T-Force Vulcan 3000 16-18-18-38 (my OC thread)

[brainsoft]

I'm working my way through the RAM OC process. I'm right at the beginning after restarting, i'm working my from 3000mhz XMP trying different speeds while leaving everything else at XMP specs.

Without loosening the timings it appears that 3030 (BLCK=101) is the highest speed I can go with XMP timings of 16-18-18-38.
I can move up to 3400 with timings at 18-20-20-54 and the benchmarks seem quite happy.

I will come back and edit this post with more details shortly, but answer me a few questions.

1) AIDA64 cachemem test shows bandwidth and latency. How to I determine my theoretical maximum and efficiency?
2) Organizationally, any advice on taking screenshots, storing, name files?
3) Tracking changes. Has anyone made a truth table spreadsheet to work through and check things off at different speeds to help keep track of this wonderful process?

For the first high-speed testing I was able to use memtest86 to narrow my window, and using aida64 I have a good rough idea of this run is less than that run (45k, 44k, 42k ~70ns is a good run), but I am too systematic for that and i'm jumping around because I'm not tracking changes well.

EDIT:
Here's my system:
AMD Ryzen 2600 @ 4.0Ghz w/ Captain 240mm AIO
Asus Rog STRIX B450-F Gaming
16GB T-Force Vulcan 3000mhz CL16-18-18-38 TLRED416G3000HC16CDC01
MSI Radeon RX570 8gb
512GB Samsung EVO 970 M2
EVGA G3 650W Gold+

Starting point benchmark is posted below.

 

[Johan45]

I would aim for a speed between 32-3600 and try to maximize the bandwidth and reduce latency as much as possible.
Take some notes about changes you make and what they accomplish also don't make too many changes at one time or you can lose track of what helps and what doesn't.
Tuning memory takes a lot of patience and I would suggest using a disposable OS because you can corrupt it quite easily

 

[brainsoft]

Not worried about the OS! Once I get all the values I will wipe and reinstall and put it into primary service, hopefully before Christmas. Right now it is sitting on the desk using one of my two monitors so things are a little cramped.

Main thing is working out the best way to keep notes and what metrics to keep track of for the changes I make. Small change, make note, test, test test, save profile is sucessful, repeat.

I think I'm going to stick at 3400 once this test passes and work in tightening the timings. I had to loosen them even to get to 3066, but the primary timings don't change much to get to 3400. Most everything else is on auto since the ryzen calculator settings don't seem happy with my ICs.

Is this good for screenshots? HWinfo64 sensor data. HWi64 system info (shows ram timings and cpuid stuff), aida64 cachemem.

 

[brainsoft]

Okay I went all out and made a spreadsheet to track every single value, I've got a stable speed at 3400 CL18-20-20-54, running the ram at 1.4V. so I'm going use that as a good bandwidth to work with and start tightening up the timings. I will play with the primary timings first because it should be pretty quick as I think I did that already to get to this point. Any suggestions on where to start for the next phase? Right now everything other than primary is set to full auto I think.

This board has full timing access so I have a lot of numbers to choose from. I may start by taking the secondary timings from the ryzen dram calculator moving half way to them to see if I can shave off a few steps. I'd like to pick a target value for a single attribute and if it doesn't work, cut it in half and try again to quickly zoom in on a target before starting full scale testing to confirm stability.

 

[infrared404]

It might be worth taking a look at a program called "DRAM Calculator for Ryzen". You select your configuration, click R-XMP to read the spd data, then it'll give you suggested timings for whatever frequency you're trying to run. Clcking Safe,fast,extreme will give you a different set of timings and voltage recommendations. It's really useful for setting the sub-timings, getting these right can really help with stability.

18-20-20-54-78 are very loose timings, I'd suggest seeing how far you can overclock with the timings from the 3000mhz xmp profile (16-18-18-38-56) and give the aida benchmark a try again. If you can do 3200mhz with those timings you should be golden.

If you still want to play around with the looser timings, try 18-20-20-42-62. 54 tRAS and 78 tRC are looser than necessary imo.

edit - another program called Ryzen Timing Checker can be useful to confirm timings have applied correctly.

 

[brainsoft]

Looks like this was binned right at the top of the scale. I was not even able to run 3066 without memory errors unless I opened the timings, even at 1.4V.

The dram calculator is not reliable for this micron Ram using any of the die selections available. The dram checker would not run on Win10, I did not spend any time trying to figure out why though.

 

[brainsoft]

Okay I returned to square one again now that I have made a wonderful spreadsheet to help me track. I decided to do some benchmarks with the board in full auto @ 2400mhz, then again at 3000mhz XMP, full auto. I've not done layers of stress testing at for the 3200 stuff, but I've got my most recent 3400 running over night and I'll play with a few more quick benches in the morning to try to get my bearings in this timing game.

2400 CL16-16-16-39-55 (Full auto)


3000 CL16-18-18-38-68 (XMP Auto)


3200 CL 16-19-18-38-73 (crashed at 16-18-18-38. Poor CB15 score seems to be caused by manual cpu over voltage!)


3400 CL18-19-19-40-60 (seems stable, lowest latency so far and good scores)

 

[Johan45]

Looks better, those Microns won't tighten up a whole lot so you're on the right track

 

[brainsoft]

How can I calculate my efficiency? Should I move back to 3200 and see how tight I can close things up instead?

 

[Johan45]

Wherever you can achieve the best latency which appears to be your 3400 settings

 

[brainsoft]

I've done many rounds of memtest at various settings and most of my recent benchmarks throw memtest errors. I'm reconfirming my last confirmed stable setup but I've had to open things right back up again to 18-20-20-48-62, might have to go back down to a lower speed and see where the trade off is

 

[brainsoft]

If memtest keeps failing on test 7 - "moving inversions, 32-bit pattern" does that point to a specific parameter I should look at loosening?

 

[Johan45]

I would try tRC tRFC and tFAW first.

 

[brainsoft]

Okay, here's an update. I am testing working my way through 3200mhz with a couple memtest rounds each day and a good stress test over night. I've worked down to 18-19-19-37-56 with tRFC=560 on auto, and it is stable. I lowered tRFC to 550 and it failed memtest. I raised it to 555 and I got only a couple of errors.

Once again the errors are from Test 7 moving inversions 32-bit, but I noticed something else. The majority of errors are displaying as CPU: 9 as well. Is that a weak core or something? I've got the CPU on full auto, with +0.05v offset to give it just a touch more juice. SOC is manual at 1.1V, and DRAM is 1.39. I have lots of ventilation so I don't mind going a bit higher on it if it will sort that core out.

OR is it just a coincidence that it just happens to hit a particularly tough pattern in the exact same order every time because it's always running the same pattern of patterns?

I've otherwise still left everything on auto, gear down mode is enabled. I've been reading non-stop to try to get a grip on some of the meanings of these values but it is slow going to actually remember what anything does. I've added a couple of check cells to my spreadsheet to try to pre-calculate a few rules of them so I don't waste time pushing lower than I should.

Can some one spot check these? I've got a few conflicting rules, though i know "rule" is often an approximate guideline.

I've read a lot that tRAS should be CL+tRCD+tRP, but that just doesn't add up for any of the SPD/XMP profiles on this ram.

* tRAS=CL+tRCD seems to work pretty solid for me. For me it's currently tRAS(37)=CL(18)+tRCD(20)
* tRC=tRP+tRAS also seems to work pretty soli. For me it's tRC(56)=tRP(19)+tRAS(37). I was not able to get CL, tRCD or tRP any lower at 3200 so I guess that sorts out the primary timings?
* tFAW=4xtRRD... I'm wondering about this one. My ASUS board has a tRRDS and tRRDL (short and long delay) and they're currently set at 6 and 8 respectively. This would make tFAW no lower than 24 or 32. Auto sets it at 34, any thoughts?

Are there any other common "rules I can watch for? I have tRFC, as well as tRFC2 and 3, is there a ratio for those? I think tRFC will lower from 560 on auto to 556 but it won't currently go any lower than that.

Do I just keep working my way down the list, find the lowest value I can get it stable at then move to the next parameter, or should I be doing a couple of them slowly one at a time? I haven't found yet a good straight forward list of where primary/secondary/tertiary timing labels are divided or what every single one means, but I'm sure it's out there.

- - - Auto-Merged Double Post - - -

This is the last screenshot I grabbed. I've also found that there must be something running in windows that is scewing the benchmarks, I see FPS in cinebench drop ~10fps sometimes. I've worked out that if I stresstest p95 512k-4096 for the day then benchmark (computer will have been up for many hours instead of a fresh boot) then I get better results then if I boot up and benchmark any reasonable time after boot.





I've been working on one column until I get a solid memtest, then prime95 and then bench and record final numbers including auto values, lots of colours to help me track changes.

 

[brainsoft]

This post on ASUS Form had some good tidbits. Setting tRFC back to auto and making my way down the list

 

[Johan45]

Don't worry too much about formulas, they're more of a loose guide, tFAW at 34 seems about right. Just keep working at it the way you are. You're making headway but as I said tuning RAM can be a long process.

 

[brainsoft]

keep running up against "CPU 9". This time i tried lowering tRRDS from 6 to 5 and got a handful of errors in test 9 modulo 20, again on CPU9. It's always cpu9. Any suggestions? more CPU voltage? I have lots of headroom on that as it's running full auto now. It will run 3900 full stable without too much voltage. I bumped up SOC one step from 1.1 (1.1125 I think) but that has not helped.

I am having a lot of fun, I love numbers and optimizing, it's as close as I can get to car tuning in the winter.... change something and see how she runs! Nothing is making any real change to latency, I imagine these ICs don't have a whole lot of headway on them. If they performed slightly worse they'd be a great value in the next lower bin, but they barely made it into the 3000mhz xmp bin it looks like.

 

[brainsoft]

I wanted to do faster cycles so I got the pro version of memtest so I can save the config file and I've been doing just test 5/7/9/13 and just for the area of memory surrounding the block that gives every error (0x3C25Dxxxx). i've been able to pump through the tests very quickly now, but I noticed that test 12 in the pro version (128-Bit test) throws errors. I never would have known with only the free version... should I be concerned? I've continued for the moment with that test disabled...