Risks of unstable RAM/fclk overclock?

[seward]

I've got a pretty high overclock on mid-tier RAM: 4200mhz, with mostly same timings as 3600mhz, at same volts (1.35v). I'm using 2x8GB of Corsair Vengeance LPX DDR4 3600, with a Ryzen 5 5600G cpu, on an Asus Prime B550-PLUS motherboard. I'm pushing RAM and fclk settings because I rely on the chip's igpu for gaming (I play old games); fclk = igpu memory clock, and 2:1 RAM/fclk ratio is preferred.

This isn't high-performance RAM, though, so I was surprised that I booted with a 4000mhz/2000mhz overclock, and even more surprised when I booted with a 4200mhz/2100mhz overclock. Auto-voltage remains 1.35v for RAM, same as 3600mhz, and 1.25v SOC, which is also same as. It seems like my equipment is being nice to me, but I don't want corrupted files. I haven't stress-tested these settings yet. Is it a mistake to rely on being able to boot with a high RAM/fclk overclock, as a sign of viability? Is my data coming apart?

 

[NattyKathy]

Definitely do stability testing. Booting does not mean it's stable and yes using an unstable ram/IF OC for an extended period of time can cause data corruption in addition to weird system behavior and random crashes.

 

[kirbyrj]

IIRC, the Ryzen FCLK was only good up to ~3800Mhz, so I would think you're losing performance by changing it to a 2:1 ratio just to increase RAM speed.

 

[seward]

IIRC, the Ryzen FCLK was only good up to ~3800Mhz, so I would think you're losing performance by changing it to a 2:1 ratio just to increase RAM speed.

I've been wondering about keeping the 2:1, looking for more info about decoupling and performance. I think the Corsair Vengeance RAM is overachieving here, which is okay (I was able to boot at 4400/2200, but it was a strain). I may decouple and raise fclk and run some benchmarks.

 

[kirbyrj]

I've been wondering about keeping the 2:1, looking for more info about decoupling and performance. I think the Corsair Vengeance RAM is overachieving here, which is okay (I was able to boot at 4400/2200, but it was a strain). I may decouple and raise fclk and run some benchmarks.

Definitely. I'm talking more from a "best practice" perspective, but your individual needs might be better met by decoupling. Probably worth doing some benchmarks to see.

 

[riev90]

I've been wondering about keeping the 2:1, looking for more info about decoupling and performance. I think the Corsair Vengeance RAM is overachieving here, which is okay (I was able to boot at 4400/2200, but it was a strain). I may decouple and raise fclk and run some benchmarks.

1. You are actually running 1:1 as the ram is running at 2000 not 4000 (ddr).
And for ryzen 5600g (cezanne) it's a very common thing to have the IMC limit higher compared to Vermeer (2000, max 2033).
I have tested 2 samples of ryzen 5600g and both can run ram fclk : mclk 1:1 at 2266, resulting dram frequency of 4533mhz.

2. Don't do decouple, for AM4, decoupled of fclk will give worse performance on all of the apps/games.
better check your IMC limit, for example you said that it can boot dram/flck/mclk at 4400/2200/2200. then I would pursue stability on that settings.

3. Do memtest on it, you can select the appropriate apps to test.
I usually use this (TM5, use the absolute config):
https://mega.nz/file/vLhxBahB#WwJIpN3mQOaq_XsJUboSIcaMg3RlVBWvFnVspgJpcLY

4. This is one of my previous 5600g result from overclocking (dram/flck/mclk at 4266/2133/2133):

It can do 4533 but I prefer to SoC Voltage to not exceed 1.3v for daily usage.
If you have the same memory chip (Hynix C-Die) same with the screenshot I posted above, you may try setting the memory timings and sub-timings referencing to mine.

 

[seward]

1. You are actually running 1:1 as the ram is running at 2000 not 4000 (ddr).
And for ryzen 5600g (cezanne) it's a very common thing to have the IMC limit higher compared to Vermeer (2000, max 2033).
I have tested 2 samples of ryzen 5600g and both can run ram fclk : mclk 1:1 at 2266, resulting dram frequency of 4533mhz.

2. Don't do decouple, for AM4, decoupled of fclk will give worse performance on all of the apps/games.
better check your IMC limit, for example you said that it can boot dram/flck/mclk at 4400/2200/2200. then I would pursue stability on that settings.

3. Do memtest on it, you can select the appropriate apps to test.
I usually use this (TM5, use the absolute config):
https://mega.nz/file/vLhxBahB#WwJIpN3mQOaq_XsJUboSIcaMg3RlVBWvFnVspgJpcLY

4. This is one of my previous 5600g result from overclocking (dram/flck/mclk at 4266/2133/2133):
View attachment 585659

It can do 4533 but I prefer to SoC Voltage to not exceed 1.3v for daily usage.
If you have the same memory chip (Hynix C-Die) same with the screenshot I posted above, you may try setting the memory timings and sub-timings referencing to mine.

Thank you for reminding me about ratio, I keep forgetting about DDR. Sorry if I confused anyone. I did more reading about my new build (general AMD stuff + Ryzen-specific info), and went ahead and swapped the Corsair RAM out for 2x16GB G Skill Ripjaws V 3600 CL16, which is dual rank (if CPUZ can be trusted). When I bought the Corsair CL18, I didn't know how important RAM is with these Ryzen chips, particularly the apus and their igpus. I figure that if I can get to 4000/2000 with CL18 Corsair Vengeance, I may be able to get some tighter timings, and perhaps better latency, with CL16 G Skill Ripjaws V. I've never loved overclocking RAM, I'm too impatient, but I can't ignore how much memory affects Ryzen performance, including igpu. As far as increasing memory from 2x8GB to 2x16GB: when I only had 2x8GB, I had a little bit of system instability when I allocated 2GB of RAM to igpu. With 2x16GB of RAM, I've allocated 4GB (UMA Frame Buffer in BIOS), and there are no hiccups. The G Skill is set to DOCP, and the setup has passed 3 error-free hours of OCCT's memory test. I've got memtest on a USB, haven't run it yet. I haven't started overclocking the new RAM yet, that'll be work.

I should mention again that I'm currently playing games like Supreme Commander 2 and Factorio, on an old-ish 27-inch 1080p IPS monitor. Gameplay and image quality in these low-graphic-demand games really seem to improve with each little tweak and overclock that I apply to igpu and RAM. I don't think the igpu will run things as well at 1440p, though. That's gonna take a discrete gpu, and a new monitor, and it won't be cheap. So I'm staying at 1080p for now, and I'm trying to optimize my 1080p experience. Because of the nature of these AMD apus, RAM overclocking is a big part of that.

 

[DogsofJune]

I don't think the igpu will run things as well at 1440p, though. That's gonna take a discrete gpu, and a new monitor, and it won't be cheap. So I'm staying at 1080p for now, and I'm trying to optimize my 1080p experience. Because of the nature of these AMD apus, RAM overclocking is a big part of that.

I run DDR4 4000 sticks with my 5700G. The igpu does like the faster ram, and absolutely would struggle at 1440 for games. Even 1080 can be a struggle depending on the game and eye candy settings.

 

[riev90]

Swapped the Corsair RAM out for 2x16GB G Skill Ripjaws V 3600 CL16, which is dual rank (if CPUZ can be trusted).

Can you specify the timings of your new GSkill Ripjawz? 3600 16-16-16-16? if so then you've got Samsung B-Die which is the best ddr4 dram chip.

From my screenshot above, the dram chip is Hynix C-Die.
But I also have the result for Samsung B-Die using the same procie & combo:

1. This one using GSkill OC World Cup 2016 3600c16, with optimized timings 18-19-8-19-39 trfc 348, vdimm 1.44v only, SoC set at 1.28v. I didn't use direct fan for ram airflow because at 1.44v, the max temp dram reached was 44.8c.

2. This one using GSkill Royal 3600c16, with optimized timings 16-16-8-16-32 trfc 288, vdimm 1.54v , SoC set at 1.28v. I used direct fan for ram airflow to maintain dram temp under 45c for daily use.

I've never loved overclocking RAM, I'm too impatient, but I can't ignore how much memory affects Ryzen performance, including igpu. The G Skill is set to DOCP, and the setup has passed 3 error-free hours of OCCT's memory test.

Oh you will absolutely love it.
If you find it hard to manually tuning the timings, then you can look at the sample of timings from other user, like the one I posted above.
For APU, going from freq 3600 to 4266 I showed above, the performance uplift ~20%, and it can increase to ~30% once you overclock the APU Core to, for example 2200mhz (from 1900mhz) at 1.2v.

 

[seward]

Can you specify the timings of your new GSkill Ripjawz? 3600 16-16-16-16? if so then you've got Samsung B-Die which is the best ddr4 dram chip.

From my screenshot above, the dram chip is Hynix C-Die.
But I also have the result for Samsung B-Die using the same procie & combo:

1. This one using GSkill OC World Cup 2016 3600c16, with optimized timings 18-19-8-19-39 trfc 348, vdimm 1.44v only, SoC set at 1.28v. I didn't use direct fan for ram airflow because at 1.44v, the max temp dram reached was 44.8c.
View attachment 586286

2. This one using GSkill Royal 3600c16, with optimized timings 16-16-8-16-32 trfc 288, vdimm 1.54v , SoC set at 1.28v. I used direct fan for ram airflow to maintain dram temp under 45c for daily use.
View attachment 586288

Oh you will absolutely love it.
If you find it hard to manually tuning the timings, then you can look at the sample of timings from other user, like the one I posted above.
For APU, going from freq 3600 to 4266 I showed above, the performance uplift ~20%, and it can increase to ~30% once you overclock the APU Core to, for example 2200mhz (from 1900mhz) at 1.2v.

I wish I had those timings! My Ripjaws are 16-19-19-39. Looser than those b-dies, but tighter than the Corsair. Thanks very much for timings btw. I'm still undecided about whether to tighten 3600mhz timings, or work to get 2000 FCLK, like I did with the CL18 Corsair, but with better timings.

The Corsair CL18 would boot into Windows at 4200mhz/2100FCLK, at 3600mhz DOCP timings and volts, although it wasn't stable. I was pretty stable at 4000mhz/2000FCLK (3600 timings/volts), although I didn't run very long memory stress tests on it, like I now know I should. The G Skill CL16 runs great at 3600mhz DOCP, and passes multi-hour memory stress tests, but it won't boot into Windows at 4000/2000 with 3600mhz DOCP settings. If I want to do that, it's going to be a little more work (as expected), but if I can get the G Skill to run 4000/2000 with timings that are tighter than the Corsair, I'd be pretty happy. If 3600mhz really is the sweet spot for these AMD chips, and if 4000/2000 takes a lot of work (and volts) and doesn't scale up performance that much, then I'd be okay with optimizing 3600. Your results make me think that I should go for 4000/2000, though - the performance improvements are there (and I bought this RAM to over-volt it, for god's sake). Happy to hear any experiences with either approach.

Edit: I always overcomplicate things...I basically want to overclock the G Skill to 4000mhz, in order to support a 4000/2000 RAM/FCLK setting. 16-19-19-39 won't boot at 4000mhz/1.35 volts. I'm prepared to loosen timings, not sure where to start with volts. I figure the G Skill should be able to at least run at 4000mhz with better timings than the Corsair, though that might mean CL18.

 

[riev90]

Edit: I always overcomplicate things...I basically want to overclock the G Skill to 4000mhz, in order to support a 4000/2000 RAM/FCLK setting. 16-19-19-39 won't boot at 4000mhz/1.35 volts. I'm prepared to loosen timings, not sure where to start with volts. I figure the G Skill should be able to at least run at 4000mhz with better timings than the Corsair, though that might mean CL18.

Your GSkill Ripjaws with 3600 16-19-19-19 probably use Hynix C/D Die.
At 4000 mhz and dual rank, I would go with 18-22-22-22 at 1.35v.
For 16-19-19-19 or 16-17-19-21 you may need 1.45v at 4000mhz.

Like I said previously, better find the limit of your 5600g IMC (4200/2100???) and then start from there.
I can testify that even running 4200 18-22-22-22 for APU performance, it will be better than running 3600 14-14-14-14 with super-tuned sub-timings and tertiary.

 

[dasa]

Could also be Micron e die which can often hit ~4800-5200+ ~1.4-1.45v and while you will lose CPU performance by running outside 1:1 I believe you will still gain IGP performance which is what matters most for you so it may be worth a test with it decoupled and seeing just how far you can go.
Test both with something like 3dmark to check performance.