R5 3600 overclocking and strange ram behavior

[trents]

Woomack, let me check to see if I am understanding you correctly. What I hear you saying is that:

1. There is a 20c temp offset that AMD has built into Ryzen 1 and 2. So are you talking about the difference between Tctl and Tdie? [As you can see from my attached pics, HWInfo64 indicates that there is a 10c offset (Tctl vs. Tdie). My understanding is that different models of Ryzen CPUs accross the generations have different amounts of fan curve offset, ranging from 0c to 20c. I'm running the 2700x.]
2. There is thermal throttling that takes place at around the 75c mark with Ryzen 1 and 2 because the actual temp is 95c due to the temp offset. [Is this why my core speed multiplier drops down from 4.35x to 39.5x under load as can be seen in the pics? Yet, what is curious to me is that HWINfo64 does not report any thermal throttling and it doesn't even when I drive temps up to over 90c. Is this not thermal throttling or is AMD defining thermal throttling in a different way?]

 

[Zerileous]

They don't call it throttling, but if you look into the precision boost 2 algorithm you have 3 things that can limit your clock. Max rated clock speed, power draw (which is again broken down into 3 parts PPT, TDC, EDC), and temperature. If you have an x chipset you can set the limits of the power draw through the roof (for example PTT set to 740W, draw under prime95 small FFT 103W, TDC 480A, draw = 72A, EDC 650A, draw = 136.5A). So the stock limits being 128W PPT, 80A TDC, 125A EDC, only the EDC is exceeded. Disabling the stock limits does give me about 50MHz under the load (still prime95 small FFT). What all those numbers mean is explained here

You can see this in Ryzen Master if you're curious.

In other words if it's not max clock and it's not current, it has to be temperature.

 

[trents]

So if you were to exceed TJMax such that you get throttling, would it be a more drastic lowering of the speed? And if we are talking about the same thing when TJMax is not exceeded, if not throttling, what do we call it?

And why does HWInfo64 not report any "thermal throttling" when that is exactly what seems to be going on?

 

[Woomack]

Woomack, let me check to see if I am understanding you correctly. What I hear you saying is that:

1. There is a 20c temp offset that AMD has built into Ryzen 1 and 2. So are you talking about the difference between Tctl and Tdie? [As you can see from my attached pics, HWInfo64 indicates that there is a 10c offset (Tctl vs. Tdie). My understanding is that different models of Ryzen CPUs accross the generations have different amounts of fan curve offset, ranging from 0c to 20c. I'm running the 2700x.]

Yes, depends on the chip it's between 10 and 25°C. For example, TR 1000 has 25°C.

2. There is thermal throttling that takes place at around the 75c mark with Ryzen 1 and 2 because the actual temp is 95c due to the temp offset. [Is this why my core speed multiplier drops down from 4.35x to 39.5x under load as can be seen in the pics? Yet, what is curious to me is that HWINfo64 does not report any thermal throttling and it doesn't even when I drive temps up to over 90c. Is this not thermal throttling or is AMD defining thermal throttling in a different way?]

Throttling appears when anything exceeds values in the specification like way too high temps or power. At first, I was overusing this term but AMD protected themselves so everything was in the specified range, even using their pretty bad stock coolers. In real it's like when the CPU has rated clock between let's say 3200-4200MHz and because of temps it runs at the average of ~3500 but no less than 3200 then it's no throttling. It's just their "amazing technology".

It's really hard to exceed Tj just because much before that, the CPU already lowers its voltage and clock. On 1000-2000 Ryzen it's much easier to keep optimal temps but on 3000 it's easy to hit 95°C and this is the throttling point for the latest chips.
As long as the CPU is set at auto then no matter what cooler you use but at least AMD stock one (assuming there is any airflow in the case) then the CPU will reach ~94.8°C (the same max temp reading in hwinfo64 on at least 15 different coolers that I tested when the CPU was at auto) but will be stable and won't drop the frequency below its rated value.
What does it mean for the end-user? Only that the CPU won't run at a higher clock and won't ever boost up to its maximum boost clock which is specified by AMD. However, that's a full load scenario and in games or anything else even the AMD cooler can keep Ryzen 3000 at no more than 70-80°C.

So when will you see the throttling on new AMD? When you use fully manual settings and disable power saving options while the cooler won't be good enough to handle these settings. It depends on the motherboard but usually when you set manual CPU clock then it automatically disables some options. When the CPU will be too hot or the power limit will be exceeded then you will see a lower CPU clock than in the specs and then we can call it throttling. Most gaming or overclocking series motherboards have a higher power limit so usually, you won't see a throttling because of that. Maximum temps are usually set to 95°C but on some OC series motherboards you can set it up to 125°C.

 

[trents]

So I gather from what you say then that end users running "Auto" bios settings can expect to settle into a higher clock under extended heavy load if they have superior cooling, particularly with Ryzen 1 and 2. By that I mean still somewhat less than max turbo but skewed more toward that end than users with stock or equivalent cooling. Is that correct?

 

[Woomack]

Pretty much yes. It's the same as with new graphics cards. Even though there is some frequency range and in the specs you see some clock then depends on temps and other factors, the frequency is higher or lower. I don't remember exactly but I think it's 75°C and less for new Ryzens that let to achieve the highest clock. Once the CPU pass it then the clock won't boost to the maximum value. The main problem is that there is no Ryzen 3000 that will run below 75°C under full load like in Prime95 or AIDA65 FPU/CPU tests on popular AIO or air coolers. The best cooler with fully loaded Ryzen 3600/3600X/3700X that I was testing, could keep CPU's temp around 80-82°C. I wasn't testing anything better than Noctua D15 with 2 fans or 240/280 AIO. Maybe on good custom water cooling, it will be slightly better.

One example based on my 3900X as I was testing it in the last days in the modded ITX case.
75°C and less - boost up to 4.55GHz, randomly but it was registered
80-90°C - boost up to 4.25GHz
90-94°C - constant 4GHz (hwinfo64 was showing 39.8 CPU ratio)
95°C+ (not really much above) - 3.8GHz

Every CPU clock drop is also connected with a CPU voltage drop.

So when my 3900X was hitting 95°C then it was matching the frequency which is listed in the specification. At the same time, the CPU's voltage was ~1.15-1.19V.
So when would thermal throttling start in this case? ... I assume that when the CPU would drop the voltage to its minimum possible voltage which is below 1.1V. On my board, it's registering 1.05V as the lowest at which CPU is booting.

Of course, everything above is related to auto CPU settings.

Overclocking has some point and gives better results as long as you can keep low CPU temps.
One more thing to add. In a typical environment, 240 AIO is as good as for example Noctua U12S with 2 fans so expect close to 90°C under full load on 3600/3600X/3700X/3800X. On 3900X it will already reach 95°C. On the other hand, no one uses 100% CPU load daily and even if then not for an extended time.

 

[trents]

So, if I moved my system out to the garage on a cold winter day where ambient temps were say, 7c, I should have a good chance of seeing continuous turbo frequencies while doing a stress test at stock settigns. Correct?

 

[Woomack]

In theory, yes but I doubt it will be constant, more like it will boost much more often to a higher frequency. I guess that guys who were testing Ryzen 3k on a single stage, dry ice or LN2 can say some more but then I doubt they were even checking how these chips act like at stock settings. I would just spend my time testing how high it will OC

Your Ryzen 2700X is rated at lower temps and actually it runs at lower temps so there can be other factors that are affecting a higher boost.
I'm not sure if Ryzen 3000 is so big step forward compared to 2000. Performance, in theory, is higher on 3000 series but in real they boost to about the same clock during a typical work and heat up much more. So the main advantage is a chance to have 12-16 cores on this "cheap" platform.

 

[trents]

But isn't the IPC performance better on the 3000 series. From my research, the 3700x gives about 13% higher CB R20 scores than the 2700x at stock.

 

[Kenrou]

But isn't the IPC performance better on the 3000 series. From my research, the 3700x gives about 13% higher CB R20 scores than the 2700x at stock.

Higher IPC but what shines is the much higher memory clocks which gives higher FPS. I'm running 3800mhz synced with infinity circuit with no issues and I'm sure I could push it higher if there were any meaningful gains with my setup. Maybe if I ever get my hands on a 2080ti or next gen equivalent...

 

[Zerileous]

I'm pretty sure memory clocks has nil impact on FPS. As in < 1% from 3200MHz to 3600MHz. Probably even less at 3800MHz. I can't find the official AMD graphic but this real world testing from TPU should clear things up: https://www.techpowerup.com/review/amd-zen-2-memory-performance-scaling-benchmark/4.html.

Someone did test the stock behavior of Zen2 on LN2 and did find increased boosting, but using an all core load. https://www.gamersnexus.net/news-pc/3492-ryzen-cpu-thermals-matter-coolers-and-cases

 

[Woomack]

In all my tests based on games, the difference between RAM settings is like 1-3% max where that 3% is between something like 2666 vs 3600 and 1% is like 3200 vs 3600. Timings make 0-1% difference and not always tighter timings give a higher FPS.

Ryzen 3000 is slightly faster but I wouldn't say it's 13% overall. Maybe in some specific benchmarks but it also depends on CPU's boost clocks and other things. The difference can be up to ~10% average between Ryzen 1000 and 2000.

In Cinebench R20 you can see 200-300 higher scores because of power delivery, LLC and CPU voltage settings. I'm able to drop 20W from the CPU with additional settings and the CPU clock looks exactly the same but somehow Cinebench R20 scores 200-300 points worse. So besides the CPU clock itself, there is something more. I assume the same are acting some other benchmarks. I haven't seen the difference in benchmarks built-in popular games.

One more thing about memory. Depends on the CPU and its chiplet configuration, higher or lower capacity or frequency is acting a bit different. On 3900X/3950X there is some more point to set 1:2 IF ratio and a higher memory frequency. Also, with higher capacity, the performance is scaling better.
Here is one example:
On 3600/3700X I had a similar performance at:
- 3600 16-16-16 IF 1:1 / 2x8GB
- 4800 18-24-24 IF 1:2 / 2x8GB
2x16GB kit was already slower at DDR4-4266 CL18-24-24

On 3900X I had a similar performance at:
- 3600 16-16-16 IF 1:1 / 2x8GB
- 4800 18-24-24 IF 1:2 / 2x8GB
- 4266 18-24-24 IF 1:2 / 2x16GB
- 4133 18-23-18 IF 1:2 / 2x32GB <-- this is my 24/7 setting right now

This is out of synthetic bandwidth benchmarks but based on games and PCMark 10 scores.
It also explains why on ROG series motherboards there are competitive overclocking profiles at DDR4-4133 with IF 1:2.
On Ryzen 3000 there are simply too many variables to say straight what is better and in theory, small specification differences may affect results more than we expect.

 

[Kenrou]

The "something more" is the infinity fabric (fclock), and how it interacts with the memory speed/timings. Ofc it only shows in games that are not GPU bottlenecked but it can be 5%-10% improvement on top of clock speed and IPC. I think it was Linus that said in one of he's videos that Cinebench shows no difference whatsoever to memory scaling but it shows in games.

 

[Zerileous]

How many people are running CPU bound games and caring about the performance on Zen2? I will give you the TPU article I linked shows a 5%-10% improvement in Wolfenstein II at 720p (from 2400MHz to 4000MHz), with a difference from 350fps to 400fps. In this case yes FPS increased by a measurable amount but you'll never no the difference. Note that even at 1080p this game has dried up to a difference of 3% and is still making 300fps. What game is going to be CPU bottlenecked at FPS that matter? (DCS is one, but I have a hard time thinking that's a typical use case). My point is it's not appropriate to generalize that faster memory gives more FPS in this platform. It can in certain uncommon circumstances, but it's the exception not the rule.

 

[Kenrou]

I agree up to a point, but I thought the point of this forum was “yay i/it can” instead of “meh i/it can sometimes” ? Gamers Nexus, LTT, Hardware Unboxed and even der8auer at one point or another recommended 3600mhz memory instead of 3200mhz for a reason, and the reason is higher FPS no ? Am I missing something ?

 

[Zerileous]

Even AMD recommends it, that's probably where they're getting it from. I would also recommend it if I was picking components for a new build. Sure this is about being the best, for some, for others its about getting the most out of what you have / can afford. Nobody should be unsatisfied gaming at 3200MHz, which is spec. Maybe a little bummed out that they don't get the numbers they wanted, but it's not going to ruin your rig.

Really what I object to is your claim,

Higher IPC but what shines is the much higher memory clocks which gives higher FPS. I'm running 3800mhz synced with infinity circuit with no issues and I'm sure I could push it higher if there were any meaningful gains with my setup. Maybe if I ever get my hands on a 2080ti or next gen equivalent...

You basically stated that memory speed has more to do with performance than IPC. This is in no way true. I understand you're hyped to by running 3800MHz and congrats, that's pretty cool and something I would be thrilled about as well. But let's not pretend that it matters as much as the IPC improvements. Yes it can also run the infinity fabric faster, which also helps, but again the impact is not anywhere near the same.

Don't believe me? Pick any test, we'll both set our CPUs to 4.0GHz, memory to 3200MHz CL16 and I guarantee you're going to come out ahead. Then let me set my memory to 3600MHz and see if I can make up the difference (I won't).

 

[trents]

For me, the issue is whether I'm building a new Ryzen 2 system from scratch or simply upgrading a Ryzen or Ryzen+ system with a Ryzen 2 CPU. If starting from scratch, I would definitely go for 3600 mhz RAM since it's just a little more expensive than 3200 mhz. But if I'm just upgrading, I would not likely spend the money to move from 3200 mhz RAM to 3600 mhz RAM since any performance gain, if there is any, would be small.

 

[Woomack]

The "something more" is the infinity fabric (fclock), and how it interacts with the memory speed/timings. Ofc it only shows in games that are not GPU bottlenecked but it can be 5%-10% improvement on top of clock speed and IPC. I think it was Linus that said in one of he's videos that Cinebench shows no difference whatsoever to memory scaling but it shows in games.

I'm usually testing games at 1440p or higher and then there is 0-1% difference because of memory speed. However, I usually test memory kits at 3000+ on Ryzen so there can be that +1% if I would add 2400/2666 memory to the comparison.
As it was said, 5-10% you get in games at 720p/1080p and lower details where the CPU does much more work and FPS are already high.

I agree up to a point, but I thought the point of this forum was “yay i/it can” instead of “meh i/it can sometimes” ? Gamers Nexus, LTT, Hardware Unboxed and even der8auer at one point or another recommended 3600mhz memory instead of 3200mhz for a reason, and the reason is higher FPS no ? Am I missing something ?

3600 RAM is generally recommended because of many things. It's a maximum clock at which you can set 1:1 IF ratio on pretty much every Ryzen 3000 (excluding APU). Some chips go up to 3800, some not, but every CPU should run with memory at 3600 and that was confirmed by AMD, even though they guarantee a maximum of 3200.
3600 is also really easy to set on current IC as every modern IC runs at 3600, regardless if you use 8 or 16GB modules. 32GB modules are also designed to reach 3600.
Another thing is that 3600 memory costs as much as 3200, and often even less. It's because of the mentioned IC which easily runs at 3600 CL18 and can be exchanged within a product number to every brand like Micron, Samsung, Hynix, Spectek, simply everything.

So it's clear that 3600 memory is recommended for Ryzen but it doesn't change the fact that if someone has 3200 then won't see any significant difference (out of synthetic benchmarks) if upgrade RAM to 3600.

I'm trying not to read or watch any memory reviews and comparisons in the web as most are pretty bad and most reviewers have no idea how to test or compare memory. Still, all will tell you that for Ryzen, 3600 memory is optimal but some will add some BS that you need 3600 CL14 or it will be slow. This is totally not true as 3600 CL14 and CL18 performs almost the same. It was already covered by some websites like TPU.

Enthusiasts see these things in a different way so a lot depends on who is the end-user. I wouldn't recommend memory at high clocks or really tight timings to a regular gamer or someone who needs a PC for daily work. I would recommend it to an enthusiast or overclocker but this is a really low % of all users. Most of those who are on enthusiast/overclockers forums are nothing but a regular gamers.
I run memory like 2x8GB at 4800-5000 or 2x32GB at 4133-4200 just for fun and because I can. It doesn't mean it's the best option for anyone else.

 

[ManofGod1000]

Just saying, I have moved things around so that I now use the 3600 in my AB350 Gaming mITX board with the 32GB of 3200 speed ram. I left my 1600 running at 4 GHz at 16GB of 2800 speed ram because although it is a good board, it is not a good board for the Ryzen 3000 series processors when compared to my other two boards.

 

[ManofGod1000]

I figured I would insert my continued experience here, if someone is looking for a possible solution. Well, I upgraded the computer I had the Asrock B450 Fatality Gaming mITX mainboard to a 3700X from a 1600. The thing is stupid fast in no small measure but, even with the 2800 GSkill ram I have installed in the machine, a couple of days later (Today) I got one Memory Management BSOD. I have since bumped the vram to 1.25v from 1.20v and am going to see if that helps in the long run. Fact is, the problem could be the mainboard but, I think it is the firmware that is the problem on that board.