Question 7-zip crashes when compressing large files when I have XMP turned on. How to tell if the memory controller on the cpu is bad or if it's the RAM?

[SteveGrabowski]

Just did a platform upgrade with an i5-12400F, MSI Mortar B660m WIFI DDR4 board, and 2x8GB of G.Skill CAS18 DDR4 3600 RAM and when I enable XMP to get my RAM running at DDR4 3600 speeds either 7-zip fails or my computer crashes when I try to compress PS2 roms of size ~4-8GB into .7z files using 7-zip. They all compress fine when I turn XMP off and my RAM runs at DDR4 2133 speeds. Is there a way I can test whether the RAM itself isn't living up to its stated speed, or if its the memory controller on the cpu that's faulty? I know Alder Lake only guarantees DDR4-3200 speeds so I guess I can't guarantee DDR4-3600 speed since it's considered an overclock but if the RAM is the problem I'm returning it for another DDR4 3600 kit. But would hate to keep returning and ordering RAM if it's the memory controller killing me.

 

[igor_kavinski]

So he would need to do everything manually, starting in reverse from 3600 selected timings and dialing them in at lower speed, cause his memory does not know how to train them by itself.

Is it a two way thing? I thought XMP just provides the data and then the IMC/BIOS tries to tune as close as possible to provided XMP parameters.

 

[JoeRambo]

Is it a two way thing? I thought XMP just provides the data and then the IMC/BIOS tries to tune as close as possible to provided XMP parameters.

XMP has very limited amount of timings and memory voltage in it



So primaries come from XMP, with everything else on Auto, motherboard needs to select sensible values for them and then memory needs to "train" using that set of params selected. MB knows that for say for 3600 speed and say CL15 it would need 1.35V VCCSA and 1.25V IO voltage on for example Comet Lake Z490 motherboard and it feeds that, even if in reality 1.2V would work as well.

You can already see in this pic the tricks manufacturers play with XMP -> this is B-Die, but tFAW is relaxed to 39, when 16 is best performance, and something like 20 or 24 is relaxed for B-Die already. tRFC is also sky high for B-Die. But only primaries are listed on sale sites, so noone really cares. And those who do, don't touch XMP

And if someone thinks secondaries and tertiaries are not important? Well, they can have outsized impact on performance, esp on Alder Lake where memory subsystem was designed for mobile phones and L3 cache would fit right in some higher end tablet. Look no further than our own forums:

https://imgur.com/gdsrxs3

As You can see Robert896 gained 9% more FPS in that 2nd zone on exact same primaries by going from "auto" to hand tuned secondaries/tertiaries. This is of course extraordinary result due to game scaling so well and Alder Lake core being so powerful and underfed, but personally i have seen Linpack gaining 15-20% more GFLops on Skylake going from XMP to handtuned with same primaries.

 

[SteveGrabowski]

Hmm, I guess I learned something in this thread about locked Alder Lake. I would return the CPU and RAM, and go with a 12700KF and some better tiered RAM. It doesn't have to be Samsung B die, but that is a possibility.

Edit: actually, I would return the board as well. Get a Z series board these days if going with Intel. After seeing this thread, it is evident that we should avoid anything locked.

Nah I'm doing a budget platform upgrade from my 8 year old Xeon E3-1231v3 + MSI Z97S Krait + 2x8GB DDR3-2400. That would be like a $700 upgrade when factoring in an aftermarket cooler and I was looking at more like ~$400. Even at DDR4-3200 Gear 1 speeds (which passed a couple of hours stress test at CAS17 thankfully) the 12400F should be stronger than the cpu in the PS5/Series X so should be enough to last me the current console gen since I only target 60 fps. I think if I was in the market for a $700 platform upgrade I would have just waited for Zen 4 so I could get AVX-512 for RPCS3.

 

[SteveGrabowski]

Hmm, I guess I learned something in this thread about locked Alder Lake. I would return the CPU and RAM, and go with a 12700KF and some better tiered RAM. It doesn't have to be Samsung B die, but that is a possibility.

Edit: actually, I would return the board as well. Get a Z series board these days if going with Intel. After seeing this thread, it is evident that we should avoid anything locked.

You think that's bad you should see the boards. A bunch of the more budget B660 boards (eg $120 to $140) Hardware Unboxed tested throttled even a stock 12600k.

 

[Shmee]

Fair enough, even as a budget upgrade, it is still way faster than the Haswell CPU. And those boards, yikes...

 

[bononos]

XMP has very limited amount of timings and memory voltage in it
View attachment 62724


So primaries come from XMP, with everything else on Auto, motherboard needs to select sensible values for them and then memory needs to "train" using that set of params selected. MB knows that for say for 3600 speed and say CL15 it would need 1.35V VCCSA and 1.25V IO voltage on for example Comet Lake Z490 motherboard and it feeds that, even if in reality 1.2V would work as well.
.............

Is it a good idea to hand tune AM4 Zen1/2/3 as well? How different is AMD vs Intel on XMP ?

 

[DrMrLordX]

Is it a good idea to hand tune AM4 Zen1/2/3 as well? How different is AMD vs Intel on XMP ?

Yes, it is a good idea to hand-tune AM4 setups. There's an excellent tool for getting people started, too! Ryzen Memory Calculator. YMMV, but it has helped me a lot on my 3900X.

 

[RadiclDreamer]

Check your voltages, Ive seen XMP not apply the correct voltage for the frequency/timings