RAM upgrade importance withe regard to Multi-Threading

[LucShep]

3 hours ago, Bucic said:

New relevant info?

https://www.reddit.com/r/MSI_Gaming/comments/uzissc/does_ram_xmp_profile_issue_still_persist_on_the/

That is exectly what I have on my to-do list. Leave 1.4 SET voltage, bring back the frequency to 3600 and tighten the timings. Some day...

 

EDIT:
shiiiiiii@#$@#

https://www.reddit.com/r/intel/comments/sluk4z/psa_12th_gen_non_k_skus_have_locked_vccsa_voltages/

 

 

OMG.... The info I had was that VCCSA settings were locked for the user. I thought it would automatically re-adjust for XMP profile!
I'm am so, so sorry! 

In this case I need to take back what I said about your motherboard (and any other B660 mobo, that is), because it seems to be doing what it's meant by design.
Now it's explained why your process has been difficult! 

Leaving 1,40v on DRAM voltage, bringing back the frequency to 3600 and tighten the timings is a good idea, but if it actually allows you to use 3700 MHz 18-22-22-42  stable and without any issues (on stock/locked VCCSA) then there's nothing to complain about it.

Still, have to say, fully locked VCCSA settings are a very dumb decision by Intel and mobo manufacturers. B660 was supposed to allow for memory overclocking!

Edited July 15, 2023 by LucShep

[LucShep]

BTW, here's a DDR4 OC Guide:  https://github.com/integralfx/MemTestHelper/blob/oc-guide/DDR4 OC Guide.md#table-of-contents
Lots of interesting stuff, also for sub-timings.
If interested, please read it thoroughly, and do not feel tempted to test settings that go beyond the mentioned "safe" settings in it. 

Edited July 15, 2023 by LucShep

[Bucic]

35 minutes ago, LucShep said:

OMG.... The info I had was that VCCSA settings were locked for the user. I thought it would automatically re-adjust for XMP profile!
I'm am so, so sorry! 

In this case I need to take back what I said about your motherboard (and any other B660 mobo, that is), because it seems to be doing what it's meant by design.
Now it's explained why your process has been difficult! 

Leaving 1,40v on DRAM voltage, bringing back the frequency to 3600 and tighten the timings is a good idea, but if it actually allows you to use 3700 MHz 18-22-22-42  stable and without any issues (on stock/locked VCCSA) then there's nothing to complain about it.

Still, have to say, fully locked VCCSA settings are a very dumb decision by Intel and mobo manufacturers. B660 was supposed to allow for memory overclocking!

 

Well, I'll be damned! I would have never expected this outcome. Look at the comparison between the tight 3700 vs even tighter 3600!

 

29 minutes ago, LucShep said:

BTW, here's a DDR4 OC Guide:  https://github.com/integralfx/MemTestHelper/blob/oc-guide/DDR4 OC Guide.md#table-of-contents
Lots of interesting stuff, also for sub-timings.
If interested, please read it thoroughly, and do not feel tempted to test settings that go beyond the mentioned "safe" settings in it. 

 

Thanks, but after that last benchmark if it wouldn't be the pinnacle of diminishing returns I don't know what would have been

There's nothing to apologize on your part, Luc. I started in the dark and I'm leaving with lots of knowledge. Besides, every step of the way before buying the mobo I kept reading that it's very specifically NOT for OC purposes. They (MSI!) should have been a tad more specific! Is "3600 maximum" too much of screen ink or what? They even mislead by stating 4000 here and there.

[Bucic]

Now... any tips for those who don't see any performance improvement with MT vs ST DCS?

[LucShep]

55 minutes ago, Bucic said:

Well, I'll be damned! I would have never expected this outcome. Look at the comparison between the tight 3700 vs even tighter 3600!

Yep!  

So, you already understand that the 3600 (Mhz) listed there is about the speed of the memory.
But the other four main timings also represent important aspects, as you now notice, with the new 3600 17-19-19-39 mem settings.

So, in that "17-19-19-39" order....

CL: CAS Latency. The time it takes between a command having been sent to the memory and when it begins to reply to it. It is the time it takes between the processor asking for some data from the memory and then returning it.
 
tRCD: RAS to CAS Delay. The time it takes between the activation of the line (RAS) and the column (CAS) where the data are stored in the matrix.
 
tRP: RAS Precharge. The time it takes between disabling the access to a line of data and the beginning of the access to another line of data.
 
tRAS: Active to Precharge Delay. How long the memory has to wait until the next access to the memory can be initiated.

As you can realize, the shorter these timings are, the better. And it's why 3600 17-19-19-39 is a bit better than 3700 18-22-22-42.

Ideally, you want the highest speed AND shortest timings, but that does not come cheap (the DDR4 "B-Die" memory prices!) or easy, if you're into RAM overclocking.
In reality, it's not like small differences can be perceived (most times they're not) but there can be measurable differences in some aplications, as you now notice in DCS.

Edited July 15, 2023 by LucShep

[Bucic]

@LucShep Now an interesting conclusion with regard to CineBench R23. Compare the DCS results with the following CineBench results:

8240 RAMGSK3700@1.4V 18-22-22-42 701

7917 RAMGSK3600@1.4V 17-19-19-39 701

Conclusion: If you're tuning your memory for DCS, CineBench is the last piece of software you'd want to use for benchmarking

Edited July 15, 2023 by Bucic

[LucShep]

4 hours ago, Bucic said:

@LucShep Now an interesting conclusion with regard to CineBench R23. Compare the DCS results with the following CineBench results:

8240 RAMGSK3700@1.4V 18-22-22-42 701

7917 RAMGSK3600@1.4V 17-19-19-39 701

Conclusion: If you're tuning your memory for DCS, CineBench is the last piece of software you'd want to use for benchmarking

 

Yes, Cinebench is more focused on CPU performance, you're not getting the clear picture from RAM in it.

For RAM, I'd say AIDA64 Memory (paid version) and MaxxMEM2 are better to benchmark that.

[Hiob]

11 hours ago, LucShep said:

Yes, Cinebench is more focused on CPU performance, you're not getting the clear picture from RAM in it.

For RAM, I'd say AIDA64 Memory (paid version) and MaxxMEM2 are better to benchmark that.

The only problem with that is, that the results don't translate well to real world application. For just chasing high scores for a particular part of the PC, ok, but when your ultimate goal is to get the best gaming performance the only thing that counts is, how well the ram leverages the power of the cpu or in other words how well they perform together.

So, I don't say that those benchmarks are unfit to measure RAM performance, but that the results are rather meaningless imho.

[kksnowbear]

4 hours ago, Hiob said:

The only problem with that is, that the results don't translate well to real world application. For just chasing high scores for a particular part of the PC, ok, but when your ultimate goal is to get the best gaming performance the only thing that counts is, how well the ram leverages the power of the cpu or in other words how well they perform together.

So, I don't say that those benchmarks are unfit to measure RAM performance, but that the results are rather meaningless imho.

Of course, you are correct about the difference in benchmarks and 'real-world' performance, especially in any specific application/game.  I agree completely.

However, there are many instances where benchmarking is useful - at least two of which are:

1. A case where someone doesn't already have the hardware in question, and is looking to determine how best their budget is applied.  In this case, you might want to know how a given upgrade/hardware purchase will affect performance generally, as a measure of how it might affect performance specific to one app/game.  An example might be "RAM kit X seems to increase performance by (measurement)...but for 10% less cost, RAM kit Y seems to provide the same performance increase".

2. Where comparing system A to system B, either when they are same/similar hardware or even different, but looking at effect of exact same hardware upgrade on each/both.  As you say, it won't necessarily tell you what any given specific app will do on either machine being compared...but it will help establish the overall impact of the same change, comparatively, to both systems.  One example might be "Changing a 3090 to a 4090 on a VR system is worth the cost, but may be less of a 'value' if only running 1440 on a conventional monitor".

Naturally this doesn't apply to everyone, because many (most?) people don't have more than one setup to test anyway...but it does apply when considering/comparing one's setup to others, or if a person has access to more than one set of hardware.  I run a shop and have 20-30 setups at any one point that I can try various things with, and this is very helpful to me in making changes and/or recommendations to my own system, and others' (which I do a lot of).

So benchmarks in general, while not necessarily translating directly to 'real world' performance, are far from being meaningless.  If understood and used properly, they can be a very effective tool in an overall approach to system design/building/maintenance/upgrading.  And they're not just academic, depending on the circumstances.

Because I do a lot of this type work, I keep very extensive records (>1,800 test cases), all involving synthetic benchmark scores.  Without this data, I'd have much less in the way of tools to compare/consider options, and effectively advise others as to their options.

Edited July 16, 2023 by kksnowbear

[Hiob]

I don't meant that benchmarking is meaningless - I do it all the time. I meant that in the particular case of RAM, I'd rather test the system performance instead of the isolated RAM performance. Because I think (but I have no proof or a desire to start a fight over this) that other than CPU or GPU performance benchmarks, specialized RAM benchmarks barely translate to real world (gaming-) performance at all. 

Again - I didn't mean to kick off an argument about those benchmarks. I just wanted to give a heads up that those numbers doesn't neccessarily translate well into gaming performance.

But in the worst case, they are still interesting and can help to optimize timings.  

[kksnowbear]

42 minutes ago, Hiob said:

I don't meant that benchmarking is meaningless - I do it all the time. I meant that in the particular case of RAM, I'd rather test the system performance instead of the isolated RAM performance. Because I think (but I have no proof or a desire to start a fight over this) that other than CPU or GPU performance benchmarks, specialized RAM benchmarks barely translate to real world (gaming-) performance at all. 

Again - I didn't mean to kick off an argument about those benchmarks. I just wanted to give a heads up that those numbers doesn't neccessarily translate well into gaming performance.

But in the worst case, they are still interesting and can help to optimize timings.  

We agree completely Sorry if it seemed I was arguing - I guess in a way it *is* an "argument" - but not at all negative

I believe you are perfectly correct in that RAM particularly will have less of an overall impact than other major components like CPU/GPU.  The thing is, precisely because of your point, the impact of RAM on performance is very difficult to test/quantify in a system's overall game performance; it seems to get 'buried' within the whole.  So the benchmarks, especially those specific to memory testing, can help 'bring out' the difference in performance that is due to RAM itself, instead of it being lost in all the other factors.

[Bucic]

Guys, in case of Cinebench Vs DCS with regard to RAM Cine ench is outright misleading. Remembering that Cinebench score increases nicely in line with CPU OC.

All of this brings the "test as you fly" rule in aviation.

I've got a tastier case for you. With CPU OC, the dreaded Bulldozer AMD FX-6300, stock 3.5GHz. With "stress testing" programs I couldn't get the frequency to stable 4.1GHz, maybe even lower. The system was intended mainly for the damn ArmA3. Yeah ... So I applied the abovementioned rule and tested each OC step only in ArmA3. The result was I got to 4.5-4.6GHz! I used 4.4-4.5 for all future games. Now I know why some people call those "stress testing" programs *heat viruses*.

[Hiob]

11 minutes ago, Bucic said:

Guys, in case of Cinebench Vs DCS with regard to RAM Cine ench is outright misleading. Remembering that Cinebench score increases nicely in line with CPU OC.

All of this brings the "test as you fly" rule in aviation.

I've got a tastier case for you. With CPU OC, the dreaded Bulldozer AMD FX-6300, stock 3.5GHz. With "stress testing" programs I couldn't get the frequency to stable 4.1GHz, maybe even lower. The system was intended mainly for the damn ArmA3. Yeah ... So I applied the abovementioned rule and tested each OC step only in ArmA3. The result was I got to 4.5-4.6GHz! I used 4.4-4.5 for all future games. Now I know why some people call those "stress testing" programs *heat viruses*.

That is because the overall stress on the cpu depends on the command set. When you really want to stress your cpu, you take something like prime95 avx, small fft‘s…. but real life applications are rarely as demanding. Then the achievable clock speed depends on the core count that is utilized and so on…

For this reason I like using Heaven Benchmark when I test for DCS. It is DirectX11 like DCs and you can run it windowed whilst tuning afterburner (e.g.) live.

But benchmarking and overclocking (including ram) is a rabbit hole and a time consuming hobby on its own…. add water cooling to the mix and it is even as expensive as any other hobby….

[BitMaster]

I wouldn't run an OC that I know will likely fail under full load.

It must run stable under all conditions, anything else is a big risc for your data integrity over time.

[Hiob]

On 7/17/2023 at 12:39 AM, BitMaster said:

I wouldn't run an OC that I know will likely fail under full load.

It must run stable under all conditions, anything else is a big risc for your data integrity over time.

When it fails under load, it isn't a stable overclock. Also, (over-)overclocking* a productive system you need to do actual work on or store important data, shouldn't be overclocked at all (and doesn't need to be in the first place). 

*You're aware that applying an XMP-Profile aka using the advertised speed on a ram kit IS in fact overclocking? Also modern CPUs have auto-overclocking features out of the box. So where do you draw the line?

Overclocking sounds scary to many, where it is in fact just finding and using the individual headroom, that is otherwise lost and wasted to the loose specifications that come from mass production. When Intel pre-selects silicon for its KS series e.g. they do exactly the same (in a different way of course) - so any top tier CPU is in fact overclocked compared to its average brother.

The really dangerous fields, where you apply way to much power and use impractical cooling solutions like chillers or even liquid nitrogen or similar is of course a completely different story. No one in their right mind would use such a system for important data.

Edited July 18, 2023 by Hiob

[BitMaster]

These days I wouldn't necessarily call it overclocking cause it is more the opposite of what it used to be. Overvolting became undervolting and thus everything changed in regards of CPU and GPU "tuning".  RAM overclocking  is still the old fashioned way ala "give more juice and see how far it goes" kind of thing.

So when I look at my own Ryzen, it's undervolted with an -26 Offset and that is all I did for the CPU, I wouldn't call this overclocking but "tuning this specific die to it's (lower) native volts".

I have watched and followed many tips of how to squeeze some more MHz out of it but it is not worth the time and possible gains. Likely this is the same with many new CPU's.

Keep them as cool as possible, may tune the volts down and see how much more it can deliver by itself. Actually, a blessing to do it that way.

 

For my RAM, yeah..that's a tad different. I wanted to play with them, I wanted 3600 and the lowest CL that runs stable no matter what I throw at it and I achieved it, easily too.

 

Regarding data, I did not mean scientific data, no one should do that work on an overclocked DCS rig...hell no.

What I mean is corrupting your NTFS, crippling Windows and forcing a reinstall or at least a "sfc /scannow" followed by some "DISM" commands if you are lucky.

Especially when you do RAM testing, DO NOT use your usual Win install, it will become corrupted if you walk at the edge of BSOD etc.., use a bootable Linux and memtest etc...   if it's stable there, you may try your windows.... just keep an eye open for NTFS corruption if you still BSOD out of the blue.

Edited July 18, 2023 by BitMaster

[diego999]

59 minutes ago, BitMaster said:

Especially when you do RAM testing, DO NOT use your usual Win install, it will become corrupted if you walk at the edge of BSOD etc.., use a bootable Linux and memtest etc...   if it's stable there, you may try your windows.... just keep an eye open for NTFS corruption if you still BSOD out of the blue.

 

The only blue screen in 6 years using windows 10 was when I got too greedy with memory latency.

I didn't break anything, but the lesson was learned.