig based on this... lol

[Movieman]

MM, you are gonna have to change avatar! I wet my pants everytime I see this board! I wish I could get a student loan for a rig based on this... lol

You? How the hell do you think I feel?
At least your young with a strong bladder pal.
I'm 58 and in a week wiped out the entire New England supply of Depends!

[Angmaar]

That board is hot! Any idea how many PPD it will push out?

[Movieman]

That board is hot! Any idea how many PPD it will push out?

My best guess is that if I can hit 4350mhz I can make app 100,000PPD..
That's just a educated guess.

[=[PULSAR]=]

100,000PPD on what? WCG I would think roughly 15k w/OC.

[Movieman]

100,000PPD on what? WCG I would think roughly 15k w/OC.

I mean WCG points..7 WCG=1 Boinc..