Sacrifice MT for CL?

@Rogue-agent
A word of caution there… a few years ago I purchased a 4133 memory kit for my new Ryzen 3000 build foolishly thinking it would future proof my next upgrade. I had calculated the expected latency at 3866 to be CL17 but at that time I didn’t understand geardown mode and its rounding of tCL and other timings to even values. Long story short, after days of tweaking timings the best CAS latency I could get at 3866 was CL18 (after being rounded up from CL17). I tried disabling geardown and desynchronising MCLK and FCLK but that just led to even worse performance. In the end the memory benchmarked highest at 3600 CL16…

TDLR; choose a kit with an XMP profile or do your homework first. I bought some really expensive “3600” memory.

3 Likes

yeah you should have a setting to disable gear down mode.
i ran into that when i got cas 15 ram and it would only allow 16 with GD enabled.

like i said he should check the jdec of the 4000 ram to make sure it can run at 3866.
it should as its only 1 step above and doesn’t have the markup of 3866. which is why im thinking its an attractive alternative.

its also generally easier to underclock ram than over clock it.
thats not to say you cant run into issues, just way less often.
add in the likes of ryzen dram calculator which takes all the guess work out of underclocking.
i still think this is the best option over the 3600 or the 3866. :slight_smile:

2 Likes

Yes, I’ve looked at it before but was way too busy so I just ended up running at XMP speeds. This time I’ll try to figure it out and see how it goes.
But in any case in LTTs video they said it should be as simple as increasing voltage by .1 and then tightening the timings by 1 and tRAS by a few.
We’ll just have to see.

1 Like

You really don’t want to fuss with high frequency and low latency for a workstation. Reliability is what you want. There’s very few workloads that will appreciate from memory frequency in a workstation, and even fewer that really genuinely benefit from lower CL.
Chasing overclocks on your memory is for purely gaming PCs where you don’t care if it bluescreens now and then for a few months or years.

3600 CL16 is already pretty good, and lower latency than 3866 CL18 already. Generally, if you’re at 10ns or lower, or in other words, 1/200th the advertised MTs for DDR4, you’re already doing pretty well.

2 Likes

I’d go further than that and say chasing these things is for competitive benchmarking. Yes maybe you can compare frame rate metrics and see an improvement but honestly the difference is usually difficult to detect when playing the game. You simply aren’t going to take something from unplayable to playable.

Few things are more annoying than pc crashing mid game. Less critical but almost as annoying as a workstation doing it.

And as much shit as Microsoft cop for that, if you’re seeing regular crashes in windows in 2022 it is hardware or driver related.

Depending on the platform, overclocking memory can be pretty substantial in games, especially for things like minimum frame times, which impact realtime software performance, like games, which are very sensitive to that, but doesn’t really matter in most professional softwares, where it often won’t even make that kind of substantial impact anyway.
I mean, you could possibly make a case for something very latency sensitive like fast sketching, but the added potential instability and lost work isn’t worth it at all.

For games, though? You can sometimes see quite substantial gains from a bit of memory tuning. It’s an enthusiast thing for sure, but not really that unreasonable to appreciate. I’ve seen linear gains in a game before. Went from 62~65fps up to 85~90fps in one game, which only responded positively to lowering tRAS and nothing else.

1 Like

Sure, but you can get 90% of that pretty easy without tweaking to the nth degree.

I’m not saying run JDEC speed :smiley:

More like, don’t obsess over chasing speeds higher than 3000 or so if it proves difficult - and again especially on the 5900 and 5950 you’ve got double the cache which helps to hit the memory less in the first place.

70 M of cache is no small amount.

5800X-3D? Even less point…

Very easy to spend more time tweaking and crashing than playing games, i guess that’s my point :slight_smile:

My process: try XMP, and then progressively lower clock until its stable, then tweak CAS based on what HWINFO64 shows as valid profiles for my RAM.

E.g., on this 5900X i’m running DDR4-3600 at 3200, based on this:

They’re dual-rank 3600 sticks but those timings highlighted are validated and work for me. Took like half an hour and works well.

2 Likes

Yea true. But this will be a custom Linux

Picosecond calculations. Where I’ll have a nuclear clock to my workstation, so i kind of need very low latency on computer.

Yea, but it is something new to learn and experience. Some day I’d hope to be able to use this knowledge.

1 Like

I thought R5k decoupled stuff from the memory controller

you can force it to decouple but you face a latency penalty, you also can make it run gear 2 if you want to run your ram stupid fast

if only it was that simple… :smiley:

3 Likes

looking at your timings.
your bus speed looks to be drifting by nearly 1% to 99.1 when it should be 100 +/ 0.2% (99.8-100.2)
is your cpu bus ratio timing set to auto?
if so set it to manual and 100. it should lessen the drift and give more stability.

you will notice im well outside the jdec. the ones listed in the jdec only sort of worked so had to use different more expansive timings.

but also look at my bus speed. (blue box)

to get my clocks, i used the ryzen timing checker and got different results than the jdec.
they are data stable and tighter than the jdec specs on the ram.
at cas 14 its sub 10ns first word return.

2 Likes

Yeah basically everything on my board at the moment is auto save for

  • turning on AMD virtualisation
  • XMP then tweaking timing
  • setting XFR to default rather than whatever gigabyte think
2 Likes

I was recommended this, is this sufficient?

KF436C16RB1K2/32 is a kit of two 2G x 64-bit (16GB)
DDR4-3600 CL16 SDRAM (Synchronous DRAM) 2Rx8, memory
module, based on sixteen 1G x 8-bit FBGA components per module"

They’re dual rank (like most 16G sticks I think). So you may have trouble hitting 3600 at least not without a bit of tweaking.

Which brings up one thing: if you are planning a heap of tweaking make sure to buy a board with a clear cmos button, ideally on the rear IO shield.

Otherwise if you try to push too hard and prevent your system from being able to post it is a pain in the balls to open the case and possibly remove your GPU or a stick of ram or short some pins (hopefully correct ones) to clear CMOS.

You can tweak memory on any system but if you plan on trying too hard with it a clear cmos button preferably outside the case will make it so much less painful.

Yes that may blow your board cost out for you but it’s worth it.

1 Like

Isn’t it easier to just connect your reset button to clear CMOS jumper?
Its not always easy to reach the backplate.

Depends if you use your reset button more than your clear cmos.

In my case I also have a cat who sometimes walks on the case and it’s exposed power and reset buttons :joy:

I tend to just take out the battery and hold pwb while the power is disconnected

not sure what you think your doing with that mate.
if you pull the battery and trurn the pc on all you will do is boot into a default bios.
as soon as you put the battery back in the bios chip will read the cmos and load your previous settings.

to clear the cmos via the removal of the battery you have to:
remove the battery, turn the pc on and off again to power the motherboard with residual power.
then short the +/- terminals in the battery.
put the battery back in and power on. you will clear both bios and cmos.

Yup, that’s also an option, but sometimes pulling a stick of ram to get into the BIOS with too aggressive ram timings is easier than getting to the CMOS battery if its under your gpu :smiley:

1 Like