Trying to understand Ryzen 3000 series boost speed variations

[lorry]

I have a 3900x and have enabled PBO, XMP in the bios and use Ryzen high performance power setting within Windows, but not much more than that.
On some boots it will boost to 4.600 on its two best cores, at other times to 4.575 (like today) and at other times 'only' 4.550. All the other cores do not seem to be affected like this, reaching their 'usual' peak boost. This info I am taking from HWINFO64 (one time it reported that the best cores boost was 4.650). Simple benchmarking scores are affected by this of course. When the cores boost to 4.600 the TDC reaches 93% of its limit, EDC 100% during benchmarking, but when the peak boost is the slightly lower figure, the TDC will only reach 89%.

I am not so much interested in what scores and peak speeds that I can reach (although they are enjoyable), but in what affects and limits those cores from reaching their slightly better speeds and what might be done to reduce them. It doesn't look to be temperature, as with the lower speeds that reaches 71c and with the highest speed, 72c (I presume that is due to the reduced max TDC?).
I ask because I can boot once and get a high of say 4.575, reboot and 'only' reach 4.550, try again and obtain that 4.600. Is that down to pure chance or is there something that can be done to improve things?

I was wondering if it might be Windows background processes or programs that I have installed lowering the max speed on that boot? If that is so though why then does that alter on the next boot when they are all still running yet not seemingly affecting top performance? As I said, I am not asking 'just' to chase my best speed, but more so that I can understand what is happening and what may or may not be able to be done to ensure best performance.

 

[Deleted member 185158]

A lot of the frequency drop isn;t the multiplier. It's the reference clock. You'll see it dip down a little from 100mhz as that's never steady. That's where most of that lower dips in clock speed is coming from. No worries, it's very normal.

There are a lot of settings for power control on AMD Ryzen chips. 3 different AUTO OC approaches, and they all do something different.

You can also alter system power management. This also will have an impact on performance.
For example, parking cores, then running a benchmark, the cores need to unpark. This takes time and energy.

TDP is Thermal design point. This is the removal of wattage through BTU. The average spec varies depending on overclocking, and it's a wide range.

At the cpu processor highest P-State, you will find power Wattage at load to be very near the rated TDP for the CPU. So between c-states, p-states and idle system to Boosted OC full load on an average, you'd have TDP rated spec.

For example, running my Zen+ 2700X at 3.7ghz, the usage wattage is just about exactly what the CPU is rated for.
Remember Boosting is an overclock. Even if it comes factory stock, it's an overclock.
So don't mind trying to compare rated TDP to actual electrical wattage usage. It's going to be way off.

So PBO, CPB and XFR are 3 separate features.
XFR is single core boost only and works when CPB and PBO are set to auto. When you alter either, XFR is disabled.
Manual (static o.c.) will disable all three (setting multiplier manually)
PBO, CPB and XFR are super efficient at power management. There's actually no real need to alter unless going full manual OC, or overclocking the PBO or CPB settings.

 

[lorry]

A lot of the frequency drop isn;t the multiplier. It's the reference clock. You'll see it dip down a little from 100mhz as that's never steady. That's where most of that lower dips in clock speed is coming from. No worries, it's very normal.

There are a lot of settings for power control on AMD Ryzen chips. 3 different AUTO OC approaches, and they all do something different.

You can also alter system power management. This also will have an impact on performance.
For example, parking cores, then running a benchmark, the cores need to unpark. This takes time and energy.

TDP is Thermal design point. This is the removal of wattage through BTU. The average spec varies depending on overclocking, and it's a wide range.

At the cpu processor highest P-State, you will find power Wattage at load to be very near the rated TDP for the CPU. So between c-states, p-states and idle system to Boosted OC full load on an average, you'd have TDP rated spec.

For example, running my Zen+ 2700X at 3.7ghz, the usage wattage is just about exactly what the CPU is rated for.
Remember Boosting is an overclock. Even if it comes factory stock, it's an overclock.
So don't mind trying to compare rated TDP to actual electrical wattage usage. It's going to be way off.

So PBO, CPB and XFR are 3 separate features.
XFR is single core boost only and works when CPB and PBO are set to auto. When you alter either, XFR is disabled.
Manual (static o.c.) will disable all three (setting multiplier manually)
PBO, CPB and XFR are super efficient at power management. There's actually no real need to alter unless going full manual OC, or overclocking the PBO or CPB settings.

Thank you, you've given me a lot to look into and more
Seriously though, you have given me pointers, which is brilliant. I'm very new to all of this, so have Lots to learn.
When you say reference clock is that the bus clock as used in HWINFO64?

Any good pointers as to where to start reading at all please?

 

[Deleted member 185158]

There are quite a few guides out there. Here's one that I really like alot.

FRONTPAGE - AMD RYZEN Overclocking Guide

The time is finally here, AMD has launched a game changing CPU with incredible leaps in performance compared to their previous FX CPU line. My intention here is to aid you in leveraging just a bit more out of the AMD Ryzen CPU. One great thing that AMD carried forward is that all Ryzen CPUs have...

www.overclockers.com

Well yea I have pointers for sure.
Performance gains with memory would be any overclockers concern.
Your manual OC probably won't beat the factory OC any ways......
So you'll read there and see what kind of options you have for tweaks.
Come back with questions and off we go!!!

 

[lorry]

There are quite a few guides out there. Here's one that I really like alot.

FRONTPAGE - AMD RYZEN Overclocking Guide

The time is finally here, AMD has launched a game changing CPU with incredible leaps in performance compared to their previous FX CPU line. My intention here is to aid you in leveraging just a bit more out of the AMD Ryzen CPU. One great thing that AMD carried forward is that all Ryzen CPUs have...

www.overclockers.com

Well yea I have pointers for sure.
Performance gains with memory would be any overclockers concern.
Your manual OC probably won't beat the factory OC any ways......
So you'll read there and see what kind of options you have for tweaks.
Come back with questions and off we go!!!

oh memory I know is going to be 'fun' as what I have I've discovered is decent but not brilliant, Corsair dominator platinum cl16 4 x 8GB

 

[Deleted member 185158]

Work with what you've got. Then it's good you don't have super high expectations.

Ram speed highest obtainable is your main direction. Then tweak timings for better performance.
A little over-volting never hurts. I don't run less than 1.410v mem when overclocking.

I have my SK Hynix in the 2700X right now. It's a pain in the butt to tweak, but not bad. I'm running XMP right now, can tighten these up with some sacrifice depending how hard I push.

So right now, true XMP stability at 1562mhz reading 16-17-17 timings, super loose. But very stable for me. 3200mhz is my best supprted speed on this motherboard, so tweaking at this frequency always seems the easiest.

 

[lorry]

Work with what you've got. Then it's good you don't have super high expectations.

Ram speed highest obtainable is your main direction. Then tweak timings for better performance.
A little over-volting never hurts. I don't run less than 1.410v mem when overclocking.

I have my SK Hynix in the 2700X right now. It's a pain in the butt to tweak, but not bad. I'm running XMP right now, can tighten these up with some sacrifice depending how hard I push.

So right now, true XMP stability at 1562mhz reading 16-17-17 timings, super loose. But very stable for me. 3200mhz is my best supprted speed on this motherboard, so tweaking at this frequency always seems the easiest.

And don't try running until I can walk I'm sure. Thank you, I have lots to get my head around I'm sure. It's good to know that i can ask and learn

 

[Deleted member 185158]

Todays processors are much different than what I started on, but all the same principles apply.

No 2 rigs are the same, we dont really have golden 1 shot answers for that reason.
Tweak and test tweak and test.
Thats what it takes to dial in a gaming PC.

 

[tabascosauz]

I ask because I can boot once and get a high of say 4.575, reboot and 'only' reach 4.550, try again and obtain that 4.600. Is that down to pure chance or is there something that can be done to improve things?

As of recent revisions, HWInfo shows Effective Clock parameters. Your "effective clock" readings are going to be much more accurate of performance than peak clock, because your 3900X isn't holding a 46x multiplier, even on a single core, for any meaningful period of time. It will "spike" to 46x for a split second, then spend most of its time at no more than 45x. Effective Clock essentially takes an average of the frequency over a short period of time, as HWInfo's polling rate will never reflect the 1-2ns reaction time of Ryzen 3000.

The peak clock reading is useful for judging differences in boosting behaviour between AGESA versions, and pretty much nothing else. FWIW, there are days when my rig sees a lot of multi-core load from startup to shutdown, and at the end of the day cores 5 and 7 will still be hitting 43.5x max. On other days when it's idling for 5 hours at a time, HWInfo will record 44x on those two cores. Which is a good example of how useless peak clock actually is, on stock boost behaviour.

Cores 5 and 7 are my best on my 3700X. Most of the time, if I have my PC running for the whole day, the multipliers will get up to 44x on those two. But like I said, it's pointless and has no bearing on performance. I'm afraid this is what AMD's ambitious marketing and consumers' expectations from 22nm and 14/12nm have resulted in - paranoia over something that actually has zero relevance. I would much rather see Effective Clock hovering between 4200MHz and 4300MHz when under load, than the magical 44x that means nothing in benchmarks or everyday usage.

On Ryzen 3000, the load is juggled every couple of seconds so that a single core will never bear the brunt of the load indefinitely. Even though the 1usmus plan minimizes the switching, it's inevitable and just how it works.

A lot of the frequency drop isn;t the multiplier. It's the reference clock. You'll see it dip down a little from 100mhz as that's never steady. That's where most of that lower dips in clock speed is coming from. No worries, it's very normal.

It might vary from board to board, but mine is always a static 99.8MHz in usage. That accounts for about a 8MHz discrepancy at 44x - 4392MHz. Any more than that, and it's the multiplier that's responsible. For example, 25MHz is probably going to be a multiplier difference of 0.25x.

 

[lorry]

Todays processors are much different than what I started on, but all the same principles apply.

No 2 rigs are the same, we dont really have golden 1 shot answers for that reason.
Tweak and test tweak and test.
Thats what it takes to dial in a gaming PC.

Tweak, test and learn perhaps?

 

[Deleted member 185158]

Tweak, test and learn perhaps?

You'll be doing lots of that. The suggestions ideas and opinions at this forum from everyone will put you in a good direction learning and actually doing.

 

[lorry]

As of recent revisions, HWInfo shows Effective Clock parameters. Your "effective clock" readings are going to be much more accurate of performance than peak clock, because your 3900X isn't holding a 46x multiplier, even on a single core, for any meaningful period of time. It will "spike" to 46x for a split second, then spend most of its time at no more than 45x. Effective Clock essentially takes an average of the frequency over a short period of time, as HWInfo's polling rate will never reflect the 1-2ns reaction time of Ryzen 3000.

The peak clock reading is useful for judging differences in boosting behaviour between AGESA versions, and pretty much nothing else. FWIW, there are days when my rig sees a lot of multi-core load from startup to shutdown, and at the end of the day cores 5 and 7 will still be hitting 43.5x max. On other days when it's idling for 5 hours at a time, HWInfo will record 44x on those two cores. Which is a good example of how useless peak clock actually is, on stock boost behaviour.

Cores 5 and 7 are my best on my 3700X. Most of the time, if I have my PC running for the whole day, the multipliers will get up to 44x on those two. But like I said, it's pointless and has no bearing on performance. I'm afraid this is what AMD's ambitious marketing and consumers' expectations from 22nm and 14/12nm have resulted in - paranoia over something that actually has zero relevance. I would much rather see Effective Clock hovering between 4200MHz and 4300MHz when under load, than the magical 44x that means nothing in benchmarks or everyday usage.

On Ryzen 3000, the load is juggled every couple of seconds so that a single core will never bear the brunt of the load indefinitely. Even though the 1usmus plan minimizes the switching, it's inevitable and just how it works.



It might vary from board to board, but mine is always a static 99.8MHz in usage. That accounts for about a 8MHz discrepancy at 44x - 4392MHz. Any more than that, and it's the multiplier that's responsible. For example, 25MHz is probably going to be a multiplier difference of 0.25x.

I have 1usmus plan installed and am using it (another piece of good advice from here).
I'm not looking for fantastic figures as such but do want my setup to be as good, or as near good as it can be. After all I've invested a tidy sum in it, going to go down the open loop eventually, so I want to not only have a properly specced rig, but know Why it is.

You'll be doing lots of that. The suggestions ideas and opinions at this forum from everyone will put you in a good direction learning and actually doing.

That's good to hear, thanks

 

[tabascosauz]

I have 1usmus plan installed and am using it (another piece of good advice from here).
I'm not looking for fantastic figures as such but do want my setup to be as good, or as near good as it can be. After all I've invested a tidy sum in it, going to go down the open loop eventually, so I want to not only have a properly specced rig, but know Why it is.

"As it can be"? We're on the first iteration of TSMC's 7nm process. Your chip knows exactly how good it can be based on its own binning, as CPPC tells itself exactly that, and is already giving you as much performance as it can while respecting its original boost and Pstate behaviour. This isn't 2014, this isn't 22nm. No longer do we have that kind of headroom to experiment with.

You ask why? Because running 46x for an extended amount of time is not feasible on this silicon. There's not much more to it.

It might be easier to demonstrate this visually:

Here, Cores 5 and 7 being the best silicon on die is plain to see. But while the Peak is reported as a mythical 44x and 43.5x (well, not so magical, since it goes up to 43.8x and 44x) respectively, effective clock tops out at roughly 41.25x on both. Why? But aren't I supposed to be getting 4392MHz out of at least my two best cores?

Because I haven't done anything with my PC today, that's why. The 41.25x is merely reflective of CPU-Z's quick multi-threaded benchmark, which pushes all the cores up to 40.5-41.5x. Meanwhile, Core 0 (unfortunately, my worst core, but that's just how the lottery goes) has the highest effective clock of them all because CPU-Z defaults to Core 0 for the single-threaded bench. The 44x has indeed been reached, but it's such a transient event that it's all but pointless.

Think of it like the 8086K's turbo tables. Yes, it can reach 5GHz, but that's only for a single core, and in 2019 how many instances are there in which you will be running any sort of load in Windows 1909 on just one core?

I mean, if you want that affirmation of the 46x multiplier to make yourself feel a little more comforted, then the simple method is to run your PC for longer periods of time. Leave it on for the whole day, and you'll get your two 46x cores under Peak Clock. But all it takes is one quick look a bit down the HWInfo list to see that it's a more complicated ordeal than the 46x multiplier seems to imply.

 

[lorry]

"As it can be"? We're on the first iteration of TSMC's 7nm process. Your chip knows exactly how good it can be based on its own binning, as CPPC tells itself exactly that, and is already giving you as much performance as it can while respecting its original boost and Pstate behaviour. This isn't 2014, this isn't 22nm. No longer do we have that kind of headroom to experiment with.

You ask why? Because running 46x for an extended amount of time is not feasible on this silicon. There's not much more to it.

It might be easier to demonstrate this visually:

View attachment 138844

Here, Cores 5 and 7 being the best silicon on die is plain to see. But while the Peak is reported as a mythical 44x and 43.5x (well, not so magical, since it goes up to 43.8x and 44x) respectively, effective clock tops out at roughly 41.25x on both. Why? But aren't I supposed to be getting 4392MHz out of at least my two best cores?

Because I haven't done anything with my PC today, that's why. The 41.25x is merely reflective of CPU-Z's quick multi-threaded benchmark, which pushes all the cores up to 40.5-41.5x. Meanwhile, Core 0 (unfortunately, my worst core, but that's just how the lottery goes) has the highest effective clock of them all because CPU-Z defaults to Core 0 for the single-threaded bench. The 44x has indeed been reached, but it's such a transient event that it's all but pointless.

Think of it like the 8086K's turbo tables. Yes, it can reach 5GHz, but that's only for a single core, and in 2019 how many instances are there in which you will be running any sort of load in Windows 1909 on just one core?

I mean, if you want that affirmation of the 46x multiplier to make yourself feel a little more comforted, then the simple method is to run your PC for longer periods of time. Leave it on for the whole day, and you'll get your two 46x cores under Peak Clock. But all it takes is one quick look a bit down the HWInfo list to see that it's a more complicated ordeal than the 46x multiplier seems to imply.

Yes i was looking at the effective clock speed numbers and wondered about them, but decided that would best be looked into tomorrow (I'm already upstairs with the rig turned off). Again though, more pointers of what to look for, thanks

I am lucky I guess then, in that core 0 is one of my best from what I can see

As of recent revisions, HWInfo shows Effective Clock parameters. Your "effective clock" readings are going to be much more accurate of performance than peak clock, because your 3900X isn't holding a 46x multiplier, even on a single core, for any meaningful period of time. It will "spike" to 46x for a split second, then spend most of its time at no more than 45x. Effective Clock essentially takes an average of the frequency over a short period of time, as HWInfo's polling rate will never reflect the 1-2ns reaction time of Ryzen 3000.

The peak clock reading is useful for judging differences in boosting behaviour between AGESA versions, and pretty much nothing else. FWIW, there are days when my rig sees a lot of multi-core load from startup to shutdown, and at the end of the day cores 5 and 7 will still be hitting 43.5x max. On other days when it's idling for 5 hours at a time, HWInfo will record 44x on those two cores. Which is a good example of how useless peak clock actually is, on stock boost behaviour.

Cores 5 and 7 are my best on my 3700X. Most of the time, if I have my PC running for the whole day, the multipliers will get up to 44x on those two. But like I said, it's pointless and has no bearing on performance. I'm afraid this is what AMD's ambitious marketing and consumers' expectations from 22nm and 14/12nm have resulted in - paranoia over something that actually has zero relevance. I would much rather see Effective Clock hovering between 4200MHz and 4300MHz when under load, than the magical 44x that means nothing in benchmarks or everyday usage.

On Ryzen 3000, the load is juggled every couple of seconds so that a single core will never bear the brunt of the load indefinitely. Even though the 1usmus plan minimizes the switching, it's inevitable and just how it works.



It might vary from board to board, but mine is always a static 99.8MHz in usage. That accounts for about a 8MHz discrepancy at 44x - 4392MHz. Any more than that, and it's the multiplier that's responsible. For example, 25MHz is probably going to be a multiplier difference of 0.25x.

Two questions -

Does setting PBO max boost to 200 actually do anything? (as I cannot see much in the way of change)

My effective clock seems to be 4159 - 4161 for the majority with the best core being 4317 and another at 4224 and a bus clock of 100.
Is that normal, good, bad? Or do I have a setting not right?



Just noticed that the infinity fabric is at 1600, should that be set to 1799? If so, how please?

I wish that the CPU OC MAX was true mind! lol

There are quite a few guides out there. Here's one that I really like alot.

FRONTPAGE - AMD RYZEN Overclocking Guide

The time is finally here, AMD has launched a game changing CPU with incredible leaps in performance compared to their previous FX CPU line. My intention here is to aid you in leveraging just a bit more out of the AMD Ryzen CPU. One great thing that AMD carried forward is that all Ryzen CPUs have...

www.overclockers.com

Well yea I have pointers for sure.
Performance gains with memory would be any overclockers concern.
Your manual OC probably won't beat the factory OC any ways......
So you'll read there and see what kind of options you have for tweaks.
Come back with questions and off we go!!!

just finished reading the article, very interesting, i'll come back with questions after I've read it again and can pick out parts where I have a query.
In the meantime can you have a look at my last comment about my effective clock speeds etc if possible please when you can I mean.

 

[Zach_01]

InfinityFabric clock should match the DRAM clock for minimizing a latency penalty between the CPU and memory. Your ram now runs at 3200MT/s so the actual clock is 1600MHz and the IF match it. So you are ok for know. Just check in RyzenMaster that all subsystems match speed.
RAM:MemoryController:InfinityFabric
Remember the video I linked you a week ago about ZEN2 memory subsystem?
It explains a lot...

Just check in RyzenMaster that all subsystems match speed.
RAM:MemoryController:InfinityFabric
(MCLK:UCLK:FCLK)

And something about the effective clock by the author of HWiNFO

FEATURE - Effective clock vs instant (discrete) clock

It has become a common practice for several years to report instant (discrete) clock values for CPUs. This method is based on knowledge of the actual bus clock (BCLK) and sampling of core ratios at specific time points. The resulting clock is then a simple result of ratio * BCLK. Such approach...

www.hwinfo.com

 

[lorry]

Yes I remember you saying that, so I did wonder if it was linked to that but wasn't certain, hence my next read being 'Theory: Infinity Fabric and types of Ram' from this forum.
I am using 6.20.403 HWINFO, here are a couple of screenshots I just took



You can see the differences between the two readings (although it's a shame that you cannot see all 12 cores at once in RM)

The current values are closer together than the peak values though

It's all Very new isn't it, I'm sure that there are many confused people around, in fact I have seen various posts where even i could see how the poster isn't getting it, there's a lot to take in.

 

[Zach_01]

Yes there is confusion about the clocks and the boost behavior. All this couldn’t be explained while marketing the chip, and we as users are used to very different things and definition of boosting.
While AMD did not lie about boosting... did not disclosure the whole truth about boost (and it’s factors) either. For me the bottom line is that the performance is real as promised.

 

[lorry]

@Zach_01 what about both RM and the summary page of HW both reporting the max OC speed as being 4850 (4650 + 200 MHz max CPU boost clock)?

Ans I am still working through that video from BZ, this will be about the 6th time (or more) that I have seen it

Yes there is confusion about the clocks and the boost behavior. All this couldn’t be explained while marketing the chip, and we as users are used to very different things and definition of boosting.
While AMD did not lie about boosting... did not disclosure the whole truth about boost (and it’s factors) either. For me the bottom line is that the performance is real as promised.

Yes, I have said several times now that their PR could have done a lot better job I felt in preparing the users about what was going on, but I also realise that until you actually have one and experience what they are saying, can you appreciate what they say.

 

[Zach_01]

@Zach_01 what about both RM and the summary page of HW both reporting the max OC speed as being 4850 (4650 + 200 MHz max CPU boost clock)?

Have you seen 4850MHz at HWiNFO or RM? I'm not aware of any CPU auto clocking/boosting that high... 200~250MHz over rated boost. Maybe some rare chips can do +25~50MHz over stock max boost.

ZEN architecture is unique (so far) to the layout of cores, SoC/IO subsystems and the way they are implemented are introducing high latency among core-to-core (out of a CCX) and core-to-ram communication. AMD compensates in a way this high latency with large amounts of Level3 cache. This helps with the core-to-ram communication (reduces frequent access to ram) but still the core-to-core "talk" (out of CCX) is delayed. In order to do what best for this core-to-core (out of CCX) cross talk the infinity fabric must run as high as possible because this is the channel that cores talk through to each other. But IF cant clock all that high. Most chips can raise this clock between 1800~1900MHz. But you cant change this IF speed by its own. In order to not delay further the core-to-ram access you need to match with IF the speed of ram and the speed of the UMC (UnifiedMemoryController) that it is contained into the SoC along with IF.

Another important thing about effective clock that may not be clear. The reported eff clock of HWiNFO is not the actual clock of the cores. It contains also the halted(sleeping) states of the cores. As said by the author of HWiNFO there is no software that can catch every state change of cores perfectly accurate. The dynamics of the cores can change whithin a few mseconds and thus can change states a few hundread times within a whole second.

The pic below is after an hour of browsing work load. I have marked the 3 high quality cores in blue lines and the 3 less quality in orange. Threads are 2 per core. Core 0 has Thread 0(Physical core and colored) and Thread 1(logical core and black).
Watch the "perf # x/x" numbers. The second perf number is the order of the quality cores as the CPU reports them and as RM reports too (with stars and dots). The first perf number is the order of cores windows chooses to load.
1usmus's power plan and later maybe some windows/UEFI updates (rumoured) helped windows to choose the right cores to load first. Prior to this the loads where all over the cores and windows's scheduler did not pay well attention to core perf# order.
You can see that while the high quality cores are the 1/2/3 (in blue/red square) the loads are mostly on the cores 0/1/2 (green square) and core 3 has lower eff clock (86.1) than core 0 (217.9). This in theory may improve performance by win scheduler neglecting core 3 and choosing core 0 instead because the cores 0/1/2 are in the same CCX and this way may improve the core-to-core cross talk, even tho core 0 is 4th in actual perf# order.

All this is very new for all (users, software/windows developers). The platform is still in the early stages of development and it can be improved further from gen to gen along with software/windows.
ZEN3 (4000series) will continue to improve IPC and performance. What is heard so far is that cores within the same CCD (6 or 8) will be unified and the 2x 3/4core CCXs will be 1 CCD/CCX of 6 or 8 cores sharing the same resources. This will reduce the core-to-core latency for up to 8 cores instead of up to 4 cores that exists today. And I believe that they need to increase IF clock capability beyond 1800~1900MHz.
We'll see...

 

[lorry]

Have you seen 4850MHz at HWiNFO or RM? I'm not aware of any CPU auto clocking/boosting that high... 200~250MHz over rated boost. Maybe some rare chips can do +25~50MHz over stock max boost.

ZEN architecture is unique (so far) to the layout of cores, SoC/IO subsystems and the way they are implemented are introducing high latency among core-to-core (out of a CCX) and core-to-ram communication. AMD compensates in a way this high latency with large amounts of Level3 cache. This helps with the core-to-ram communication (reduces frequent access to ram) but still the core-to-core "talk" (out of CCX) is delayed. In order to do what best for this core-to-core (out of CCX) cross talk the infinity fabric must run as high as possible because this is the channel that cores talk through to each other. But IF cant clock all that high. Most chips can raise this clock between 1800~1900MHz. But you cant change this IF speed by its own. In order to not delay further the core-to-ram access you need to match with IF the speed of ram and the speed of the UMC (UnifiedMemoryController) that it is contained into the SoC along with IF.

Another important thing about effective clock that may not be clear. The reported eff clock of HWiNFO is not the actual clock of the cores. It contains also the halted(sleeping) states of the cores. As said by the author of HWiNFO there is no software that can catch every state change of cores perfectly accurate. The dynamics of the cores can change whithin a few mseconds and thus can change states a few hundread times within a whole second.

The pic below is after an hour of browsing work load. I have marked the 3 high quality cores in blue lines and the 3 less quality in orange. Threads are 2 per core. Core 0 has Thread 0(Physical core and colored) and Thread 1(logical core and black).
Watch the "perf # x/x" numbers. The second perf number is the order of the quality cores as the CPU reports them and as RM reports too (with stars and dots). The first perf number is the order of cores windows chooses to load.
1usmus's power plan and later maybe some windows/UEFI updates (rumoured) helped windows to choose the right cores to load first. Prior to this the loads where all over the cores and windows's scheduler did not pay well attention to core perf# order.
You can see that while the high quality cores are the 1/2/3 (in blue/red square) the loads are mostly on the cores 0/1/2 (green square) and core 3 has lower eff clock (86.1) than core 0 (217.9). This in theory may improve performance by win scheduler neglecting core 3 and choosing core 0 instead because the cores 0/1/2 are in the same CCX and this way may improve the core-to-core cross talk, even tho core 0 is 4th in actual perf# order.

View attachment 138959

All this is very new for all (users, software/windows developers). The platform is still in the early stages of development and it can be improved further from gen to gen along with software/windows.
ZEN3 (4000series) will continue to improve IPC and performance. What is heard so far is that cores within the same CCD (6 or 8) will be unified and the 2x 3/4core CCXs will be 1 CCD/CCX of 6 or 8 cores sharing the same resources. This will reduce the core-to-core latency for up to 8 cores instead of up to 4 cores that exists today. And I believe that they need to increase IF clock capability beyond 1800~1900MHz.
We'll see...

Re the reported boost of 4850, yes, I posted screenshots above and do so here to save going hunting for them


I know that this is only PBO being enabled and then Max CPU Boost Override being set to 200MHz (is that the way to do it? Or is it one or the other and not both?), but that could confuse people as they might then expect the max boost to actually be 4850. BTW even before applying the Max CPU Boost Override, my 'CPU OC Max in HW was still showing 4650, which I thought somewhat odd?
My bus clock is always on 100MHz, that never changes at all from what I have seen.

So to potentially improve speeds I would need to increase the timings of my Ram from 3200 to hopefully 3600? I say hopefully as I know that Bill Zoid calls the Corsair B die Dominator ram 'rubbish' so I don't know what to expect when/if I get around to that.

It would seem that CCD0 is a pretty good chiplet as it would seem as though 1 of the cores (core 0), 2 of the cores will hit 4.575 and 3 reach 4.550, according to HW.
I noted on that thread on HW site that the coder said that 1usmus's power plan did not cure the problem and I'm not too sure why he came to that conclusion. The workload does seem to be spread fairly evenly across my first four cores

I am wondering just how many more bios updates there will be before they are considered 'fixed', as from various threads that I have read, there does seem to be a number of confused users.

 

[tabascosauz]

Re the reported boost of 4850, yes, I posted screenshots above and do so here to save going hunting for them

I know that this is only PBO being enabled and then Max CPU Boost Override being set to 200MHz (is that the way to do it? Or is it one or the other and not both?), but that could confuse people as they might then expect the max boost to actually be 4850. BTW even before applying the Max CPU Boost Override, my 'CPU OC Max in HW was still showing 4650, which I thought somewhat odd?

So to potentially improve speeds I would need to increase the timings of my Ram from 3200 to hopefully 3600? I say hopefully as I know that Bill Zoid calls the Corsair B die Dominator ram 'rubbish' so I don't know what to expect when/if I get around to that.

It would seem that CCD0 is a pretty good chiplet as it would seem as though 1 of the cores (core 0), 2 of the cores will hit 4.575 and 3 reach 4.550, according to HW.
I noted on that thread on HW site that the coder said that 1usmus's power plan did not cure the problem and I'm not too sure why he came to that conclusion. The workload does seem to be spread fairly evenly across my first four cores

I am wondering just how many more bios updates there will be before they are considered 'fixed', as from various threads that I have read, there does seem to be a number of confused users.

You have to expand the first text column in HWInfo to see the CPPC and scheduler core rankings. Though from what your RM screenshot shows, you're doing pretty well for yourself. My Core 0 is dead last in both rankings, so in applications that won't have their loads shifted off Core 0, I'm at an inherent slight disadvantage on account of losing the lottery.

3200MT/s is the "frequency" (strictly speaking, 1600MHz is) not your timings. 16-18-18-36 are your main timings, and you'll need to figure them and the subtimings out in BIOS and Ryzen Master. 3200 with timings around 14 as opposed to 16 will get you into the range of the other normal B-die kits out there.

"Rubbish" from buildzoid is something you'll have to find out for yourself, as he deals mostly in pretty hardcore overclocking as is evident from the channel name. A lot of Corsair's B-die is poorly-binned B-die, or in other words, the B-die that the more elite companies like G.Skill wouldn't touch with a ten-foot-pole. That plays a part in why the XMPs on Corsair B-die are rated about the same speeds and latencies as other companies' non-Bdie SKUs.

Get those timings down to 14, and it'll be pretty damn respectable. Even at 3200/16, you're doing way, way better than the vast crowds out there running 2133 or 2400 on Ryzen. Just to put things into perspective.

 

[lorry]

You have to expand the first text column in HWInfo to see the CPPC and scheduler core rankings. Though from what your RM screenshot shows, you're doing pretty well for yourself. My Core 0 is dead last in both rankings, so in applications that won't have their loads shifted off Core 0, I'm at an inherent slight disadvantage on account of losing the lottery.

3200MT/s is the "frequency" (strictly speaking, 1600MHz is) not your timings. 16-18-18-36 are your main timings, and you'll need to figure them and the subtimings out in BIOS and Ryzen Master. 3200 with timings around 14 as opposed to 16 will get you into the range of the other normal B-die kits out there.

"Rubbish" from buildzoid is something you'll have to find out for yourself, as he deals mostly in pretty hardcore overclocking as is evident from the channel name. A lot of Corsair's B-die is poorly-binned B-die, or in other words, the B-die that the more elite companies like G.Skill wouldn't touch with a ten-foot-pole. That plays a part in why the XMPs on Corsair B-die are rated about the same speeds and latencies as other companies' non-Bdie SKUs.

Get those timings down to 14, and it'll be pretty damn respectable. Even at 3200/16, you're doing way, way better than the vast crowds out there running 2133 or 2400 on Ryzen. Just to put things into perspective.

is this what you mean? It didn't look as it they would expand at first. You say both ranking? Do you mean Ryzen and Windows rankings? My core 0 is perf #1/2, core 1 perf #1/1, are they the two rankings?


Oh, I'm aware that I either got lucky and I'm not complaining in the slightest. What I am doing is learning about what is going on, and more importantly, why.

The ram was the very first thing that I bought for this rig and at the time I only knew enough to realise that it was 'respectable' and even then it was only two sticks. I then found out about the different dies and found another B die to 'match', considering that it's not on Gigabytes vendors list I'm happy that it even runs at 3200, lol. If I can get the timings down and the speed up to say 3400 I'll be damn pleased, if not it's still way way better than anything I've had. My last build to this was back in the mists of time, an Am486 DX4-100 ! Only had laptops until now, so this was quite a jump in learning what was needed and what worked with what!

 

[Zach_01]

You hit the lottery because you

Re the reported boost of 4850, yes, I posted screenshots above and do so here to save going hunting for them

I know that this is only PBO being enabled and then Max CPU Boost Override being set to 200MHz (is that the way to do it? Or is it one or the other and not both?), but that could confuse people as they might then expect the max boost to actually be 4850. BTW even before applying the Max CPU Boost Override, my 'CPU OC Max in HW was still showing 4650, which I thought somewhat odd?
My bus clock is always on 100MHz, that never changes at all from what I have seen.

So to potentially improve speeds I would need to increase the timings of my Ram from 3200 to hopefully 3600? I say hopefully as I know that Bill Zoid calls the Corsair B die Dominator ram 'rubbish' so I don't know what to expect when/if I get around to that.

It would seem that CCD0 is a pretty good chiplet as it would seem as though 1 of the cores (core 0), 2 of the cores will hit 4.575 and 3 reach 4.550, according to HW.
I noted on that thread on HW site that the coder said that 1usmus's power plan did not cure the problem and I'm not too sure why he came to that conclusion. The workload does seem to be spread fairly evenly across my first four cores

I am wondering just how many more bios updates there will be before they are considered 'fixed', as from various threads that I have read, there does seem to be a number of confused users.

Yes but you dont actually getting speeds anything above 4600... are you?

Like mine. CPU OC Max says 4400 but never saw anything above 4200.

 

[lorry]

You hit the lottery because you

Yes but you dont actually getting speeds anything above 4600... are you?

Like mine. CPU OC Max says 4400 but never saw anything above 4200.

View attachment 138986

No, as I said, the highest I've seen is 4.600 according to HW and I think about 4.400 according to RM.
I know that it won't go to 4.85 (gawd, I wish! lol) but I was wondering why both HW and RM stated that my CPU OC Max was 4850 when it obviously cannot go that high.
Also why when the 'Max CPU Boost Clock Override is left at zero the CPU Max OC is then 4650 and not 4600? Surely that cannot mean that this CPUs' limit is 4650???

Are they just bad reporting by both apps? As in not taking the 200MHz into account, or what?

Question - I know that I am reporting on the 'normal' clock speeds now, but is it normal to have two cores that reach 4.6?
As I now have two at 4.600, two at 4.575 and two at 4.525. (I'm not bragging or anything, but I am asking if that is usual.

 

[tabascosauz]

I know that it won't go to 4.85 (gawd, I wish! lol) but I was wondering why both HW and RM stated that my CPU OC Max was 4850 when it obviously cannot go that high.
Also why when the 'Max CPU Boost Clock Override is left at zero the CPU Max OC is then 4650 and not 4600? Surely that cannot mean that this CPUs' limit is 4650???

Question - I know that I am reporting on the 'normal' clock speeds now, but is it normal to have two cores that reach 4.6?
As I now have two at 4.600, two at 4.575 and two at 4.525. (I'm not bragging or anything, but I am asking if that is usual.

View attachment 139012

It's like the theoretical power draw limits under manual OC - when will you ever reach the limit of 1000W of power draw on anything? It's pointless to fret over, and has no bearing on what you actually get. Better for AMD to set an unattainable ceiling than one that will actually limit performance.

If you're using 1usmus or 1909, the scheduler tries its best to keep lightly threaded loads within the same CCX for latency purposes. For you, a CCX is populated by 3 active cores. Seeing as your two best cores are in the first complex, it's not hard to see why the scheduler would want to keep most of the load in there.

Your Core 0 quality makes it easy, because as far as the scheduler is concerned, there is literally no reason ever to migrate load out of the first CCX unless more than 3 cores are needed, which is probably why you see lower freqs and usage on the other 9 cores. On the other hand, if your chip is like mine with Core 0 and the best core at opposite ends of the silicon, the scheduler will feel a bit torn at times.

I described Corsair as "poorly-binned" B-die, but it's still B-die. Throw a little bit extra DRAM voltage at it and you'll likely be able to comfortably get a little frequency and tighter timings out of it. Ryzen DRAM calculator has everything you need, maybe check the V2 profile instead for a more relaxed set of timings.