Single core or All core that is the question...

[xSOSxHawkens]

OK guys. Recently upgraded to the wonderful 3900x, and it seems I have a decent win on the silicon lottery side, but as I am on the stock prism max I will have to compromise on the OC for the time being till I can get better cooling...

At all stock setting my chip is in the 86th percentile overall vs other 3900x's. Under full load it consistantly holds 4.0-4.1 with stock cooling and all settings, with normal boost to 4.4 under low loads.

From what I have seen I can drop a CCD and get closer to 4.5-4.6 boost from the single CCD for gaming, but have not yet tried that.

In terms of OC, I seem to be able to push my two best cores to about 4.8-4.9 at stock voltages without hitting stability issues, but I do start to hit thermals into the 90's.

Alternatively I can seem to pull a solid 4.3 all core at 1.25 vcore and a solid 4.4 all core at 1.32v. Both of those give me the best overall multi-core score in all benchmarks, but I am always trailing about 3-5% from my stock single core score for some reason...

I dont think I will be dropping to a single CCD much, as even when gaming I often have use for the extra cores (but I might on ocassion go "game mode").

But for the rest of the time, do I run 3.8ish base with 4.5 boost, or do I instead run the all core @ 4.3/4.4 locked?

Rest of system:
32GB/3600/cl16
Vega 64
MSI Meg Unify x570

Opinions?

[r.linder]

All overclocking with the 3900X is barely better than PBO+AutoOC, and is still beaten in gaming by PBO+AutoOC+LCM (aka Game Mode) due to not being held back by latency between CCD 1 and 2 (as well as the lower clocked CCXs in CCD 2, as task scheduler can easily decide to use one of those cores if they aren't being used)

I got 4.6 to work with one CCD on all 6 cores, but it runs hot (and fails stress testing), hotter than stock while running 12 cores and 24 threads. That says something, because the amount of cores and threads has its own effect on power draw, and heat as a result. I can almost guarantee that 4.6 all-core with only one CCD will be too hot for the Prism to handle, even if you drop SMT.

CCD 2 will also have a hard time running 4.3~4.4, as is the case with my 3900X. It POSTs, and passes stress tests, but games crash, and the voltage was already set to 1.4v with LLCs on Turbo.

TL;DR

I bounce between 12C/24T with CCX OC (4.55 CCX1, 4.45 CCX2, 4.35 CCX3, 4.25 CCX4 @ 1.38v) for things that need the cores, but I use Game Mode from time to time.

https://valid.x86.fr/mt77j1

[Bad 💀 Motha]

Just run it stock until you have better cooler, such as BeQuiet Dark Rock Pro 4

[xSOSxHawkens]

Originally posted by Bad 💀 Motha:

Just run it stock until you have better cooler, such as BeQuiet Dark Rock Pro 4

Right now thats what I am thinking of doing. Stock ballances out the cores both ways, and is still in the 86th percentile for usermark results with a solid 4.0-4.1 all core.

[Napoleonic S]

Originally posted by Escorve:

All overclocking with the 3900X is barely better than PBO+AutoOC, and is still beaten in gaming by PBO+AutoOC+LCM (aka Game Mode) due to not being held back by latency between CCD 1 and 2 (as well as the lower clocked CCXs in CCD 2, as task scheduler can easily decide to use one of those cores if they aren't being used)

I got 4.6 to work with one CCD on all 6 cores, but it runs hot (and fails stress testing), hotter than stock while running 12 cores and 24 threads. That says something, because the amount of cores and threads has its own effect on power draw, and heat as a result. I can almost guarantee that 4.6 all-core with only one CCD will be too hot for the Prism to handle, even if you drop SMT.

CCD 2 will also have a hard time running 4.3~4.4, as is the case with my 3900X. It POSTs, and passes stress tests, but games crash, and the voltage was already set to 1.4v with LLCs on Turbo.

TL;DR

I bounce between 12C/24T with CCX OC (4.55 CCX1, 4.45 CCX2, 4.35 CCX3, 4.25 CCX4 @ 1.38v) for things that need the cores, but I use Game Mode from time to time.

https://valid.x86.fr/mt77j1

What exactly is LCM (Game Mode)?

For me I even turned off all the auto oc, using - 0.1v offset voltage and 1usmus ryzen universal power plan mode.

I don't see the difference in clockspeed while gaming, in fact the real difference I got was from switching BIOS, with my X570 Aorus Master I got slightly better clockspeed when using 1003 ABBA AGESA BIOS than the 1004 version.

I'd just recommend Ryzen 3000 users to just leave the settings on stock, there's no point for oc in daily usage because unless you go extreme oc there's no significant difference.

Tweaking RAM and Infinity Fabric may actually yield much more significant results.

[r.linder]

Originally posted by Napoleonic S:

Originally posted by Escorve:

All overclocking with the 3900X is barely better than PBO+AutoOC, and is still beaten in gaming by PBO+AutoOC+LCM (aka Game Mode) due to not being held back by latency between CCD 1 and 2 (as well as the lower clocked CCXs in CCD 2, as task scheduler can easily decide to use one of those cores if they aren't being used)

I got 4.6 to work with one CCD on all 6 cores, but it runs hot (and fails stress testing), hotter than stock while running 12 cores and 24 threads. That says something, because the amount of cores and threads has its own effect on power draw, and heat as a result. I can almost guarantee that 4.6 all-core with only one CCD will be too hot for the Prism to handle, even if you drop SMT.

CCD 2 will also have a hard time running 4.3~4.4, as is the case with my 3900X. It POSTs, and passes stress tests, but games crash, and the voltage was already set to 1.4v with LLCs on Turbo.

TL;DR

I bounce between 12C/24T with CCX OC (4.55 CCX1, 4.45 CCX2, 4.35 CCX3, 4.25 CCX4 @ 1.38v) for things that need the cores, but I use Game Mode from time to time.

https://valid.x86.fr/mt77j1

What exactly is LCM (Game Mode)?

For me I even turned off all the auto oc, using - 0.1v offset voltage and 1usmus ryzen universal power plan mode.

I don't see the difference in clockspeed while gaming, in fact the real difference I got was from switching BIOS, with my X570 Aorus Master I got slightly better clockspeed when using 1003 ABBA AGESA BIOS than the 1004 version.

I'd just recommend Ryzen 3000 users to just leave the settings on stock, there's no point for oc in daily usage because unless you go extreme oc there's no significant difference.

Tweaking RAM and Infinity Fabric may actually yield much more significant results.

1. Legacy Compatibility Mode, it's the same as choosing how many CCDs to have enabled in BIOS (only available with CPUs that have more than one chiplet, AKA Ryzen 9 and TR)
It's included with the Game Mode preset in Ryzen Master, and removes potential issues with game performance as games aren't designed to work with multiple chiplets, and neither is windows task scheduler.

2. It takes a lot of fiddling to get the advertised clocks, but it's easiest to run a CCX OC in BIOS if you have the latest BIOS with AGESA 1.0.0.4+. It's not an overclock usually in the sense that you're not going above advertised speeds, but it's rather a manual control of the clocks based on what your CCDs are actually capable of; Ryzen 9 3900X's CCD 1 usually only handles 4.5~4.6, while CCD 2 struggles with 4.3~4.4 unless the lotto was kind to you.
The highest stable working CCX OC I can get properly working is 4.55, 4.45, 4.35, and 4.25 @ 1.38v, and I'm not willing to go above 1.4 as it's already toasty at 1.38v with turbo LLC.

3. As long as the IF is a 1:1 ratio with your RAM, it's fine. Best performance comes with 3600 MHz and an 1800 MHz FCLK, but there's compatibility issues if you're using more than 2 sticks of RAM, and it can still have issues with only 2 regardless. It depends on the motherboard, as lower priced boards are more likely to struggle with RAM.

[Napoleonic S]

Originally posted by Escorve:

Originally posted by Napoleonic S:

What exactly is LCM (Game Mode)?

For me I even turned off all the auto oc, using - 0.1v offset voltage and 1usmus ryzen universal power plan mode.

I don't see the difference in clockspeed while gaming, in fact the real difference I got was from switching BIOS, with my X570 Aorus Master I got slightly better clockspeed when using 1003 ABBA AGESA BIOS than the 1004 version.

I'd just recommend Ryzen 3000 users to just leave the settings on stock, there's no point for oc in daily usage because unless you go extreme oc there's no significant difference.

Tweaking RAM and Infinity Fabric may actually yield much more significant results.

1. Legacy Compatibility Mode, it's the same as choosing how many CCDs to have enabled in BIOS (only available with CPUs that have more than one chiplet, AKA Ryzen 9 and TR)
It's included with the Game Mode preset in Ryzen Master, and removes potential issues with game performance as games aren't designed to work with multiple chiplets, and neither is windows task scheduler.

2. It takes a lot of fiddling to get the advertised clocks, but it's easiest to run a CCX OC in BIOS if you have the latest BIOS with AGESA 1.0.0.4+. It's not an overclock usually in the sense that you're not going above advertised speeds, but it's rather a manual control of the clocks based on what your CCDs are actually capable of; Ryzen 9 3900X's CCD 1 usually only handles 4.5~4.6, while CCD 2 struggles with 4.3~4.4 unless the lotto was kind to you.
The highest stable working CCX OC I can get properly working is 4.55, 4.45, 4.35, and 4.25 @ 1.38v, and I'm not willing to go above 1.4 as it's already toasty at 1.38v with turbo LLC.

3. As long as the IF is a 1:1 ratio with your RAM, it's fine. Best performance comes with 3600 MHz and an 1800 MHz FCLK, but there's compatibility issues if you're using more than 2 sticks of RAM, and it can still have issues with only 2 regardless. It depends on the motherboard, as lower priced boards are more likely to struggle with RAM.

How do software and games handle asymmetrical clockspeed across different threads though? Wouldn't that cause problems with games? If you have main game thread running at 4.6 GHz and AI/Physics/others threads running at slower speed, wouldn't it cause async issue or bottlenecking the game engine?

[r.linder]

Originally posted by Napoleonic S:

Originally posted by Escorve:

1. Legacy Compatibility Mode, it's the same as choosing how many CCDs to have enabled in BIOS (only available with CPUs that have more than one chiplet, AKA Ryzen 9 and TR)
It's included with the Game Mode preset in Ryzen Master, and removes potential issues with game performance as games aren't designed to work with multiple chiplets, and neither is windows task scheduler.

2. It takes a lot of fiddling to get the advertised clocks, but it's easiest to run a CCX OC in BIOS if you have the latest BIOS with AGESA 1.0.0.4+. It's not an overclock usually in the sense that you're not going above advertised speeds, but it's rather a manual control of the clocks based on what your CCDs are actually capable of; Ryzen 9 3900X's CCD 1 usually only handles 4.5~4.6, while CCD 2 struggles with 4.3~4.4 unless the lotto was kind to you.
The highest stable working CCX OC I can get properly working is 4.55, 4.45, 4.35, and 4.25 @ 1.38v, and I'm not willing to go above 1.4 as it's already toasty at 1.38v with turbo LLC.

3. As long as the IF is a 1:1 ratio with your RAM, it's fine. Best performance comes with 3600 MHz and an 1800 MHz FCLK, but there's compatibility issues if you're using more than 2 sticks of RAM, and it can still have issues with only 2 regardless. It depends on the motherboard, as lower priced boards are more likely to struggle with RAM.

How do software and games handle asymmetrical clockspeed across different threads though? Wouldn't that cause problems with games? If you have main game thread running at 4.6 GHz and AI/Physics/others threads running at slower speed, wouldn't it cause async issue or bottlenecking the game engine?

Logically, yes, but I don't think it has much of an effect based on the fact that all threads won't be utilised in gaming so the cores will bounce around, but task scheduler picks what cores are used in real-time, so one core could be 4.5, while another one could be 4.2, and another could be 3.8.

What harms gaming performance more is multiple chiplets, because there's latency when cores from more than one chiplet are communicating. If you limit it to CCD 1 or use a program like Prio to limit cores to ones that are on CCD 1 or 2, the performance actually improves, and depending on the game, that improvement can be significant. (BDO is a huge example because it hates Ryzen)

A CCX OC can prove useful for the Prio method, because you can tell a specific process to only run on cores that you tell it to, such as the first 6 threads, or the first 6 even-numbered cores (which matters if SMT is enabled, as it appears that performance improves by doing so when SMT is enabled while there's no effect if it isn't)
Why it's useful is that it's still a manual OC.