Overclocking RAM, Gear 1 vs Gear 2

[ir_cow]

Anything over 6000 will automatically have the motherboard set UCLK DIV1 mode to UCLK = MCLK/2.

It seems to vary per motherboard series and vendor. For example the MSI X870E Carbon does 1:1 on auto up to 6400.

I've also seen 2:1 at 6000 for the A620.

 

[ShrimpBrime]

It seems to vary per motherboard series and vendor. For example the MSI X870E Carbon does 1:1 on auto up to 6400.

I've also seen 2:1 at 6000 for the A620.

What's the max Fabric clock?

 

[freeagent]

What's the max Fabric clock?

I am running mine at 2200 right now

 

[ir_cow]

What's the max Fabric clock?

Dont know. It was 2000 Mhz for the default auto for every X670 B650 MB. I've always had bad luck with CPUs not being stable above 2066 for y-cruncher. Must not be pushing the voltage hard enough

Now I'm seeing a optional setting in the MSI X870 that will match it to the memory "optimally". Just enabled or Disabled as choices.

 

[ShrimpBrime]

Dont know. It was 2000 Mhz for the default auto for every X670 B650 MB. I've always had bad luck with CPUs not being stable above 2066 for y-cruncher. Must not be pushing the voltage hard enough

Now I'm seeing a optional setting in the MSI X870 that will match it to the memory "optimally". Just enabled or Disabled as choices.

Understandably, why push voltage for short gains. So do you bclk for that 66mhz? Is that a thing for most/some people. like 101, 102 nothing crazy?

 

[ir_cow]

I leave the BLCK alone. It's been a while, but 102 was the limit for my 7950X. Haven't tried it with the 9 series I have.

 

[ShrimpBrime]

I leave the BLCK alone. It's been a while, but 102 was the limit for my 7950X. Haven't tried it with the 9 series I have.

Ah ok. Cause I'm under the impression that you can't lock 100mhz bclk so it's always like 99.x mhz. Or can you disable spread spectrum?

 

[ir_cow]

You can disable spread spectrum. Smarter to leave it on unless you are setting a world record and need that last 1mhz.

The way I understood it is that the 1-3mhz range was to prevent harmonics / resonate with other electronics. So spread is a good thing.

 

[JustBenching]

You could have said this in general, not just for X3D chips (AMD)
I used to overclock RAM with Intel and it's the same, you just earn better benchmarks scores with no real benefits.
You could be happy to have a temporary world record in RAM OC. You could sell the RAM kit for better than you paid, that's it.

My 12900k gets 30+% performance from ram tuning. EG. in Ratched and clank. Not all games scale like that, but still you will see a 20+%. Of course assuming you are not gpu bound

I also don't run a really high Trfc either. On auto, it's around 6650, I typically don't go over 11,000. ummm well..... I should really say, this depends on what I'm doing and the benchmark specifically. Many all respond quite differently, and also a persuasion on stability with certain clocks and timings.

You mean TREFI, set it to 65535 and let it be. It doesn't affect stability. Well it does affect temperatures which might affect stability if dram temps get too high but that's about it. Trefi and TRFC are the best ways to reduce latency

 

[Chrispy_]

You could have said this in general, not just for X3D chips (AMD)
I used to overclock RAM with Intel and it's the same, you just earn better benchmarks scores with no real benefits.

I think in the early Zen2/Zen3 days, getting the FCLK/MCLK/UCLK up to 1800-1900MHz actually showed significant gains - in the order of 5-10% average fps and up to 25% on the 1% lows. Then the 5800X3D came along and proved that even with slow DDR4, gaming performance simply didn't change regardless of what speed your ran your FCLK/MCLK/UCLK as long as you were using at least DDR4-3000 or more.

Zen4 and Zen5 have clearly had some architectural tweaks to Infinity Fabric that means it's less sensitive to latency and more tolerant outside of 1:1 FCLK/UCLK ratios. I'm sure you'll always be able to find some niche use case where more bandwidth really makes a difference, but realistically even 5600 seems to be enough bandwidth and the more important things to focus on are to get your access times down and to get good ratios of FCLK and UCLK. The easiest "optimum" ratio for Zen5 seems to be 2:3 so a 2GHz infinity fabric driving a 3GHz memory controller outperforms most other ratios for a wide spread in either direction.

Given that the FCKL should ideally run as fast as it can, within stability, that typically means somewhere around 2GHz, which lends itself nicely to DDR5-6000 at the optimum 2:3 ratio. Yes, 6200 and even 6400 will give you marginal gains from those extra 66MHz and 133MHz extra fabric clocks, but only on the strict conditions that you don't upset the 2:3 ratio and don't have to increase your absolute access time via loosened CL memory timings. I'd love to see a hypothetical 1:1:1 with a CPU capable of running 2400FCLK with very tight DDR5-4800. I have no idea how it would perform, but my hunch is that in games it would likely be better than the sweet spot of 3:2:3 using DDR5-6000 CL30.

 

[ShrimpBrime]

My 12900k gets 30+% performance from ram tuning. EG. in Ratched and clank. Not all games scale like that, but still you will see a 20+%. Of course assuming you are not gpu bound


You mean TREFI, set it to 65535 and let it be. It doesn't affect stability. Well it does affect temperatures which might affect stability if dram temps get too high but that's about it. Trefi and TRFC are the best ways to reduce latency

For me, seeing the best latency gains is cache overclocking.

Yes, meant Trefi. Thank you.

And it does effect stability at some point, I suppose temps could get difficult sometimes. Dont always have a fan right on them. Good point.

 

[freeagent]

My 5800X3D loves fast IF with the mem clocks to match, it also loves running a 2R config with C14 timings. But so does my 5600X and 5900X..

My 9900X doesn't seem to care too much.. 1:1, 1:2 it doesn't seem to matter latency wise as long as you run everything as fast as you can, as tight as you can.

 

[AleXXX666]

It was mainly because I was locked into a dead-end socket, with a DDR4 motherboard. If I'd had a DDR5 board I would have just picked up a 14700k because it has a larger level 3 cache than the 12700k and more performance in general. Would have been an energy hog but beastly for productivity as well as competitive in gaming.

And for what it's worth, the new setup does seem to have smoothed out Helldivers 2 somewhat.


Could it be that the 68 TRAS is more stable for him because he left his ram at CL36?

I'm currently on 14600K with DDR4, want to change mobo to DDR5 and.. STAY with it

 

[AVATARAT]

What's the max Fabric clock?

With the 9000 series, there are some unicorns that can make 2233 Fabric stable.
My chip can get into bios at 2233 but that's it but since the last few generations most chips can do 2200, some are bad and are 1-2 steps lower.

 

[Chrispy_]

What's the max Fabric clock?

I have a 9700X that's unhappy at 2133. I haven't pushed my personal 9900X but it is quite happy at 2133 running DDR5-6400 in a 3:2:3 config.

Dont know. It was 2000 Mhz for the default auto for every X670 B650 MB. I've always had bad luck with CPUs not being stable above 2066 for y-cruncher.

Is y-cruncher the best stress test for fabric clock? I assume 2000 is the default in AGESA code sent to board manufacturers, meaning that AMD are 100% confident it's stable on every single product they sell.

Knowing what test is going to display instability first is useful, because I'm making the assumption that my DDR5-6400/9900X/B650 Tomahawk are happy at 2133, but haven't tried y-cruncher - just linpack and memtest86...

 

[ir_cow]

Larger Y-cruncher sizes like 2.5b and 10b is a really quick way to find out if something is unstable, but not enough to crash the computer outright.

For example I might be able to run games, but immediately fail y-cruncher. If I ran a OCCT (CPu+men) long enough I would get a error. Y-cruncher intensive uses CPu and memory making it ideal for a quick indication that your on the right path or not for that OC.

 

[Chrispy_]

Larger Y-cruncher sizes like 2.5b and 10b is a really quick way to find out if something is unstable, but not enough to crash the computer outright.

For example I might be able to run games, but immediately fail y-cruncher. If I ran a OCCT (CPu+men) long enough I would get a error. Y-cruncher intensive uses CPu and memory making it ideal for a quick indication that your on the right path or not for that OC.

Wow, okay - my 6400 (2133FCLK) 9900X isn't stable at stock voltages.

That's the problem with memory stability. It can be borderline and go days/weeks without a single hiccup. I guess DDR5's single-bit ECC means you could be "stable" whilst actually losing performance as ECC hides the evidence.

 

[freeagent]

Chances are you need VDDG IOD. Mine will do 2133 at stock, or expo or whatever.. but VDDG IOD is what let me run 2200.

 

[Chrispy_]

Chances are you need VDDG IOD. Mine will do 2133 at stock, or expo or whatever.. but VDDG IOD is what let me run 2200.

Probably, but this is a temporary CPU and this Kingston DDR5-6400-CL32 kit runs happily at 6000-CL28.

I was more posting to say that 2133FCLK isn't always stable, thinking one of the two AM5 chips I've tried maxing the fabric clock on couldn't handle it at stock voltages. It would appear that actually both of them are unstable at 2133 once I introduce y-cruncher.

I'm sure they'll make it and maybe even go higher with voltage bumps in various places, but talking about stock voltages here - otherwise where do you draw the line?