DDR3 memory stability on X38 with C2Q

[pr0t0typ3]

Hi there,

This is my first post here on the Anandtech forum and among the first on any forum for that matter. Even though I have been reading about electronics, computers and overclocking religiously for about 7 years now, I just haven't had the means to get a decent PC for overclocking. This has changed about two months ago. I'm still limited financially so the PC I have is almost a relic but for the first time, I've had the chance to not just read but actually get my hands dirty.

Needless to say, it's been a long two months; I have learned a lot and had even more fun. I have read countless articles and forum posts and realized something that is probably nothing new to the veteran folks out there: the deeper you go, the less info is available about the questions you have, and the more confusion and misinformation you need to sort through. I have more and more questions that I am unable to find a definitive answer to, or even worse, ones I have found multiple answers to, all presented as definitive...

For most of the progress I've made, I have Anandtech to thank for; the scientific value of the site's and its forum's content is just unbelievable, and this is why I have made this account. I'm hoping to have thought provoking conversations with veterans and maybe clarify a few things for others in the process as well. My main goal is not to squeeze out the very last bit of performance out of my rig but to gain an in-depth understanding of how it works.

Sorry for the long intro but I wanted to share where I'm coming from. That's done so let's dive in!

My current configuration:

- Motherboard: Asus p5e3 X38
- CPU: Intel Xeon X3353 OC'ed to 3.6 GHz (2.66 GHz stock with a 8x multi, 45nm quad with 12 MB of l2 cache)
- Memory: 4x2GB Kingmax DDR3 1333MHz (9-9-9-24 CR2)
- GPUs: ASUS GTX 960 strix 2 GB and Gainward GT 610 1 GB (only for its VGA port...)
- Misc: crappy 550W generic PSU, Samsung 840EVO 120 GB, 1 TB+300 GB old SATA-II hard drives

I have purchased the motherboard, RAM and PSU from a friend. He used it for about 5 years with a Q6600 at stock speeds (no OC whatsoever, at least prior to my purchase, that is). He experienced no memory or stability issues. However, as soon as I changed the CPU to the X3353, the system posted but failed to boot. Here's a list of everything that had been changed from the original Q6600-based system configuration:

- storage configuration
- number and configuration of fans (the system was moved to a different case)
- CPU
- CPU socket (needed for the LGA-771 Xeon CPU)
- BIOS updated and modified (micro code extension needed for the Xeon)

For reasons I have mentioned above in the intro, I have spent hours tweaking the system every day throughout the past two months and tried countless combinations of BIOS settings. However, no matter what I did, I was unable to achieve and maintain stability with the memory set to its rated stock frequency. All I got was a POST or boot failure. A few times (like twice, maybe three times) the system successfully booted into Windows but froze or rebooted shortly after. I am unable to achieve stability at any effective memory frequency above 1000-1100MHz, but even 1100MHz is flakey.

What I have tried, to no avail:
- resetting BIOS and leaving everything as is, mostly 'auto'
- setting everything to factory defaults manually, including CPU frequency (FSB and multi), CPU voltage, memory frequency and timings, etc.
- setting everything to default manually but loosening memory timings
- playing with all available timing- and latency-related settings (AI transaction boost, AI clock twister, etc.)
- increasing memory voltage, up to 1.7v (1.5v stock)
- increasing NB voltage, up to 1.65v (1.25 stock)
- playing with the rest of the available voltage settings (PLL, FSB termination, vCore, SB, etc.)
- a variety of different DRAM skew values on channels A and B
- reseating DIMMS, cleaning RAM slots with pressurized air
- countless different combinations of the settings above

At lower frequencies however, it was easy to achieve stability. I could even tighten timings beyond recommended values (according to MemSet, the DIMMs are rated at 666 MHz (9-9-9-24 CR2 33-74-4-10-5-5), 592 MHz (8-8-8-22 CR2 30-66-4-9-5-5) and 444 MHz (6-6-6-16 CR2 22-49-3-7-4-4)). Performance is not bad either, here's an example:

FSB: 440 MHz
Memory bus: 528 MHz
Effective memory frequency: 1056 MHz
DRAM:FSB ratio: 12:10
Timings: 6-5-5-14 CR1

Read: 11373 MB/s
Write: 9346 MB/s
Copy: 9754 MB/s
Latency: 67.4 ns

Given how low quality my RAM is, these numbers are not bad (or are they?).

It's also worth mentioning that I had issues even at low memory frequencies. Whenever I selected 'stress system memory' in the AIDA64 stability test, within a few seconds the test stopped and a message appeared saying 'hardware failure detected'. Without stressing the RAM, the stability test ran for hours without a problem and other stress tests (Prime95, Intel Burn Test, OCCT) ran without errors as well. I couldn't find a solution for this for a long time but it was resolved after the first 3-4 weeks of tweaking the system and I haven't experienced it ever since. Unfortunately, back then I wasn't logging the settings I changed in the BIOS and I can't remember what solved the problem…

I know my memory is of pretty poor quality, this is probably the cheapest DDR3 RAM available. But it had worked with no issues running at 1333 MHz for 5 years, yet with the new CPU the system suddenly doesn't boot. It just can't be a coincidence. Also, I have run memtest86 countless times and got no errors at all.

In a couple of forum threads here and there it is mentioned that running four DDR3 DIMMs in mobos with older chipsets heavily stresses the MCH and could cause FSB OC limitations. They say similar things about running C2Q's in general. Is this true? Even though my issue persists with or without overclocking, could it be related to this? Or could my memory be too crappy? If so, how could it handle running at stock for all those years? Q6600's have a different stock FSB, so the memory divider was not the same. Could it be because of that? Or is it processor-related? After all, the X3353 is an LGA 771 chip so it was designed to be paired with ECC memory but there are so many people running LGA-771 Xeons on 775 mobos, yet I haven't read anyone talking about such issues.

Am I missing something here? Any input and would be highly appreciated. I would like to understand what's going on.

 

[tcsenter]

- Memory: 4x2GB Kingmax DDR3 1333MHz (9-9-9-24 CR2)

Module timings are usually marketed based on population assumption of only one or two modules. Load up all four DIMMs and things change, particularly with cheaper modules that aren't vetted in performance configurations. So you're less likely to get the "advertised" frequency at those timings when populating four modules.

Also, running the latest BIOS?

 

[VirtualLarry]

Honestly, you're chasing ghosts here.

On Core2-era platforms, running RAM with a faster memfreq than FSB, is pointless.

If the RAM is stable, with your CPU overclock, at a lower frequency, leave it there.

<-- veteran Core2Quad OCer, though I've not done a Xeon mod.

 

[myocardia]

Module timings are usually marketed based on population assumption of only one or two modules. Load up all four DIMMs and things change, particularly with cheaper modules that aren't vetted in performance configurations. So you're less likely to get the "advertised" frequency at those timings when populating four modules.

Indeed.

If the RAM is stable, with your CPU overclock, at a lower frequency, leave it there.

No doubt. Pr0t0, you are chasing one or two tenths of a frame/sec in the games that you're playing. If that couple of tenths of an FPS are that important to you, overclock your GPU by an extra 15 or 20 Mhz, and you'll get much more than a few extra tenths.

 

[pr0t0typ3]

Thank you for your answers!

Module timings are usually marketed based on population assumption of only one or two modules.

Interesting! I was not aware of this. It makes me wonder though: how does populating more modules affect the performance of the individual modules?

Also, running the latest BIOS?

Yes, version 1503. It was released in December, 2009... Kind of old, but it's the latest.

On Core2-era platforms, running RAM with a fast memfreq than FSB, is pointless.

running the latest BIOS?

Because of the benefits of 1:1 FSB : DRAM ratio, right? Summary from this forum thread:

A ratio of 1:1 provides the best level of stability, since the memory controller, which is an integral part of the northbridge chipset for Intel processors, does not need to translate data flow across the FSB between the memory modules and the processor(s). Also, since memory and processor FSB clocks are synchronous at 1:1, (400:400 or DDR 800), there is no additional latency introduced.

The thing is though that for some reason, at 1:1 (which translates to 900 MHz effective memory frequency in my case) I get worse throughput and higher latency than with my best stable setting with different ratios. Dunno why...

If the RAM is stable, with your CPU overclock, at a lower frequency, leave it there.

If that couple of tenths of an FPS are that important to you, overclock your GPU by an extra 15 or 20 Mhz, and you'll get much more than a few extra tenths.

Again, fixing this is not my main priority; rather, I'd like to understand the reasons behind the issue as well as possible. To be honest, I really don't mind about having my memory run at a lower frequency. Maybe I'll get four higher-end sticks for better performance but at this point, I shouldn't; I'm hoping to build a new rig by the end of the summer and I don't think it makes sense to configure it with 4x2GB RAM nowadays. Besides, I'll go with Skylake if I can afford it, so DDR3 might turn out to be a complete waste of money.

Pr0t0, you are chasing one or two tenths of a frame/sec in the games that you're playing.

Agreed, with one interesting exception: in GTA V, I noticed a significant performance improvement when I found my ideal stable memory settings. By significant I mean getting constant, very smooth 51-60 FPS from regular drops to 32-38 FPS, with very severe micro stuttering (horribly uneven frame timing). From this configuration guide by Gamers Nexus (very useful BTW) it turns out that it's because I have enabled the 'ignore suggested limits' option due to the rather small 2GB frame buffer of my GPU, which makes the game use my RAM as a swap for vRAM 🙂

 

[tcsenter]

Interesting! I was not aware of this. It makes me wonder though: how does populating more modules affect the performance of the individual modules?

Consequences of the multi-drop parallel bus on electrical loading and impedance (and also stuff like jitter, skew). The more modules, the more chips (devices) connected to the bus at any given time, the higher the electrical loading across the entire bus. This is what registered (or buffered) DIMMs are designed to overcome (as well as serial point-to-point bus technologies). Good discussion/overview on the first few pages of this paper:

Fully-Buffered DIMM Memory Architectures: Understanding Mechanisms, Overheads and Scaling

And more general/basic (readable) overview of DRAM technologies from this HP white paper:

Memory technology evolution: an overview of system memory technologies