[SOLVED] X79 - Memory performance differs with CPU straps?

[pr0t0typ3]

If you have an answer to this, you are a true OG. I've looked everywhere and found no explanation.

My build:

• Asus Rampage IV Formula (X79)
• Intel Xeon e5 1650 v2 (=4930k, unlocked multiplier)
• 4x8GB Kingston Hyper X Savage (quad channel kit, 1866Mhz CL9)

I spent a LOT of time recently learning about RAM OC, and have been experimenting for weeks to push the limits of my config.

The stock settings of my memory are:
1866MHz 9-10-11-27 (CR2).

I have found my sweet spot stable memory OC with the best effective latency at:
2500MHz 11-13-13-30 (CR1), all secondary and tertiary timings manually tightened.

The x79 platform was among the first which had BLCK tied to PCI-e frequency, something that is still true on the newest Intel platforms. BLCK overclocking is therefore limited to the usual +5-7 MHz (if you are lucky), and for this reason straps were implemented on the platform, which change the PCI-E : CPU and PCI-E : DRAM frequency ratios for more OC flexibility.
The straps are 100 (stock), 125, 166 and 250. When set, they effectively become your BLCK and get multiplied with the multiplier you want, and RAM frequency options also change accordingly.

My question:
How is it possible that with the stock 100 CPU strap, my OC is stable but performance is not nearly as good as with the other straps? Write performance is actually worse than stock, and latency is only marginally better (with my previous 2400MHz 11-13-12-30 OC it was actually worse at around 62ns).

Testing:

• I kept CPU frequency as close to 4.5 GHz as possible with each tested setting, but the focus was on keeping RAM config the exact same, so there is some variance in CPU speed
• I used my memory's stock performance as reference, everything auto except CR which I set to 1 (I hate to leave it at 2)
• for all other tests, I tuned memory frequency to exactly 2500 MHz and used the same set of manually configured primary, secondary and tertiary timings
• I used AIDA64 and ran each test 5 times individually from the benchmark menu (always do this instead of the cache & memory benchmark, because this way you don't have to wait for the irrelevant cache benchmarks to finish every time and, more importantly, for some reason run-to-run variance is much lower when testing each benchmark individually).

I made two simple tables with the results, however I can't directly insert pictures here and Google Photos links were not accepted. I linked the images below:

Results summary is available here.
Result details available here.

So what am I missing here folks? I can't figure out what the reason for this is and it's driving me crazy... It doesn't really matter since the other 2 straps work wonderfully, but I hate to not understand things as it is killing me.

PS: Unfortunately I was not able to include the 250 CPU strap, even though it worked without issues yesterday... For some reason, today my rig failed to POST with code 00 no matter what settings I tried.

Any input would be much appreciated.

 

[jimmysmitty]

https://www.overclock.net/forum/6-i...125mhz-bclk-hurts-memory-performance-x99.html

Might be an issue with x79. There was a post on Asus forums but I can't open it. It may also be the brand. I can't find anything definitive on your issue just that others have run into it as well.

Also I was able to link your images directly, you just have to use the image itself to do so like this:

 

[pr0t0typ3]

I followed your advice and was able to insert the images to my OP by linking to them directly, but I wasn't able to save the changes. I tried 10+ times and kept getting a popup saying a server error has occurred.

I found only one thread where this memory behavior was mentioned over at the techpowerup forums:

Also worth noting: You'll see a pretty big improvement in your AIDA numbers across the board if you use the 125 CPU strap to reach the same RAM speeds (128.625 BCLK using the 1866 RAM divider will give 2400 exactly). For some odd reason, on X79 the 2400 RAM divider on 100 CPU strap gives lower performance numbers than it should. Cadaveca pointed this out to me when he was helping me get situated on the X79 platform, and sure enough, my own testing showed this pattern.
If you insist on using the 100 strap, or are simply limited to that strap by your CPU, try running the RAM at 2133 and really tightening those timings. Seeing that you are able to run C9 at 2400MHz, I wouldn't be surprised if you could run C8 at 2133MHz.

It looks like this is not brand-specific and has to do with X79 specifically, but I don't understand the reason behind it. I validated earlier that a more significant performance difference only exists at and above 2300 MHz. As opposed to what the quote may suggest, the performance difference is not exclusive to AIDA64, I saw a similar story in Geekbench 4 and 5, and all other memory tests I ran.

Thanks a lot for the link, it took me here which I assume is the ASUS post you mentioned you were unable to open. I haven't read the entire thread yet but it appears to be discussing my issue specifically. It seems to be specific to Ivy-e and the 36:3 memory divider (which gives a memory frequency of 2400 MHz with the stock 100 CPU strap, exactly as shown in my testing).

 

[jimmysmitty]

That is the post, yes. It was giving me a 502 yesterday so I assume server issues. It opens now.

I thought it might be X79 specific as one post I saw a guy went to X99 and didn't have the same issue. Probably just a bug in the system that was never addressed by Intel.