Discussion Guys, just had a bizzare experience with my 3900x: Tighter RAM timings = Stability Issues Solved

[leepox]

I have just received a new new Patriot Memory Viper Steel PVS416G440C9K paired with a 3900x and a B550 MSI Itx gaming motherboard. I did not want to do some crazy OC so played with timings at around 1.38v. My original timings on the right resulted in memtest86 errors so decided to loosen it a little bit more by reducing the TFaw and TWr as per the DRAM calculator (basically going for a safe preset rather than fast). And the impact it had on membench time was ridiculous which either means a. it's unstable at that looser timing (did memtest86 on this settings and no errors)OR b. changing the TFaw and TWr from 12 to 24 have a massive impact.
to summarise:

Left Pic: Looser TFaw and TWr no memtest errors, stable but huge membench performance degradation. and Right Pic: Tighter TFaw nad TWr but memtest errors, and was unstable (esp gaming).


It doesn't end there. I decided to go back to tighter timings and this time follow the DRAM calculator's fast settings for the sub-timings (especially tRFc) closely but kept the main timings to CL15 with trCDRD at 16 (TRas and TRc also manual values to reflect) because with CL14 main timings the system doesn't really boot at 1.38v and not really keen on putting more voltage/stress.
And guess what, these even tighter timings at the same voltage actually produces no errors (2x 4 passes memtest, karhu, etc. and passes stability testing. I'm getting some really impressive benchmark results... Is it a case that the CPU/RAM prefer the tighter timings??? It just defies logic! I mean... WTF??!!


above: Stable system with tighter timings at the same voltage. Although increased VDDP and VDDG by 0.1v each as they were reading 0.895 and 0.945 respectively to match the recommended 0.9 and 0.95 minimum. SOC is 1.1 for all cases.

 

[A///]

According to Google that particular model of RAM is Samsung B-Die. Gets a lot of love on various tech discords like the overclock one. Are you going to push the RAM further and see what it performs like? If you don't mind me asking, how much did it set you back? There's a healthy jump in price from B-die 3200 to 3600 which is... to those of us wanting to see more normal RAM pricing. Hopefully if that rumor article turns out to be true about surplus...

 

[Shmee]

That is a pretty nice OC. I settled on a 3600 MHz OC with my Gskill at around 1.45 V.

 

[leepox]

This is a 4400Hz rated B die you're correct. I am tempted to push further but it's a lost cause at this stage as I am severely IF limited. The reason I bought it was to play around at 1800 IF. But after some testing it was clear that I did not win the bin lottery for my 3900x. It runs at 3733 but with signs of instability. So I'm just going to stick with this and wait for Zen 3, hopefully win the die lottery then especially with the rumoured 2000Mhz IF clocks.

 

[A///]

That is a pretty nice OC. I settled on a 3600 MHz OC with my Gskill at around 1.45 V.

Wasn't the general consensus on clocking that 3600 is a sweet spot, sometimes 3733 depending on the application and anything past that a negative ROI?

 

[A///]

This is a 4400Hz rated B die you're correct. I am tempted to push further but it's a lost cause at this stage as I am severely IF limited. The reason I bought it was to play around at 1800 IF. But after some testing it was clear that I did not win the bin lottery for my 3900x. It runs at 3733 but with signs of instability. So I'm just going to stick with this and wait for Zen 3, hopefully win the die lottery then especially with the rumoured 2000Mhz IF clocks.

This is the first time I'm hearing of a 2000 Mhz IF clock. Who said this? In the past 7 months I've built more Zen2 systems for family and friends than I'd care to count and I'd test the RAM clocks on each build. At 3600 you're pretty good to go unless you get a golden goose and get a 3733 going with very minimal improvements. The sweet spot seems to be going from good quality 3200 and OC'ing it to 3600. I know a few people who bought into 3000 and go to stable 3600. I've gotten some hush rumors of future Zen parts including 3 but I choose to not place any faith in them because I would prefer being surprised like a child on Christmas morning rather than place AMD's product on a pedestal only to accept crushing disappointment later on. Depending on Zen 3's success, it may be the last Ryzen generation before we see a sizeable increase in price as AMD aligns themselves to be the premium CPU maker.

 

[leepox]

It's more of a rumour mill based on the Renoir rumour mill... so a rumour of a rumour but nothing else :

AMD Ryzen 4000 test chips already look to be as fast as Zen 2

The rumour mills go into overdrive as more details for AMD's 'Renoir' APUs and 'Vermeer' CPUs emerge

www.pcgamer.com

https://twitter.com/x/status/1276050313472962560

Basically everyone is hoping this translates to Zen 3.

But you're right, it's diminishing returns from here onwards and we have to manage expectations. I hope AMD stays true to keeping enthusiasts not bankrupt. It's crazy how much NVDIA and Intel are pricing their products and still have demand. Life without AMD doing well would be tragic for us consumers.

 

[A///]

Oh yeah. I used to use Chip Hell a lot but my translate extention stopped working, but I was using it on Opera GX to keep it away from everything else I've got. I'm expecting a lot of under the cover changes for AM5, fingers crossed of course. Zen3 will be my chosen generation, and then I'll wait for Zen 5 under AM5 before upgrading. 2-3 years depending on how long between releases. The platform should be stable by then if any issues popup next year under AM5's launch.

 

[leepox]

same, thinking of ugrading to Zen 3 - the last of the AM4 socket next year when AM5 gets officially releases and Zen 3 prices hopefully plummet as a result. Then same as with you, I'll stick to my Zen 3/AM4 until the end of times.. more realistically, skip a generation. XD

 

[Gideon]

It's more of a rumour mill based on the Renoir rumour mill... so a rumour of a rumour but nothing else :

Well you're right that anything regarding zen 3 is still a rumor, but higher Renoir FCLK is confirmed, as Pro versions are out and reviews have been done.

In the Guru3d Renoir Pro review Usmus1 got even 4533CL16 stable (on a Corsair Vengeance LPX 4133C19 B-die kit).
On a 100$ budget E-die he got 4333CL18:

 

[leepox]

Those are some ridiculous IFs. That 4533 look like heaven... I can only imagine what the latency and bandwidth results look like if Zen 3 reaches similar numbers!

 

[A///]

Took me a while to notice the die variety and the i vs d gpu. Impressive results. God, I'm going to love Zen 3.

*plays Lady in Red by Chris De Burgh*

 

[inf64]

Zen3 will be brutal for bursty workloads, not even counting the IF improvements we see in Renoir. It will be another year+ where AMD will do a rampage roflstomping action, I hope prices will not go thru the roof.

 

[Gideon]

Just to reign in expectations a little bit: Unlike Renoir, Zen 3 is using chiplets, therefore latency will be a bit worse at the same FCLK (couple of nanoseconds at least). As the fabric is spanning multiple dies it will also use cosiderablt more power and might still hit a wall earlier than Renoir.

Even keeping all of that in mind, it should still be way better than Zen2.

EDIT: I'm personally hoping for 55ns latency in the 4200-4400 mhz DDR4 range. Wouldn't mind at all of they do better, but this (coupled with the unified L3) is what they need to really trade blows with Intel's best @ similar timings.

Ina all likelyhood, they'll still lose some benches to the 40ns Core i9s, but they should finally win others and hopefully, some of them quite comfortably. I mean, get the results they get with CS:GO now (where Zen 2 rules the roost) in multiple more relevant games

 

[A///]

I've heard some really interesting figures for Zen3 cache latency. Nothing worth repeating publicly or privately. And only because what's the point?