DDR5 on Z790 & X670|B650?

[Gouhan]

DDR5 is really starting to show its stuff with the latest chipsets. Even though AMD has to catch up being 1 generation behind. Where the X670/B650 Chipset is starting isn't a bad place at all. I've been fortunate to run 6400 solidly and of course post and boot into windows 6600. Definitely not a bad effort for the 1st go around. I expect AMD to only get better here fast (or at least that's the hopium speaking/writing).

As for Z790 though, the improvements to these boards and matching CPUs are massive. To the point where what was once challenging to do on 4 DIMM boards if not impossible (DDR5 7000+) is pretty much pedestrian for Z790.
Fully expect to see DDR5 8000 kits before EOY or latest Q1 2023 (assuming the world doesn't explode).

Really impressed with the latest boards and their ability to push the memory.

 

[dainluke]

There's a YT reviewer who got access to the TG A-die kit before it went out of stock, and Igor also had some testing on Zen 4. I've done some looks at Zen and RPL boards and it seems that SKU for SKU memory topologies are a copy and paste. It seems board pricing has increased across the board due to overall higher PCB layer counts and more QC that they learned was necessary after a lot of poor showings by Z690s.

Anyhow, the X670E Hero seems to be more than capable of 7000+ when the memory controller ratio is switched to 1:2 similar to Intel. It seems that 3200MHz is a realistic maximum for most Zen 4 chips and 3300MHz is potentially possible on some samples. In that regard, it's more a question of how AMD develop the IMC for Zen 5. Board-wise the topologies and PCB adjustments have been locked in for signal integrity.

According to some OCN users and results I've seen, it seems that AMD have adopted a different design to Intel's interleaving IMCs and that the ADL and RPL IMCs are just far more efficient. I think as it stands though, all X670, B650 and Z790 boards with 8 or more PCB layers should be capable of 7000-7466 in 2T (1T if AMD), with good H16M and more readily with H16A. It seems H16A itself has the effect of not requiring as much of a signal boost and thereby not really as at risk of data corruption as with H16M and S16B when at higher data rates. I've also seen some 2x8 A-die kits making me question if SK Hynix didn't perhaps make X8 H8A ICs as well this time because perhaps they identified the need for it right now. CJR came in different configurations as well so it would be quite possible.

Some things I've heard though are that Hynix A-die can't really do 1T and that timings don't righten up on it to the same degree that they do on Hynix M-die at low data rates. I'm wondering how much of Z790 is a topology revamp and how much is just validations of new A-die kits. It would be interesting to see M-die tested on Z790 to see if there actually has been an overall re-design of each vendor's topologies.

 

[Gouhan]

On AMD's side the IF bandwidth becomes the limiting factor in raw bandwidth terms, something that simply isn't an issue or less so for Intel. especially with a rumored boosted MC clock. I tried switching ratios and both boards (AM5) would have none of it. And it is telling because you'll see that all ROG boards for instance, don't have a DRAM profile above 6400.

It simply isn't worth it, if you can run it at all to use such a ratio. Most certainly 1T isn't happening for me anytime soon. Dropping frequency to the low 5000/5200 may do it though I've yet to attempt this. ADL's IMC wasn't/isn't as strong as RPL-S it seems as well. DDR4 boards for Z790, even cheesy cheddar melt boards like the 790M PG Lightning that are 6-layer offerings do DDR4 5333+ where this was not easy at all with Z690 DDR4.

 

[dainluke]

On AMD's side the IF bandwidth becomes the limiting factor in raw bandwidth terms, something that simply isn't an issue or less so for Intel. especially with a rumored boosted MC clock. I tried switching ratios and both boards (AM5) would have none of it. And it is telling because you'll see that all ROG boards for instance, don't have a DRAM profile above 6400.

It simply isn't worth it, if you can run it at all to use such a ratio. Most certainly 1T isn't happening for me anytime soon. Dropping frequency to the low 5000/5200 may do it though I've yet to attempt this. ADL's IMC wasn't/isn't as strong as RPL-S it seems as well. DDR4 boards for Z790, even cheesy cheddar melt boards like the 790M PG Lightning that are 6-layer offerings do DDR4 5333+ where this was not easy at all with Z690 DDR4.

The IF thing does actually make sense.

This is the video I saw it done in. It's always a possibility that he faked results, but Igor was also able to do 7000 with the change of the IMC ratio. I'm pretty sure the board spec pages are referencing the conventional conditions and not 1:2, but hey, it could be possible. I just don't know what the reason for the limitation would be. 7000+ should theoretically be entirely possible on AM5 boards bar the 3300MHz IMC clock limit.

When you say 5333+, I assume you mean Gear 2? Funnily enough, I have heard of Z690 boards struggling to do 5333 but I'm unsure if this is to do with the IMC given that 5333 in G2 is only 1333.67MHz on the IMC. I'd imagine at 5333 one would need to jack up TX VDDQ kinda substantially, but SA would be in the realms of 0.9V without issue. I'm not entirely sure if TX VDDQ is related to the RAM voltage or the data rate difficulty of the IC.

Also, it seems 1T on Intel is a board limitation. Not sure if it's primarily layer-dependent, but the 2-DIMM boards do a much better job of it. I almost have 7000-1T stable but I'm not sure what's preventing it from being rock solid. ADL RAM stuff is verifiably weird. The behavior seems to be very much related to memory trace layout. Funnily enough, ASUS redesigning their 2022 Z690 boards resulted in substantially higher memory support across the board. Even Extremes from M0UAY1 were doing 7000 with relative ease, and A-die testing on the Z690 Hero was reporting 7000 results without issue as well.

I'd actually be glad if it's primarily A-die responsible for overall higher data rates as there are a lot of 12th-gen IMCs and Z690 boards that should be able to do at least 7000. I've also seen enough results of 7466-7800 to deduce that a fair portion of 12900Ks should be good for 8000. It seems I'll be moving to RPL anyway so I won't be able to test this on my current 12900K, but I expect DDR5-8000-8400 to be doable on the Z690 Dark with A-die and a 13900K depending on how high the IMC can scale. for 8400 you'd need a 2100MHz IMC already for G2, which in itself would be a tough ask.

 

[Oj0]

And it is telling because you'll see that all ROG boards for instance, don't have a DRAM profile above 6400.

Depending on IC, profiles may only go as high as 5400. NO IC has profiles for 6600 or above though.

That said, Asus REALLY screwed up (IMO) by making the profiles harder to find than ever, with absolutely nothing in the manual telling you why they’re missing.