Ryzen APUs (5600G), memory speed and fabric clock, what to aim for?

[Mark Kaine]

3 hours ago, ShrimpBrime said:

Most people have issues just setting XMP on 3200mhz memory kits. 

To be fair he has the infamous samsung e-dies(?) though…

 

 

 

 

2133 OEM to 5000! Oof. : D

 

 

[Guest]

1 hour ago, Mark Kaine said:

To be fair he has the infamous samsung e-dies(?) though…

 

2133 OEM to 5000! Oof. : D

Ain't that something! He does some impressive videos. I like Buildzoid  

[Mark Kaine]

18 minutes ago, ShrimpBrime said:

Ain't that something! He does some impressive videos. I like Buildzoid  

yeah, its the only video i watched but i think i definitely learned something and he's pretty laid back … Im most impressed by how fast his BIOS is compared to mine, which seems mostly to run at negative fps… lol.

[Alvin853]

9 hours ago, ShrimpBrime said:

sometimes an odd cas latency does well too. 

This MSI MAG B450M Mortar Max that I'm currently using with the 5600G and the Kingston RAM doesn't accept odd cas latency values at all. I can input 17, 19, 21 etc, but it'll just apply the next even value, i.e. 18, 20, 22.

 

My main system that has the Corsair RAM is on a MSI MPG B550I Gaming Edge, which not only has a newer chipset, but also is higher on the MSI tier list, and it only has 2 DIMM slots which is commonly regarded as being better for RAM overclocking. But on the other hand, the 3700X that is running on that board maxes out at about 1867MHz on the FCLK, so I'd lose the 1:1 for now. Might upgrade to the 5800X3D later this year which would make higher memory speed more desirable.

[freeagent]

15 minutes ago, Alvin853 said:

10 hours ago, ShrimpBrime said:

too. 

This MSI MAG B450M Mortar Max that I'm currently using with the 5600G and the Kingston RAM doesn't accept odd cas latency values at all. I can input 17, 19, 21 etc, but it'll just apply the next even value, i.e. 18, 20, 22.

That’s because you have to turn gear down mode off, you should set your command rate to 2T when you do this, as 1T might be too tight.

[Alvin853]

2 hours ago, freeagent said:

That’s because you have to turn gear down mode off, you should set your command rate to 2T when you do this, as 1T might be too tight.

Wow never made the connection that odd CL timings are related to command rate. Thanks for the info

 

I now have the primary timings dialed in (up to tRAS, everything else is auto):

As expected for E-dies, the tRCDRD needs to be very high, any lower and it'll throw lots of errors. The other ones I just kept lowering until I either got errors within the first minute of memtest86+ or no POST, then backed up by 2, running extended stress test now to see if everything is stable, then I might go and try to lower the VDIMM a little.

 

Meanwhile my BIOS went full YOLO and is now pumping 1.3V into the SoC on "Auto" settings, really concerned for my 5600G. That's the one component of this system I would rather not want to damage, everything else is just parts I had laying around anyway that I can replace if needed. 

 

Next weekend I'll try to optimize FCLK and UCLK. I'm not planning to lower the memory speed if I can't get a 1:1 ratio, as I said before I'm optimizing for iGPU-performance, and preliminary testing has shown that iGPU performance greatly benefits from higher memory clock. OCing the iGPU itself will be next after that, haven't even looked into how that works. Can I still use MSI Afterburner, or is iGPU overclocking done in the BIOS too?

[Alvin853]

5 hours ago, Alvin853 said:

running extended stress test now to see if everything is stable

Well the result is disappointing, but not terrible. I ran 3 cycles of extreme1@anta777 on TM5 and ended up with a total of 20 errors, but no crashes, freezes, or instabilities otherwise. At least now I have a reference and I can make small adjustments and see if they make the situation better, worse or don't change anything.

 

I really don't want to go through the whole process once more, so I'll probably try a higher value for tRCDRD, and if that doesn't prevent errors, I'll drop down to 4533MHz, as that frequency seemed to be incredibly stable, that might allow me to tighten up the timings even more or drop the VDIMM by a fair amount. I don't think there are any other settings I could tweak to gain more stability, besides dropping to 2T command rate, but that's a big performance hit. 

 

 

Update: same result with tRCDRD at 28, so clearly this wasn't the issue. Went down to 4533, now my vSoC is also down to 1.24V again (apparently the board is basing its auto value off the memory speed), and everything looks stable now, almost through the first cycle with no errors. Maybe it was the high vSoC that was causing issues, I might try again with 4600 and manually setting the vSoC to a reasonable value. 

 

Update 2: looks like the high SoC voltage really was the cause of the erros, after the 4533 run completed without errors, I'm now trying 4600 again with a fixed SoC voltage, first cycle completed without errors. I never heard of a voltage being too high causing errors, maybe heat issues? But nothing is getting hot during testing, temps are in the high 40s while the memory test is running.

[Alvin853]

To take this topic to a conclusion, this is what I ended up with:

 

While 4600 on the RAM was stable, the board did not like 2300 FCLK at all. I don't really know why, the symptoms were very strange. I wasn't getting any errors in memtest, I could boot into Windows and do everything just fine, even run stress tests or benchmarks, for a few minutes, then the system would just power cycle. No error messages, no bluescreen, just straight up restart. I'm thinking my mainboard might be triggering some overcurrent protection and cutting power, or something similar. So since I had to back the FCLK down to 2267 anyway, I decided to drop the RAM to 4533 too, to keep the 1:1 ratio, and 4533:2267 did indeed perform slightly better than 4600:2267, even though my prior testing showed higher memory clock being beneficial. But with only 67 MHz difference, the boost from the higher memory speed did not exceed the loss from dropping the 1:1 ratio. 

 

Unless anybody can come up with a solution to the power cycles, I'm keeping these settings. Still a little bit worried about the SoC voltage being too high, but leaving it on "Auto" resulted in an even higher voltage, so it should be fine.

 

Edit (For anyone finding this thread in the future and running into the same issue): the graphics processing unit on modern APUs is very power hungry and also powered off the SoC rail. So as soon as I was running any GPU intensive application, my SoC voltage was dropping by up to 100mV, resulting in instabilities and hard crashes. By setting SoC LLC to the highest mode, I was able to stabilize the SoC voltage under load and the issues disappeared. But when I began overclocking the graphics unit, the issue reappeared, the SoC VRM on this board was not designed for powerful APUs like the 5600G, so I'm keeping everything at 4533:2267 for now as that's perfectly stable. Also, setting any mode of CPU LLC would disable the SoC LLC and crashes returned, so I'm very limited in terms of CPU overclocking, but with a 5600G on integrated graphics, the CPU isn't a bottleneck in any situation, so I'm fine with that. 

[ilyas2699]

On 2/5/2022 at 11:15 PM, Alvin853 said:

To take this topic to a conclusion, this is what I ended up with:

 

While 4600 on the RAM was stable, the board did not like 2300 FCLK at all. I don't really know why, the symptoms were very strange. I wasn't getting any errors in memtest, I could boot into Windows and do everything just fine, even run stress tests or benchmarks, for a few minutes, then the system would just power cycle. No error messages, no bluescreen, just straight up restart. I'm thinking my mainboard might be triggering some overcurrent protection and cutting power, or something similar. So since I had to back the FCLK down to 2267 anyway, I decided to drop the RAM to 4533 too, to keep the 1:1 ratio, and 4533:2267 did indeed perform slightly better than 4600:2267, even though my prior testing showed higher memory clock being beneficial. But with only 67 MHz difference, the boost from the higher memory speed did not exceed the loss from dropping the 1:1 ratio. 

 

Unless anybody can come up with a solution to the power cycles, I'm keeping these settings. Still a little bit worried about the SoC voltage being too high, but leaving it on "Auto" resulted in an even higher voltage, so it should be fine.

 

Edit (For anyone finding this thread in the future and running into the same issue): the graphics processing unit on modern APUs is very power hungry and also powered off the SoC rail. So as soon as I was running any GPU intensive application, my SoC voltage was dropping by up to 100mV, resulting in instabilities and hard crashes. By setting SoC LLC to the highest mode, I was able to stabilize the SoC voltage under load and the issues disappeared. But when I began overclocking the graphics unit, the issue reappeared, the SoC VRM on this board was not designed for powerful APUs like the 5600G, so I'm keeping everything at 4533:2267 for now as that's perfectly stable. Also, setting any mode of CPU LLC would disable the SoC LLC and crashes returned, so I'm very limited in terms of CPU overclocking, but with a 5600G on integrated graphics, the CPU isn't a bottleneck in any situation, so I'm fine with that. 

Hi,

I overclocked my 5600g's IGPU and my Crucial Ballistix(8G Micron E-Die) 3000 Mhz rams.  

CPU: 1.17 V 4200MHZ

IGPU : 2200 MHZ 1.15 GFX Voltage (SoC is auto)

Ram: 3933 Mhz // FCLK 1967// CL16 // 1.5V

 

I saw your thread and tried to overclock my IGPU to 2400 Mhz and overclock my rams up to the 4533 MHZ// 2267 FCLK// CL18 (all other timings are same with you) and it worked but system was not stable with 1.25 VSoC voltage so ı ıncreased VSoC voltage to the 1.28 (VSoC LLC mode 3 ) .   After increasing VSoC voltage system was stable and passed all the test ı did  (Aida64,Furmark,Time SPY,Cinebench R20,Memtest86, games)and it never got crush  But there were no big  performance difference between my  old overclock settings and new overclock settings in grapcih tests and games  so ı decided to use my old settings because of the VSoC voltage difference.(I think 1.28 v VSoC is too high for small performance difference)

 

Time Spy with old settings;( 2200Mhz IGPU// 1967 MHZ FCLK// 4.2 GHZ CPU)

 

Time SPY   2267 Fclk // 2400Mhz IGPU// 4200 MHZ CPU (Cpu settings were same but ı dont know why it has lower score in test)

 

Can you make a Time spy test with your overclock settings and share the results in this thread.

[Alvin853]

4 hours ago, ilyas2699 said:

But there were no big  performance difference between my  old overclock settings and new overclock settings in grapcih tests and games  so ı decided to use my old settings because of the VSoC voltage difference.

Hmm interesting... I might have to try if my memory can do CL16 at lower clocks and how that affects performance. I've only been trying CL18 so far. My iGPU starts artifacting at 2.4GHz, so I have it running at 2.35GHz right now.

 

4 hours ago, ilyas2699 said:

I think 1.28 v VSoC is too high for small performance difference

I still haven't found a reliable source what SoC-Voltage is considered "safe" on Cezanne. For Vermeer (Ryzen 5000X CPUs), people suggest not to go over 1.15V, but I've seen a Cezanne iGPU-Overclocking guide, where they upped the iGPU (and thus also SoC) voltage straight to 1.3V, so up to 1.3V should be safe, at least that's what I'm assuming for now.

 

4 hours ago, ilyas2699 said:

Can you make a Time spy test with your overclock settings and share the results in this thread.

I'm out of town for the week, will be back on the weekend and do a Time Spy test. What's that red exclamation mark on your screenshot?

Do you have access to Fire Strike? My final Fire Strike score was around 4550. 

[ilyas2699]

5 hours ago, Alvin853 said:

Hmm interesting... I might have to try if my memory can do CL16 at lower clocks and how that affects performance. I've only been trying CL18 so far. My iGPU starts artifacting at 2.4GHz, so I have it running at 2.35GHz right now.

 

I still haven't found a reliable source what SoC-Voltage is considered "safe" on Cezanne. For Vermeer (Ryzen 5000X CPUs), people suggest not to go over 1.15V, but I've seen a Cezanne iGPU-Overclocking guide, where they upped the iGPU (and thus also SoC) voltage straight to 1.3V, so up to 1.3V should be safe, at least that's what I'm assuming for now.

 

I'm out of town for the week, will be back on the weekend and do a Time Spy test. What's that red exclamation mark on your screenshot?

Do you have access to Fire Strike? My final Fire Strike score was around 4550. 

I did fire strike test:

 

Red exclamation says that ı use modified radeon software.Because ı am using modified 20.10.4 Radeon Software  version to use Radeon RELIVE with 5600g

 

 

 

 

[Alvin853]

20 hours ago, ilyas2699 said:

Red exclamation says that ı use modified radeon software.Because ı am using modified 20.10.4 Radeon Software  version to use Radeon RELIVE with 5600g

I think it's complaining your modified driver is restricting some of the features the benchmark uses, which makes the result invalid, as the score can't be compared to a result that has the feature unrestricted. Which means, the results can't be compared to my results, since you're not running the same workload as me.

 

Quote

Benchmark tessellation load modified, result invalid

---

This warning appears when you run 3DMark on a system with modified driver tessellation settings.

In order to produce scores that can be compared fairly, a benchmark must perform the same work on every system. Benchmark-specific driver optimizations are not allowed.

To get a valid 3DMark score for your system, return all driver options to their default settings then re-run the benchmark.

---

Some recent versions of AMD Radeon Software Adrenalin Edition include an application-specific profile for 3DMark. Running 3DMark with this profile will result in an invalid score, even if the profile does not modify any settings.

To get a valid score, open the AMD Radeon Software Adrenalin Edition application and remove the 3DMark application profile. Alternatively, you can use the Factory Reset option to restore all driver settings and profiles to the default values.

 

[Alvin853]

On 3/7/2022 at 8:38 PM, ilyas2699 said:

I did fire strike test:

 

Red exclamation says that ı use modified radeon software.Because ı am using modified 20.10.4 Radeon Software  version to use Radeon RELIVE with 5600g

 

 

 

 

Here are my Fire Strike

and Time Spy scores

 

As you can see my Graphics scores are lower than yours, but my results are valid, so most likely your modified driver is also tampering with the graphics scores. My CPU scores are higher, why are you limiting your CPU to 4.2GHz? I simply enabled PBO with curve optimizer, and it will do 4.5GHz all core and 4.65GHz single core boost and correctly clock down when idle so idle temps and power usage are lower, it's really not a good idea to do a manual overclock on Zen3 CPUs. 

 

I couldn't get your timings to run on my memory kit, I think 20 for RCDRD at 3933 is not possible on my kit, but I didn't bother fiddling around with it, so I don't know if your settings would perform better than mine. 

[freeagent]

On 1/30/2022 at 10:35 AM, Alvin853 said:

Wow never made the connection that odd CL timings are related to command rate. Thanks for the info

Gear down mode prefers to run even numbers, in order to run something like 15-15-15 you would have to disable GDM. GDM off and 1T really puts the hurt on. GDM off and 2T is for stability. It is still very quick though..