buildzoid patreon

My personal AM5 DDR5-8000 success list

🕑 Added 2024-11-29 02:02:19 +0000 UTC

Comments

Actually Hardcore Overclocking (Buildzoid)

AFAIK 20 is the register limit.

InkyBusiness

Hey bz what does tfaw 20 do? 16, 24, 32 would be my picks. Is there any performance, stability or sideeffect benefit for tfaw 20? To the best of my knowledge tfaw depends on the trrd and the mc phy interlink for the memory page size. For us 1kb. tfaw delay can only occur on act, after a burst.

Matwau

this really helped me out, thanks a bunch!

InkyBusiness

Ah while I am thinking of this in some video @buildzoid you mentioned dfe and Rx tx interconnect dfe taps. They have little to do with training, their value is dependant entirely on memory topology and how reflective it is. Read more here http://www.seas.ucla.edu/brweb/papers/Journals/BRFall17DFE.pdf I'll try to sum it up: dfe taps is how often at the stretched out signal bit after interference before the dfe high pass filter we measure the intensity of the signal to configure dfe, basically it's dividing a time period by n + 1 for sampling. This is usually good at 4 taps, but certain mobos are more stable with fewer tabs. This occurs when the 4th tap is dipping into a mirrored interference high, thereby worsening the dfe by subtracting instead of adding. Generally the vendor is smart enough to choose that one correctly as AMI-IBIS configures it automatically. You can also pirate jesd79 5 to read more on how dfe is actually applied this is chapter 4.30. it's very nicely explained there too

InkyBusiness

With -6.26% to +4.21% on 1.65VDD mobo is apparently not confrom to jsed which allows a VDD/Q -3% to +6%. Combine that with the generic CA bus with fly by routing for odt on ddr5 and high l2l vdelta on this PCB the umc phy interconnect might already be an issue. So apparently they are selling real shite

InkyBusiness

I mean its only pulling 20w, how can you mess that up so much?! I guess I'll be soldering on some input and especially output filtering to see if I can get anywhere with reducing the transient response on pull from the memory towards the powerstage. I might take it to work in January to give it a few X-rays after hours. I am interested to see where the ground plane is and isn't and how they're connecting to it for pulldown. Ive got a few suspicions on them cheaping out on manufacturing cost

InkyBusiness

I hooked up my ocy to the output of the soc power stage and saw a utterly messy voltage lotta droop, no detectable llc (no overshoot I'd expect on bursts from prime) and a switching frequency of 200mhz. Something is seriously messed up with this mobo. I read up on the fets the Asus b650m-k uses a hilo config of https://www.digikey.de/de/products/detail/vishay-siliconix/SIRA14DP-T1-GE3/3309106 trench mosfets. Being trench mosfets they have atrocious parasitic capacitances. Looked a look at the soc vrm and found almost no ceramic / low er capacitances. Furthermore it seems that the vsoc is only supplied by one phase instead of two under Karhu. A combination of those factors are a good indicator of why the mobo overclocks like shite

InkyBusiness

Probably me setting something wrong, Asus disliking my RAM or CPU or a combination of those. Do also note that that 100€ Asus b650m-k should not bottleneck either, but the OC is abysmal aswell. After a few hours of fiddling around I just chalk it up to a software thing and move on. Interesting is that CL goes down to 8.3ns which speaks to good signal quality overall, but a higher MCLK just doesn't wanna stick. The later one looks like a CPU issue, but then why does it work in ASRock and gigabyte? In your eyes what's specifically different with Asus compared to ASRock for instance? Anything to pay attention to? Anyway I'll give the 9800x3d a shot in the cheap Asus, and see if it's a CPU compatibility issue of something along those lines. Should he interesting

Actually Hardcore Overclocking (Buildzoid)

the ITX isn't daisy chain. So IDK why it was so bad for you.

InkyBusiness

Ah I had a typo x670E-I. Can a modern mobo be daisy chain, even though they are on two dim slots?

Actually Hardcore Overclocking (Buildzoid)

the X670E-E is a daisy chain. B650E-E also

InkyBusiness

Luckily it's a 2 dim, no daisy chain :)

Actually Hardcore Overclocking (Buildzoid)

I think ASUS just had the worst daisy chain topology for 600 series.

InkyBusiness

Anyone got an idea why my milage on the cheap Asus is so bad compared to the 200-300€ segment? Is ASRock just the best mobo vendor looking at the aspects memory OC, software that you don't wanna strangle, and price? . Note: for gigabyte I am limited to 1.43V VDD, -IO, -Q so pushing CL is no option. Quite a bit of performance is left on on the road here... For everything else I start my OC at 1.35V and then push CL at starting at 1.46V VDDIO, 1.67V VDD and 1.64V VDDQ.

InkyBusiness

I went thru a few motherboards overclocking my Ryzen 7500F which loves VDDP 1.080V droop, with Patriot Viper 8200MT CL47. - ASRock X870 Pro RS 220€. Stable on 8000MT with rcd 47 rtp 12 rdwr 19, after fiddling with the bus configuration a bit. Does XMP 7800 out of the box. - ASRock X870 Pro RS 220€. Stable on 6400MT FCLK 2166 cl 28 rcd 35 rtp 12 rdwr 15. Same bus config as above. - Gigabyte X870 Aorus Elite 280€. Does not do 8000MT stable at any timing configuration at all. Errors out after 20min Karhu. High voltage mode is bricked in BIOS! Does 7800MT with rcd 47 rtp 12 rdwr 19. Basically the same as ASRock does on 8000MT. - Asus rog strix X670E-1 350€. 7600MT with same timings as the Gigabyte X870 Aorus Elite on 7800MT. needs higher VDDSOC for some reason. - Asus 650M-K 100€. Does 6200MT cl26 rcd 35 rtp12 rdwr 14. Rdwr 13 errors in Karhu after 3min. Posts but bsods immediately at 7400MT. I assumed the bus configuration the mobo chose is fucked. Spent a good 5h fiddling around with that fucker. No game... I got to post 7600MT eventually ofc unstable, even 6400MT just does not want to stick here. - ASRock 650M-HDV it's in the mail I am testing it next. I have higher hopes here because my 7500F loves ASRock and gigabyte. But gigabyte software is just such a buggy mess I prefer Asus with arguably worse performance over them. Fuckers cant even get RGB to work. I got a 9800X3D in mail gonna retest a few mobos when it arrives

Matthias Galvin

For what it's worth: H24M has reliably worked, but vendor binning varies. My Patriot 8200C38 H24M kit has worse tRP and tRAS values than my Team T-Create Expert H24M kit at 8000, and the Patriot kit didn't work on one of my AM5 CPUs (7700), where the Team worked on all of them at 8000 (7700, 7900X3D, 7950X3D).

Zach Ketrow

Yeah exactly I mean the kit you tested is likely about the same tbh so I'd test all the other more interesting stuff first.

Actually Hardcore Overclocking (Buildzoid)

yeah that wouldn't surprise me though the 7800C36 kit + toma is one of the combinatios I probably won't test any time soon

Zach Ketrow

I have the 2x16 G.skill 7800 C36 XMP gskill kit and the settings in your latest 8000 XMP adie video work on my 9800x3d and x670e carbon. I figure that would probably work on your x870 tomahawk and 9700x as well.