Lowering latency with 5800X3D and Micron E Die

[Chomiq]

My current daily setup at 1900 FCLK 1:1:

PBO2 with default limits, All Core CO -30

Other then 10 total Whea 19 corrected errors during a course of 6 months I don't really have any stability issues:

Identical timings had me at around 54 ns on 5800X, any tips?

 

[Ferrum Master]

start with tRFC...

 

[Taraquin]

Try disabling GDM. That may require changing DrvStr to something like 40 20 30 20. Also you may need more SOC\iod voltage, try upping both by 20mv and see if latency improves. RFC may go a bit lower. 544 maybe.

 

[Veseleil]

My situation is different as I run 4 sticks, and this is the best I could do without errors:



I've also experimented with CL15, but 2T just puts me off. You can read about it more in this thread, not a long read, it's just slightly over one page.

Edit: Added newest AIDA64 result as a reference.

 

[dgianstefani]

My current daily setup at 1900 FCLK 1:1:

Identical timings had me at around 54 ns on 5800X, any tips?

I know it sucks to hear, but it's E die, for lowering latency by a meaningful amount, you want B die.

5800X gets lower latency because the CPU clocks higher.

tRFC is a good place to start as Ferrum mentioned.

I would increase TREFI too.

Although both of these tweaks will increase DIMM temperature a bit, so ensure you have decent airflow.

WHEA errors indicate instability, even if you're not crashing. Run Karhu to ensure you're actually stable. 1 hour tends to be sufficient, 24hrs if you want to be 100% certain.

 

[freeagent]

You will only get good results with B-Die. Literally everything else is trash. Not trash but.. for the numbers..

 

[dgianstefani]

This table can be useful.

To clarify, Samsung B die, since apparently Micron makes a B die too lol.

 

[Veseleil]

You will only get good results with B-Die. Literally everything else is trash. Not trash but.. for the numbers..

It's Micron E die 19nm, which is the best (I'd say only) alternative to Samsung B die. It's the most overclockable DDR4 RAM, and its better binned versions AFAIK still hold a WR. Just ask @buildzoid. Not tight timings as a Sammy B, but it's far from being trash.

 

[Chomiq]

WHEA errors indicate instability, even if you're not crashing. Run Karhu to ensure you're actually stable. 1 hour tends to be sufficient, 24hrs if you want to be 100% certain.

Yeah that's why I bumped my VSOC and VDDG IOD yesterday as a lot of people suggest that WHEA-19 errors are a sign of unstable FCLK.

I'll focus on ironing that out, since the original AIDA64 had 69 ns (and that's why I made the post) but subsequent runs showed around 62 ns which seems to be near what Zen 3 3D leaderboards show for similar setup.

 

[freeagent]

B

It's Micron E die 19nm, which is the best (I'd say only) alternative to Samsung B die. It's the most overclockable DDR4 RAM, and its better binned versions AFAIK still hold a WR. Just ask @buildzoid. Not tight timings as a Sammy B, but it's far from being trash.

But you don’t need speed with 58X3D. Well, you do but tight timings are better for the particular bench. Clamp tRFC down as tight as you can. I also didn’t mean that your ram was trash, but for benching with programs that love tight timings it is the way to fly. The last Aida bench showing latency at 63.9.. that is like what I get at 3200C14.

 

[Veseleil]

B

But you don’t need speed with 58X3D. Well, you do but tight timings are better for the particular bench. Clamp tRFC down as tight as you can. I also didn’t mean that your ram was trash, but for benching with programs that love tight timings it is the way to fly. The last Aida bench showing latency at 63.9.. that is like what I get at 3200C14.

I don't feel insulted, just clearing up some facts. I don't really care about benches, but the feeling of having a well tuned and maxed out system is just one part of being an enthusiast you'd agree.
Sammy B isn't just about the premium price, but also about availability around here. A couple of years back I was searching for a 32GB (2x16GB DR) Samsung B die, new or used. Not a single kit came up in search (it's easy to distinguish them just by MHz and CL numbers). So I was happy to find another identical Micron E kit, they're becoming rare too.
Sorry for the off topic, I'll walk myself out RN.

 

[freeagent]

I bought mine 4 years ago, all ram was expensive then.. but I do recall some of the guys recommending that stuff at the time.

 

[ShrimpBrime]

Yeah that's why I bumped my VSOC and VDDG IOD yesterday as a lot of people suggest that WHEA-19 errors are a sign of unstable FCLK.

I'll focus on ironing that out, since the original AIDA64 had 69 ns (and that's why I made the post) but subsequent runs showed around 62 ns which seems to be near what Zen 3 3D leaderboards show for similar setup.

Right where it should be. Can't OC the cpu, so you loose out reducing latency from the cpu overclock. Your old 5800X was probably capable of 4.8ghz with little effort, I'm guessing. It may be what the difference is. Knowing at stock the X3D os clocked lower, this will for sure be the reason.

Measurements.
I think you Know better than to rely on AIDA64 for accurate measurements. It's not consistent that I can tell.

If you want to know for sure that you have reduced latency with a benchmark, try PiFast as well. It's sensitive to memory tweaking, less so to CPU frequency.

Any hoot, Good Luck!!

 

[freeagent]

I am running 4 sticks as well. My board is a bit lame for this test, my B550 board runs this CPU much better. This board is good for 12 cores and up power wise.. but so is my Strix XE

 

[damric]

hmmn in my main rig I have 4x8GB of these Micron E. They are not fine tuned by any means, but they are 100% stable at anything including folding@home and I do not get WHEA errors.

Could try a few of the settings if you want: