RAM Put to the Test: ECC vs Non ECC Memory Performance

Jos

Posts: 3,073   +97
Read the full article at:
[newwindow=https://www.techspot.com/article/845-ddr3-ram-vs-ecc-memory/]https://www.techspot.com/article/845-ddr3-ram-vs-ecc-memory/[/newwindow]

Please leave your feedback here.
 
How weird. I didn't even know there was a myth surrounding ECC vs non-ECC performance at a common bandwidth and latency. I was under the impression that the only downsides of ECC memory is that it requires a Xeon for support (at least for Intel systems), and ECC memory is binned fairly low on bandwidth and high on latency - since speed and tight timings tend to exacerbate memory errors.

Seems I've been missing out on yet another forum war based on hearsay and assumption. I need to stay in more.
 
How weird. I didn't even know there was a myth surrounding ECC vs non-ECC performance at a common bandwidth and latency. I was under the impression that the only downsides of ECC memory is that it requires a Xeon for support (at least for Intel systems), and ECC memory is binned fairly low on bandwidth and high on latency - since speed and tight timings tend to exacerbate memory errors.

Seems I've been missing out on yet another forum war based on hearsay and assumption. I need to stay in more.

Have to agree I thought the same thing when I saw the title. I tested ECC memory back in the SDRAM days and found no difference when compared to standard memory.

Like you say its the processor that limits the support of ECC memory, otherwise we would all use it ;)
 
Great test, thanks! I love these practical comparisons on unusual parts, clears up my shopping list for the future. Also, ECC seems to be a very tasty option - secure and without (practically) any slowdown.
 
Great test, thanks! I love these practical comparisons on unusual parts, clears up my shopping list for the future. Also, ECC seems to be a very tasty option - secure and without (practically) any slowdown.
If you have the money to throw at it, sure. I'll stick with regular RAM for anything other than mission-critical hardware.
 
Great test, thanks! I love these practical comparisons on unusual parts, clears up my shopping list for the future. Also, ECC seems to be a very tasty option - secure and without (practically) any slowdown.
If you have the money to throw at it, sure. I'll stick with regular RAM for anything other than mission-critical hardware.

It's not really even that. For your desktop you would have to use server hardware. ECC memory isn't something that worry's desktop users, its not really on the cards.

Guys with Xeon processors in their desktop PC are playing mission critical games of Minecraft though.
 
Very interesting read.

How weird. I didn't even know there was a myth surrounding ECC vs non-ECC performance at a common bandwidth and latency. I was under the impression that the only downsides of ECC memory is that it requires a Xeon for support (at least for Intel systems), and ECC memory is binned fairly low on bandwidth and high on latency - since speed and tight timings tend to exacerbate memory errors.

Seems I've been missing out on yet another forum war based on hearsay and assumption. I need to stay in more.
I thought the same thing. I find it interesting also how the "known"/"thought" 2% performance difference is even mentioned by crucial... It is so mere it doesn't really matter and shouldn't be mentioned by them in the first place.
 
The answer comes pretty much can be summed up as "Depends on what your doing". ECC memory is pretty much made for server environments and designed to be pushed 24/7, not for the standard desktop user. There is a reason it costs so much and the tighter timings along with its 24/7 rated use is just the cherry on the sundae.

Great article, I knew people who buy server components just to say they had the most powerful machine and bragged about their dual Xeon setups at LAN parties with their ECC ram and stuff. Always found it funny but never knew it was a phenomenon.
 
Without knowing one way or another, I always figured the ECC had extra components to pick up the slack. The thought of one out performing the other never crossed my mind. And that is for one reason, ECC never was advertised to be faster or slower. ECC was advertised for what it is, Error Checking & Correction.
 
I tried to look at the reason for the 12% loss and being that the Kingston site does not have accurate pics of the dimms nor do they cite how the dimms are "ranked" in placement on the PCB I can only assume that possibly the way the dimms are ranked is different and that the channel interleaving might be the cause of the perf issues. Also not known is if some of those tests are NUMA aware, or not, could also sway the results if one of the apps utilized NUMA nodes and the others were NUMA agnostic.
 
Without knowing one way or another, I always figured the ECC had extra components to pick up the slack.
Standard DRAM is 64-bit, ECC is 72-bit, so if a regular DDR2/3 module has 8 IC's per bank, and ECC module has 9. The extra 8 bits per bank ensure that ECC doesn't suffer from SBE (Single Bit Errors) which are the most common errors in RAM reading/writing.
As Steve and I have noted, performance -wise RAM is RAM if the bandwidth and latency are equal. Any slowdown in ECC would likely be because of the CPU's memory controller or subsystem - workstation and servers tend to be a little more complex than desktop.
The answer comes pretty much can be summed up as "Depends on what your doing". ECC memory is pretty much made for server environments and designed to be pushed 24/7, not for the standard desktop user. There is a reason it costs so much and the tighter timings along with its 24/7 rated use is just the cherry on the sundae.
Untrue. ECC modules are generally sold with relaxed timings. You won't find ECC modules with timings as tight as gaming orientated RAM. That is to say, however tight the timings are on ECC RAM, you can find tighter (and faster) in desktop non-ECC.
Three other common reasons for ECC modules being more expensive than desktop RAM
1. The extra IC's required ( 9 instead of 8 per bank)
2. Binning
3. Many ECC modules have embedded thermal sensors.
 
Nice work Jose. as @cliffordcooley noted:
ECC was advertised for what it is, Error Checking & Correction
If you like your files to actually contain the data you create, then ECC is a component to ensure file integrity (it's quite obvious but seldomly considered :sigh: )
 
Just read the article and think this is a very good topic to check out.
As far as I can see, there is one thing that should have been noticed...
While both UDIMM and ECC UDIMM are dual Rank modules, the used Kingston KVR16R11S4/8I RDIMM is single rank, which gives the RDIMM a little disadvantage.
Also it would have been nice to include LRDIMMs to have all supported memory covered.
 
This article does not test this with any relevance to tasks the workstation is doing. I do CFD and I have learned the hard way that when you run a calculation that takes days and maybe weeks ECC is the only thing that will work at all. the stability is a real factor in heavy duty work. most XEONs can only use ECC any way and there is a reason for that.
 
Back