Hacker News Reader

Comment by bravetraveler | Hacker News Reader

With enough time sunk into it, these numbers are more... suggestions, anyway.

My system has a 7950X3D and four sticks at DDR5-6600. The module density, board trace topology, and luck-of-the-draw on IMC/SOC can matter a fair bit.

Tuning voltages/timings can make up differences. Also remember the infinity fabric clocks and such being somewhat paired

View on HN

May I suggest to also run memtest if you're gonna tune these kind of things. Your computer may appear to work normally after changing the settings, but memtest can reveal unnoticeable memory errors.

Good suggestion, though I have - Prime95, memtest, many long workloads :)

With DDR5 and really, most modern hardware, having some more robustness (that pseudo-ECC whatever, clock stretching)... it might be worth considering performance as part of your validation

Sounds good. I didn't mean to make a suggestion personally to you by the way. Sorry if it seemed that way. English is not my native language. My remark was more targeted towards the reader of the thread who gets curious and may start experimenting.

Totally fair :) I err towards defensive, no harm done

memtest is a pretty basic test and is not very reliable at triggering errors in memory pushed just past of what it is capable of. It's fine to run it as a smoke test before booting into the operating system (to make sure it won't crash right away), but after that you're gonna need something like TestMem5. I keep a Windows install just for this.

Cool, didn't know about TestMem5 until now. Thanks!

What's the first word latency/CAS latency? I think EXPO 6000CL30 kits can achieve 10ns, so if you have to run 6600CL40 or something it may be worse.

It's CAS 32 - one doesn't have to go sky high latency to get high frequency, the binning is truly incredible these days.

Just a boring Corsair Vengeance kit with OOTB settings as far as they're concerned. The CPU needed help

Nice, which ICs does the RAM kit have? I don't think your experience is usual by any means with uclk:fclk 1:1 though.

Can't say I've been brave enough to tear them apart :D I took the lazy way out and just sorted by specs on the listing.

I actually expressly didn't want to overclock the memory, just get something that ran nicely as-is.

I'm actually not at 1:1... stable. Perhaps wires got crossed? In another post I mentioned it POSTs but that's about it :( I may be able to get something there, but can't promise or over-represent it.

The memory is at 6600 but I have some divisor or whatever at memclk/2. This is super stable

6000 at cl32 is most likely going to be Hynix A die, the cas latency could be better if they push like 1.8V into the memory but it's not a particularly important timing for performance

Is this with the EXPO profiles, or are they running at the JEDEC transfer rate?

This is indeed with EXPO! There were great strides made with firmware revisions in the last year or so.

I do have to make a point to boost the supporting IMC/SOC voltages before enabling that, or I can get a bit wedged.

The profile voltages (and clocks/timings) for the sticks themselves work fine - just the board/CPU itself needs some care.

Is that running at 1:1 uclk/mclk?

Glad you asked! I just checked and tried, no/not yet.

It's at 1:2 or 2:1 or whatever makes sense, my firmware shows memclk/2 - haven't had my caffeine yet.

The memory can train/the system POSTs, but the OS fails to load. With some more fiddling this may be feasible - I'm not sure. Excited to try!

Additionally: the IF frequency is 2100Mhz, trying to go much higher doesn't work out. If I could get 2200 it'd factor nicely into the memory frequency too, but alas - chip isn't up to it.

you could try 6400 on memory synced if 6600 boots with 4 dimms which is a very good result. 6600 is into the territory of needing luck with IO die and then also tuning all the voltages.

For the infinity fabric if you're trying to push it checking for performance regressions is a good idea as it delays and retransmits as a form of pseudo ecc when it detects an error. Same with memory if you're not using all custom timings as motherboards can set some very much not good timings if left on auto and sitting on the edge of stability

I appreciate the tips, will give it a try :) When speccing out this system/building it, I chased memory frequency as we can tell.

Seeing 1:1 is potentially within reach, perhaps dropping down to 6400 is worthwhile. Thanks again, will give it a whirl!

While I have chosen the EXPO profile, history has shown boards can't reliably read the (sub)timings... so I do it/apply those :P I try not to mess with those much, I don't really know the relationships between them.

For the most part it's usually best to just go lowest that's stable, there are some "rules" online but most of them are just random things someone came up with. There are a couple JEDEC ones but the memory controller will adjust to work with those or it'd just not work.

A 1:2 sync usually starts being worth it above ~7200MTs but I don't know how viable that'd be with 4 dimms, at the very least it'd need tuning the various termination impedances

Reply on news.ycombinator.com