3866 MHz-Too much is enough.

Alaric, it will likely be 5% performance increase from fast ram and you probably won't even notice it in real life. I too saw negligible boost from faster ram OC which is unnoticable in real world apart from benchmarks, but since we're OCF, 5% is totally worth it.

Yeah, but I'm hoping someone familiar with what the numbers really mean will stop by and give me an idea how big a difference before it really matters and which differences I should focus on first.

IMO ram oc get me 1-2 fps in most games , not worth the time for me to find stable oc , I try the basic's now a days up the speed multi when it stops working thats where it says . Or stock for weeks on end if I get a crash .
Benching is another story it can get you over that last hurdle from your goal score some times .

We have a winner

After countless hours of testing and hair pulling, I have my daily driver OC. Voltage shows as 1.452v in software, which reads higher than measured, so I'm at 1.45v or less and I'm happy with that. To meet my voltage goals and OC jones, the below shot is it. I ran numerous Aida benches on the same setup to get a feel for the margin of error on reads, and 3600, and 3866 are both in the ballpark. 3733 was the only speed to ever go over 52,000 MB/s and it consistently stayed there. After various compromises for voltage requirements and tightening some timings (tRFC is 275) I hit on a combo that delivers consistent 52,000+ MB/s reads and sub 39 ns latency. The only other number that consistently changed was L2 cache write speed, and this OC package stays at the high end of that (301-320 GB/s).
Unless/until someone tells me I got it all wrong, this is the best this rig with this RAM will do on 24/7 safe voltage levels. Prior to this, every single time I changed anything having to do with memory I got instant BSOD and the dreaded "BOOTMGR.EXE is missing". I learned a lot from my efforts but that pales in comparison to what I got from everyone who contributed to my efforts. Thanks, guys.




And this was my highest stable speed, with enough voltage. Aida has problems with tRAS speed. It shows 42 below, was actually 38. 4000 MHz took too much voltage to even consider messing with it further, but it ran @19-19-19-38.



My tightest timings award goes to 3600 MHz@14-14-14-28



Besides being proud of this (all out of proportion to the accomplishment), this documentation may come in handy when I can't find my notes and need some data from this adventure. LOL
Thank you again to all who took time to help me out.

Or stock for weeks on end

Click to expand...

I have a hard time leaving anything stock. Almost anything mass produced ends up being a compromise, or worse, a collection of compromises. Just about everything can be improved. And I need a hobby. LOL

Alaric, take up model aviation! You get to be outside every once in a while instead of pulling hair behind the screen, and lithium battery management is like a hobby of its own

lol, battery powered planes... gas or go home!!!!

Uh sir, i don't mean to derail thread, but this isn't 90's. Lipo has advanced quite well nowadays..

Good work Alaric. Takes time and patience to work with memory and you managed quite well.

Johan45 said:

Good work Alaric. Takes time and patience to work with memory and you managed quite well.

Click to expand...

Apparently it had been a while since I last used my brain, and some things became obvious as I went along. Running the same set up multiple times through Aida gave me a fair idea of error margin and what to expect without fretting over minor differences and instead look at trends. My latency vs. Read speeds question was answered when I gathered up all my results and looked at the "big picture". Read speeds=MHz and Latency=timings. So I went in the order Highest speed>tightest timings, because I got better results from keeping speed the first priority. Since this is my do everything rig I had to compromise for rock solid stability and voltage as the final deal breakers. I kept notes on everything, though, so I still have some pretty fast options later on. All in all a good time was had by all. All being me. Some time today I'm going to bench the XMP profile and get the percentage of improvement from the OC. I'll post that up today.

UltraTaco said:

Alaric, it will likely be 5% performance increase from fast ram and you probably won't even notice it in real life. I too saw negligible boost from faster ram OC which is unnoticable in real world apart from benchmarks, but since we're OCF, 5% is totally worth it.

Click to expand...

More like 0-1% performance gain ... still worth it
On AMD it could be even performance drop ...
I run 32GB SODIMM@4000 in quad channel right now without any special reason. When I switch rigs then I will try to set something higher.

btw. to see how system works on memory, run command prompt and type 'winsat mem'

Woomack said:

More like 0-1% performance gain ... still worth it
On AMD it could be even performance drop ...
I run 32GB SODIMM@4000 in quad channel right now without any special reason. When I switch rigs then I will try to set something higher.

btw. to see how system works on memory, run command prompt and type 'winsat mem'

Click to expand...

I'll add the XMP one shortly.

XMP winsat mem results. This is out of the box conditions, no CPU OC and XMP enabled. I'll run it again with just the CPU OC and see if it adds anything so I can narrow down just what the RAM OC did. I'm good with an 18.75% bump, though.


A last surprise. Seems the memory didn't like the CPU OC in XMP. Ran it twice just in case.

Here's what I got
View attachment 196215

Alaric said:

I have a hard time leaving anything stock. Almost anything mass produced ends up being a compromise, or worse, a collection of compromises. Just about everything can be improved. And I need a hobby. LOL

Click to expand...

Only the ram this cpu has been on for 95% of the time i have had it and overclocked 94% of that time =)

I had about 77GB/s on quad channel 4000. You can compare winsat results to memory copy in AIDA64 and in most cases it's about the same. I guess that whatever is improving bandwidth in winsat then will affect performance of whole PC. Some benchmarks are showing higher results but they're not affecting daily work.

Patriot Viper LED 3200 / white LEDs

I guess it will make 4000+ on Intel mobos. On Ryzen runs without issues at 3466, maybe higher but I have only MSI B350 mobo for tests right now.

The final settings I have are as follows:

CPU @4600 MHz
Uncore @4200 MHz
vCORE 1.352v-measured
VCCIO-measured 1.194v
vDIMM-measured 1.450v
______________________________________________________________________________________________________________________________________________________________________
Timings Changed By Me

RAM @3733 MHz
Multiplier x37.33
CAS 15
tRCD 16
tRP 16
tRAS 31
tRC 48
tRFC 275
tREFI 14559
______________________________________________________________________________________________________________________________________________________________________
Timings Left On Auto In BIOS

tRRD 7
tWR 24
tRTR 7
tRTW 12
tWTR 9
tWTW 9
tRTP 12
tWTP 42
tFAW 40
tWCL 14
tCKE 8
tRTL 65
tIOL 4
BL 8
______________________________________________________________________________________________________________________________________________________________________
Results

Memory Read 52,490 MB/s
Memory Write 55,265 MB/s
Memory Copy 48,270 MB/s
Memory Latency 39.5 ns

winsat mem 45,038.56 MB/s, up from 35,887. That's a 25.5% improvement. Memory Read speed is up 22.5% and Latency is down 6.8%.

The overall effect on performance may not be Earth shattering, but if I had to buy the performance I now have, from scratch, it would run me $889 for a Z270 direct replacement for my mobo, a 7700k and G.Skill 3866 MHz RAM (because my timings are tighter than their 3733 MHz kits) at newegg prices. So what if my overall performance only went up a couple percentage points? I'll take $900 worth of free stuff any day. And almost everything I did, I learned at OCF.

Nice work Alaric! Thanks for all the updates.

Woomack said:

I guess that whatever is improving bandwidth in winsat then will affect performance of whole PC. Some benchmarks are showing higher results but they're not affecting daily work.

Click to expand...

Thank you,

I did notice one thing that seemed odd at first, but I'm sure that's due to all the things I don't know. With the RAM at the XMP, the rig performed better with default settings for the CPU than it did with a 600 MHz OC. So, was the cache outrunning the RAM by a wide enough margin to fudge up the works? It wasn't a big difference, but it was there. I ran winsat mem twice on the settings to confirm it. Overclocking the CPU alone actually cost me RAM performance somehow. Overclocking both picked up more than I expected. I may go back and run everything again with the RAM OC and the CPU at default speeds and see if I find any clues. Might be a good idea to try default CPU with default cache and with OC cache, too. I may not know enough to glean the answers, but if I throw enough data up here maybe someone smarter than me can get something from it. This is turning in to a research project. LOL

I also noticed that raising the cache speed gave small but definite gains in the memory read speeds, with almost zero effect on the cache read/write speeds, the exact opposite of what I expected.

I had about 77GB/s on quad channel 4000

Click to expand...

Holy $&*!

Alaric said:

I did notice one thing that seemed odd at first, but I'm sure that's due to all the things I don't know. With the RAM at the XMP, the rig performed better with default settings for the CPU than it did with a 600 MHz OC. So, was the cache outrunning the RAM by a wide enough margin to fudge up the works? It wasn't a big difference, but it was there. I ran winsat mem twice on the settings to confirm it. Overclocking the CPU alone actually cost me RAM performance somehow. Overclocking both picked up more than I expected. I may go back and run everything again with the RAM OC and the CPU at default speeds and see if I find any clues. Might be a good idea to try default CPU with default cache and with OC cache, too. I may not know enough to glean the answers, but if I throw enough data up here maybe someone smarter than me can get something from it. This is turning in to a research project. LOL

I also noticed that raising the cache speed gave small but definite gains in the memory read speeds, with almost zero effect on the cache read/write speeds, the exact opposite of what I expected.


Holy $&*!

Click to expand...

Where's Bart?