CPU vs server CPU?

[tweakboy]

What's the difference between a standard CPU you would have in a PC, vs a 'server CPU?' Don't they do the same thing? Or is there additional functionality a server CPU requires?

Servers are usually racks. THey allow for massive RAM which desktop boards dont. It can offer up to 256GB on the mobo along with a 8 core 16 thread Xeon.

The CPU does the same thing your CPU does,, its just a bunch more cores and threads and its expensive,, like 1k 2k a chip and expensive mobo...

 

[tynopik]

Shoddy ECC memory will not magically be more reliable than shoddy non-ecc memory.

1. yes it is. That's where the 'CORRECTING' part of ECC comes in

2. Even if it weren't (and it is), at least it's immediately identifiable as to what the problem is

instead of mysterious lockups for no known reason, the computer can report 'your memory is bad and needs to be replaced'

That is a huge improvement right there.

Yes, cheap, low quality ram is an issue.

false

did you see the report? Even with server grade memory, 8% of DIMMs had errors

cheaper, non-ECC parts will do just as well.

now who's talking out of their ass?

Your PC crashes due to broken memory. Note "This "breakage" apparently is caused by sub-quality memory that does not meet general specifications and can crash software."

And how do you propose to resolve the problem of 'sub-quality' memory?

Without ECC there's no way for the typical user to pinpoint what the problem is. All they know is it keeps crashing and then they get frustrated and pitch it in the trash.

 

[Idontcare]

Long ago, 10+ yrs now, I had an AMD system that had ECC ram in it, it may have even been a dual-socket rig, can't remember now. But I remember I specifically purchased it so I could use ECC ram.

I wish ECC ram had caught on. With today's systems I don't 16GB of DDR3-2133 for $100. I'd be fine with 8-12GB of DDR3-1866 or 1600 of ECC ram for $100.

We've got so much ubiquitous computing resources now, bandwidth galore, why not make it ECC and have more stable operation?

 

[Chiropteran]

Per idontcare's link. A whopping 8% of working (that is non shoddy, broken) DIMMS suffered one more more correctable bit flips within a solid year of powered on time (not a calendar year, but a year of powered on time). That means that yes, there is a 30% chance that you might have a bit flip in a 4 DIMM system in a year (which will likely cause no problem).

I don't think so.

About a third of machines and over 8% of DIMMs in our fleet saw at least one correctable error per year. Our per-DIMM rates of correctable errors translate to an average of 25,000–75,000 FIT (failures in time per billion hours of operation) per Mbit and a median FIT range of 778 –25,000 per Mbit (median for DIMMs with errors), while previous studies report 200-5,000 FIT per Mbit. The number of correctable errors per DIMM is highly variable, with some DIMMs experiencing a huge number of errors, compared to others.

When the article says "over 8%" and "at least" one error per year it means the actual error rate could be much higher on average. You can't create a valid argument by assuming those numbers are maximums when the article was crystal clear about those being the minimum numbers of errors experienced.

Also, note the date of the blog post, 2009. A lot changes in 3 years, such as average amount of memory installed per computer and amount used per user. As shown above, the errors occur per Mbit. Double the memory usage, you double your error rate. If you had 2GB of RAM in 2009 and you now you use 10GB, all else being equal you have increased your error rate 5X.

 

[Ferzerp]

Long ago, 10+ yrs now, I had an AMD system that had ECC ram in it, it may have even been a dual-socket rig, can't remember now. But I remember I specifically purchased it so I could use ECC ram.

I wish ECC ram had caught on. With today's systems I don't 16GB of DDR3-2133 for $100. I'd be fine with 8-12GB of DDR3-1866 or 1600 of ECC ram for $100.

We've got so much ubiquitous computing resources now, bandwidth galore, why not make it ECC and have more stable operation?

For the same reason TN LCDs are so popular. People want what is cheap, not what is necessarily best. No one is stopping you from buying ECC capable hardware, you're just going to pay more for it (and probably not get much out of it, unlike higher quality panels).

 

[tynopik]

For the same reason TN LCDs are so popular. People want what is cheap, not what is necessarily best. No one is stopping you from buying ECC capable hardware, you're just going to pay more for it (and probably not get much out of it, unlike higher quality panels).

With TN, what you see is what you get. It's not a surprise. TN panels don't cause your computer to lock up, they don't cause you to lose work and they don't corrupt files.

RAM quality is invisible while sitting on the shelf. Everyone assumes it will 'JUST WORK'. Consumers shouldn't have to make a decision between "here's a system, but if you actually want a system that's reliable you have to buy this one." All systems should be reliable.

Systems can be differentiated on storage size, memory size, processor speed, display size, display quality, form factor and battery life. They should NOT be differentiated based on 'likelihood to crash'.

Even the cheapest systems should be reliable.

Saying 'No one is stopping you from buying ECC capable hardware' is missing the point. Intel should want ALL consumers to have ECC and until they put that capability in their CONSUMER chips and push all manufacturers to include ECC by default, there are going to be a lot of frustrated end users out there.

 

[Ferzerp]

You are severely inflating what ECC will allow for. You cited someone's comments on people using RAM that is *broken* and assume ECC is the answer to that.

It isn't. The answer to that is non-broken memory. ECC is not going to make broken memory magically unbroken.

 

[tynopik]

You are severely inflating what ECC will allow for. You cited someone's comments on people using RAM that is *broken* and assume ECC is the answer to that.

It isn't. The answer to that is non-broken memory. ECC is not going to make broken memory magically unbroken.

Question: if MS was just concerned about 'broken' memory, why didn't they just talk about high-quality memory? Why were they pushing ECC?

 

[tynopik]

I would say someone like Toyota understands this. Their base models may not have many features or much 'capability', but what they have WORKS.

Even their cheapest models are reliable.

If Intel wants people to be happy with PCs, they need to recognize how important reliability is to the user-experience and how this is a non-negotiable point, not something where you try to squeeze a little extra margin out of enterprise customers.

 

[Ben90]

Without ECC there's no way for the typical user to pinpoint what the problem is. All they know is it keeps crashing and then they get frustrated and pitch it in the trash.

OH MY GOD THAT WOULD BE HORRIBLE! Could you imagine if that user then bought a new PC because the old one was broke. That would really screw over hardware manufacturers.

"Damn consumers keep buying our stuff!"
"Quick, lets make them last forever so they never buy another computer again"

*edit*

On a more serious note what good would ECC do for this end user that has a bad memory stick. Either way the computer is going to be locking up, and either way they are going to trash it for a new computer.

 

[Flapjacks]

I'd rather go with ole reliable.

 

[bononos]

Do normal ddr3 dimms use parity bits?

 

[tynopik]

Do normal ddr3 dimms use parity bits?

no

 

[tynopik]

OH MY GOD THAT WOULD BE HORRIBLE! Could you imagine if that user then bought a new PC because the old one was broke. That would really screw over hardware manufacturers.

"Damn consumers keep buying our stuff!"
"Quick, lets make them last forever so they never buy another computer again"

That theory works great when PCs are the only option.

When Intel is fighting off a growing wave of 'alternative devices' like tablets, it's a dumb, dumb idea

On a more serious note what good would ECC do for this end user that has a bad memory stick. Either way the computer is going to be locking up, and either way they are going to trash it for a new computer.

When ECC memory suffers an unrecoverable error, Windows logs it. Thus the end user can find out what the problem is. With non-ECC error, Windows has no way of knowing that it was the memory at fault

 

[cbn]

I like how you just post on the straight and narrow :thumbsup:

For the rest of the lurkers, the difference oftentimes comes down to the validation efforts that go into ensuring the server-features function and perform as intended and needed.

That takes time, time = money, the product costs more to purchase because it costs more the "produce" it.

And there's also the staggering difference in software costs for the server versus desktop. For us desktop users, Photoshop is probably the most expensive single app any one of us would license for ourselves to use at home. (Mathematica or Gaussian 09W may compete for that title)

So desktop users are more likely see the hardware cost of their system as being a significant portion of the TCO (total cost of ownership) for their computers.

Not so with the server crowd. Their software can run anywhere from $10k to $1m per "server", and the infrastructure costs (building + people) can add an equal amount of TCO to that equation.

For that market, whether the CPU itself costs $200 or $2000, the difference in hardware cost is nearly inconsequential in the big picture. So why not charge $2k then?

I was quite surprised to find out how some server software costs compared to the server cpu costs.

 

[Ketchup]

I have spent a good part of this week putting together servers for one of our customers, and when it comes down to it, real servers are designed around one word: reliability.

Why do they have four hard drives? reliability
Why do they have two power supplies? reliability
Why do they have ECC memory? reliability

ECC memory still holds an advantage, however small it may be, over non-ECC memory in reliability. Customers pay for the amount of RAM and CPUs based on the speed they need, but reliability needs to be there whatever the price.

 

[bononos]

no

tsk I thought they did.
I remembered parity was common or at least was an option for simm type memory since I was pretty sure I had memory with parity bits on my socket 7 system.

Did non-parity become the standard in the shift to SDRAM and dimm slots or was non-parity already the norm for consumer level PCs when we were still using simm slots?

 

[pm]

A long time ago, I bought an 875 chipset motherboard with ECC on it. It actually took a couple of tries because I remember a company (ABit? AOpen?) said their motherboard was "ECC enabled" but that meant that you could socket ECC memory into it and it would still boot... not that it actually fixed single-bit errirs But then I got another motherboard and this time ECC is actually appear to be enabled.

The board went into an always-on server running Linux in my basement and then I found out that I didn't have a driver for ECC so then I helped with the linux-ecc driver for the chipset to enable it (turned in the patch to the linux-ecc kernel group too). I turned on ecc data logging and I didn't see any failures for weeks, so then I created a failure myself (I used a bit of wire-wrap wire, stuck it in the socket, socketed in an ECC DIMM creating an antenna and it took me a couple of tries to get the length long enough but soon enough I had a failure. So I knew it worked.

Then I logged ECC errors on my server for 9-12 months. As I recall - and it's been a long time - I logged 7 errors on an always-on server with ~2GB of RAM. I'm at an altitude of 1600m. What I thought was interesting was that the errors were very bursty. I'd get 3 and then months went by and I got 4. It was something like that.

I bought a 975X motherboard after that - which I still have - and tried to replicate my experiment on my desktop system running Windows. I couldn't find out where Windows was logging it under Windows 95 and concluded that they didn't. So I wrote a Win95 driver for Windows for ECC - you need to read a couple of the chipset registers... it wasn't rocket science. But I never caught any errors with this method except for my dangling wire test - but it was on a system that I powered on/off daily, and the industry had switched from lead-solder to the new stuff which cut down on radiation as well. But I think it was 6 months and I never saw anything. But then I switched to Vista for some reason that made sense back then, but I'm not clear on nowadays, and the driver didn't work because it wasn't signed and the driver coding methodology for Vista was much more confusing... so that was the end of my ECC experimentation. When I bought my Nehalem motherboard and then my present-day Sandybridge motherboard, I didn't bother with ECC.

I will say that I'm in the middle of the fence on ECC. Just as I try to buy motherboards from solid vendors I am willing to buy motherboards with ECC and pay a bit extra for it, but not a lot extra. But even when you are doing this, my experience with Abit/Aopen/someone else (I can look it up if you are curious... I wrote them a complaint letter and mailed it), even when you buy ECC memory and an "ECC-enabled" motherboard, you may not be actually getting an ECC-enabled system. And who has time to stick random wires in their DIMM sockets to test... So my take on it was that most users shouldn't probably bother.

I'm headed out on vacation to a spot with no internet tomorrow so I doubt I'll be responding to this thread unless it's miraculously still going in two weeks time... so if you have any questions or want to tell me that I'm wrong, then I'd recommend sending a private message to me. I still should have the code and everything somewhere on my hard disk if anyone is curious. I'm a data packrat.


On the OP's original question, I thought IDontCare did his usual excellent job answering it.

Edit: I found the code for the ECC kernel for Linux and it was dated 2002. Wow. I had no idea it was that long ago. I think my chipsets references above are all wrong because the code looks like it was for the Intel i440BX chipset. Man, time goes by too fast as you get older.

 

[Ayah]

You're going to really love ECC memory when you start packing your servers with 64GB LRDIMMs, and coming soon 128GB LRDIMMs.

 

[nenforcer]

You can also usually run quite a few Intel Xeon processors in the high end desktop chipset platforms for Intel.

In some instances the chips are exactly the same, as my Xeon 3110 is identical to a Core 2 Duo E8400 other than different CPUID fields.

For instance, the ASUS P6T WS Supercomputer uses the same X58 chipset Socket 1366 platform and will accept Intel Xeon W3550 processors just as easily as Core i7 920's.

It's basically the same motherboard as the ASUS Rampage ][ Extreme or P6T Deluxe / V2 with the exception of a slightly different BIOS which may enable support for ECC memory.

I actually run a Xeon 3110 in my Nvidia Nforce 790i chipset machine.

 

[ncalipari]

For that market, whether the CPU itself costs $200 or $2000, the difference in hardware cost is nearly inconsequential in the big picture. So why not charge $2k then?

lol you really have a big mouth.


Is there a way to ignore someone on this forum?


I mean hiding someone's post.

The one with the big mouth here is you. This is your last warning before an infraction.
Markfw900
Anandtech Moderator

 

[jones377]

This is gonna be awesome

 

[Charlie98]

This is gonna be awesome

...all this from a simple question about differences between CPU's?

 

[tynopik]

It is already unacceptably high

It is (probably) the leading cause of blue screens.

This is pure conjecture on your part. You have nothing backing it up other than it is what you feel to be the case, do you?

more evidence for discussion:
http://research.microsoft.com/apps/pubs/default.aspx?id=144888

it's amazing how unreliable computers actually are . . .

another fun stat:

The table shows that CPUs from Vendor A are nearly 20x as likely to crash a machine during the 8 month observation period when they are overclocked, and CPUs from Vendor B are over 4x as likely.

 

[Ferzerp]

You still can't separate bit flip (which all ram is subject to, and what ECC addresses) and faulty hardware (which ECC won't do %$#@ for).

There is little point in continuing to discuss this until you can separate the two. ECC ram fails exactly like non ECC ram.