Newbie RAM overclocking

[junzhi2002]

A newbie here without any RAM overclocking experience.

There are informations out there saying that tighter timings and higher frequency gives better performance for the RAM. But I am wondering should I overclock my RAM to higher frequency or tighter timings?

Extra question: Can I tighten RAM timings without changing the frequency? Will increasing voltage do the job?

 

System: 

Ryzen 3700X

Asus TUF Gaming X570-Plus

Patriot Viper 16GB (2 x 8GB) DDR4-3200 16-18-18-36  <- converted to USD will be $84.11 sadly it's the cheapest 16GB 3200 kit in my country

[OLOy 32GB (2 x 16GB) DDR4-3200 16-18-18-38]  <- this is the RAM I originally spotted on and I want it XD USD $100 is a deal for me

[Jurrunio]

8 minutes ago, junzhi2002 said:

But I am wondering should I overclock my RAM to higher frequency or tighter timings?

Why not both?

 

8 minutes ago, junzhi2002 said:

Can I tighten RAM timings without changing the frequency?

Yes

 

8 minutes ago, junzhi2002 said:

Will increasing voltage do the job?

it may increase stability, it may not.

[SavageNeo]

8 minutes ago, junzhi2002 said:

A newbie here without any RAM overclocking experience.

There are informations out there saying that tighter timings and higher frequency gives better performance for the RAM. But I am wondering should I overclock my RAM to higher frequency or tighter timings?

Extra question: Can I tighten RAM timings without changing the frequency? Will increasing voltage do the job?

 

System: 

Ryzen 3700X

Asus TUF Gaming X570-Plus

Patriot Viper 16GB (2 x 8GB) DDR4-3200 16-18-18-36  <- converted to USD will be $84.11 sadly it's the cheapest 16GB 3200 kit in my country

[OLOy 32GB (2 x 16GB) DDR4-3200 16-18-18-38]  <- this is the RAM I originally spotted on and I want it XD USD $100 is a deal for me

Use ryzen master to put more mhz clock for your ram. Put everything on auto execpt the memory clock ofc. Try getting to 4000mhz or similar and then manually tighten timings.

The Patriot viper has better timings so get that since 32gb is not needed for gaming. 

[nick name]

2 minutes ago, SavageNeo said:

Use ryzen master to put more mhz clock for your ram. Put everything on auto execpt the memory clock ofc. Try getting to 4000mhz or similar and then manually tighten timings.

The Patriot viper has better timings so get that since 32gb is not needed for gaming. 

I wouldn't recommend using Ryzen Master to overclock anything as it may introduce instability when adjusting the same settings from within BIOS wouldn't.  And I can't tell you exactly what works well in Ryzen Master and what doesn't.  

 

Use the DRAM Calculator to get you started on RAM overclocking.  

https://www.overclock.net/forum/13-amd-general/1640919-new-dram-calculator-ryzena-1-6-2-overclocking-dram-am4-membench-0-8-dram-bench.html

 

And you will likely get more performance out of the box from that 32GB kit, but a 16GB kit will likely overclock better than a 32GB kit.  

[MageTank]

1 minute ago, SavageNeo said:

Use ryzen master to put more mhz clock for your ram. Put everything on auto execpt the memory clock ofc. Try getting to 4000mhz or similar and then manually tighten timings.

The Patriot viper has better timings so get that since 32gb is not needed for gaming. 

I would strongly advise against pushing for 4000mhz on Ryzen for several reasons. The biggest being the infinity fabric (FCLK) would have to run at 2000mhz in that configuration in order to maintain the 1:1 strap. This is extremely unlikely as most people top out around 1900. If you ignore the 1:1 strap, your 4000mhz memory will be slower than 3600 or 3800mhz memory with a 1:1 strap.

 

13 minutes ago, junzhi2002 said:

A newbie here without any RAM overclocking experience.

There are informations out there saying that tighter timings and higher frequency gives better performance for the RAM. But I am wondering should I overclock my RAM to higher frequency or tighter timings?

Extra question: Can I tighten RAM timings without changing the frequency? Will increasing voltage do the job?

 

System: 

Ryzen 3700X

Asus TUF Gaming X570-Plus

Patriot Viper 16GB (2 x 8GB) DDR4-3200 16-18-18-36  <- converted to USD will be $84.11 sadly it's the cheapest 16GB 3200 kit in my country

[OLOy 32GB (2 x 16GB) DDR4-3200 16-18-18-38]  <- this is the RAM I originally spotted on and I want it XD USD $100 is a deal for me

On all consumer platforms, latency is king. Unless your applications are optimized to take advantage of heavy memory bandwidth (AVX being the easiest example), you won't see much of a difference pushing strictly for higher frequencies on memory. That's not to say you shouldn't, as frequency and latency go hand in hand with it comes to calculating your total Round Trip Latency, you should focus more on reducing your latency even if it comes at the cost of actual bandwidth. There is a guide in my signature that goes over the basic principles of memory overclocking. It's not designed for Ryzen, but it should give you an idea of what you should be looking for. As for an actual guide, this might serve you well: https://www.techpowerup.com/review/amd-ryzen-memory-tweaking-overclocking-guide/

 

Understand, this isn't going to be a simple "change a few settings and I'm done" kind of process. This will take a fair bit of time to make sure that your adjustments are not only improving performance, but that they are stable after doing so. Prepare to devote a lot of time to trial and error as well as stress testing each change you make.

[junzhi2002]

8 minutes ago, Jurrunio said:

Why not both?

3600MHz and staying with the CL16 is a sweet spot?

 

9 minutes ago, Jurrunio said:

it may increase stability, it may not.

So...? Only increase it when I face instability issue or can I just push higher voltage just because I am lazy??  

[junzhi2002]

3 minutes ago, MageTank said:

Understand, this isn't going to be a simple "change a few settings and I'm done" kind of process. This will take a fair bit of time to make sure that your adjustments are not only improving performance, but that they are stable after doing so.

Good point there, noted. Buildziod said that running memory test only stress the memory but not the FCLK so should I run another test when I am in Windows?

[MageTank]

5 minutes ago, junzhi2002 said:

Good point there, noted. Buildziod said that running memory test only stress the memory but not the FCLK so should I run another test when I am in Windows?

Correct. While not really a stress test, you can do a few passes of Memtest86 to make sure the ram itself and memory controller agree upon the timing configuration you've dialed in. It won't tell you much about the board signaling disagreeing, but it's a start. Once you start to pound the memory/IO lanes of the board, you'll get an answer as to whether or not your system is actually stable. You can use your own stress tests of choice, but I personally use Prime95 for this, as you can change the FFT sizes to dictate exactly what you want to test, whether it resides entirely in cache (small FFT) or RAM (larger FFTs). 

 

I've heard that others use HCI Memtest with great success, however I myself am not too familiar with it. Friends of mine also recommend ASUS Realbench as a system stress test, which appears to be based off some of the later Linpack libraries. Just be mindful of your cooling solution if you go for a test designed to run hot.

 

EDIT: Going to address your voltage question.

8 minutes ago, junzhi2002 said:

3600MHz and staying with the CL16 is a sweet spot?

 

So...? Only increase it when I face instability issue or can I just push higher voltage just because I am lazy??  

Voltage will help if you are facing instability, but you also need to be mindful of scaling, as well as signal interference. Not all memory kits scale the same when it comes to voltage. Some will clock higher/run tighter after pumping extra voltage through it, while others may show no benefits whatsoever. You only really want to push voltage if you are on the edge of stability. Trying to force stability into something completely unstable by using voltage is likely going to backfire, and lead to further instability. I am not certain if this is true for Ryzen, but on Intel, pushing higher IO (VCCIO) and System Agent (VCCSA) voltages could actually cause instability compared to simply dialing it back or finding a specific sweet spot. Increasing these voltages put too much noise through the board, and could cause signaling issues, leading to instability. I know Ryzen has similar voltages, but I can't recall what they are. Someone more versed in Ryzen may be able to help there.

[junzhi2002]

7 minutes ago, MageTank said:

Voltage will help if you are facing instability, but you also need to be mindful of scaling, as well as signal interference. Not all memory kits scale the same when it comes to voltage. Some will clock higher/run tighter after pumping extra voltage through it, while others may show no benefits whatsoever. You only really want to push voltage if you are on the edge of stability. Trying to force stability into something completely unstable by using voltage is likely going to backfire, and lead to further instability. I am not certain if this is true for Ryzen, but on Intel, pushing higher IO (VCCIO) and System Agent (VCCSA) voltages could actually cause instability compared to simply dialing it back or finding a specific sweet spot. Increasing these voltages put too much noise through the board, and could cause signaling issues, leading to instability. I know Ryzen has similar voltages, but I can't recall what they are. Someone more versed in Ryzen may be able to help there.

So just dial in the numbers and wait for the memory test? And only increase the voltage when needed? I think I get it now  

[MageTank]

1 minute ago, junzhi2002 said:

So just dial in the numbers and wait for the memory test? And only increase the voltage when needed? I think I get it now  

That's definitely a simple view of the process, but yes, that's be essential order. You want to avoid going into Windows until you've performed some basic testing in Memtest86 as it will help minimize the risk of OS corruption when your memory decides to flake out while Windows is doing something important. If you have a spare drive to boot from, it wouldn't hurt to use that and load some stress test applications to avoid any risk to having to reinstall your OS on the main drive.

 

Also be sure to test memory performance in between your adjustments to make sure you are not going backwards in performance. AIDA64 offers a pretty decent testing catalog for memory tests, allowing you to test your bandwidth (Read, Write, Copy) as well as your latency. Again, prioritize latency. If you change a timing, and it reduces bandwidth but also lowers your latency, I would consider that a fair trade. 

[Jurrunio]

1 hour ago, MageTank said:

If you change a timing, and it reduces bandwidth but also lowers your latency

Doesnt improving the timings increase (ever so slightly) bandwidth while cutting latency noticeably?

[MageTank]

1 hour ago, Jurrunio said:

Doesnt improving the timings increase (ever so slightly) bandwidth while cutting latency noticeably?

Depends on the timing. Most primary timings have a negligible impact on bandwidth. The biggest timings that impact bandwidth are the tertiary timings. RDRD, unsurprisingly, can improve your read bandwidth if configured properly. The issue is, if you go too loose or too tight on the timing, and it impacts the way your RTL/IO-L offsets are training (This is for Intel, no idea what AMD calls theirs), you will end up losing bandwidth AND increasing your latency. Lower isn't always the best, it just depends on what the IMC will tolerate without breaking training values elsewhere.

 

The thing to note is that your actual bandwidth can never exceed your peak theoretical bandwidth. For example: DDR4 3600 has a peak theoretical bandwidth of 57,600MB/s in dual channel. No amount of timing adjustments will ever let you exceed that value, however you can improve your bandwidth efficiency to get as close to that value as possible.

 

The biggest metric for judging relative memory performance is overall Round Trip Latency. It is a culmination of every single timing as well as the frequency of the ram. You want to reduce that value as much as possible while maintaining stability. This will involve a balance of both ideal frequency as well as very tight timings. 

 

Another way to improve latency is by prolonging the time your ram remains active before a refresh, as well as increasing the speed at which it refreshes. So increasing TREFI to it's highest value (65535) and reducing tRFC as low as you can get it while maintaining stability, it's a free increase in performance assuming you can keep the DIMM's cool with some airflow.

 

I really need to sit down and work on memory overclocking for Ryzen, I just don't know where to start. I want to make a guide that encompasses all aspects of memory OCing for Ryzen, but their IMC's differ dramatically when comparing Ryzen 1000 series to the current 3000 series.

[Jurrunio]

3 hours ago, MageTank said:

Another way to improve latency is by prolonging the time your ram remains active before a refresh, as well as increasing the speed at which it refreshes. So increasing TREFI to it's highest value (65535) and reducing tRFC as low as you can get it while maintaining stability, it's a free increase in performance assuming you can keep the DIMM's cool with some airflow.

btw I dont think Ryzen lets users adjust TREFI, I can't find a number this large last time.

[MageTank]

30 minutes ago, Jurrunio said:

btw I dont think Ryzen lets users adjust TREFI, I can't find a number this large last time.

Interesting. I know different board vendors list it as different things, such as ASUS referring to it as "DRAM Refresh Interval", but it's odd that they remove the option entirely on Ryzen boards. Curious as to what is controlling the recharge interval for RAM, or if they are following the old rules of 7.8usec. That rule basically takes your ram's memory clock (if running 3200, it would be 1600) and multiply it by 7.8. In this example, that would be 1600 x 7.8 = 12480. I'll get my hands on a Ryzen platform and investigate further. 

[Otto_iii]

7 hours ago, nick name said:

And you will likely get more performance out of the box from that 32GB kit, but a 16GB kit will likely overclock better than a 32GB kit.  

32gb won't do anything at all, even if streaming and gaming, i've never seen more then about 10gb usage while running a game, even with multiple backround tasks and you only get performance degradation if ram is filled right up to or beyond max capacity.  Even at 8gb (2x4) you rarely see a performance improvement going to 16 at same speeds, only in a few newer titles that use slightly more then 8gb including backround processes (win10, steam,discord, chrome etc in backround) does it make a difference. 

32gb is for work not for gaming, and by the time games might actually use slightly more then 16gb we'll probably be on DDR5
 

7 hours ago, junzhi2002 said:

3600MHz and staying with the CL16 is a sweet spot?

 

For overclocking the sweet spot is pairing speed of ram with highest speed you can get out of infinity fabric while stable, so if possible ram tweaked to 3733 speed and 1866 FCLK, or if lucky ram at 3800 and 1900 FCLK.  3600 C16 is the sweet spot in regards to buying affordable ram right now, plug and play no tweaking.

Do you already own the vipers?
if not see what and if they sell this for in your region, look up "F4-3600C16D-16GVKC"

[nick name]

25 minutes ago, Otto_iii said:

32gb won't do anything at all, even if streaming and gaming, i've never seen more then about 10gb usage while running a game, even with multiple backround tasks and you only get performance degradation if ram is filled right up to or beyond max capacity.  Even at 8gb (2x4) you rarely see a performance improvement going to 16 at same speeds, only in a few newer titles that use slightly more then 8gb including backround processes (win10, steam,discord, chrome etc in backround) does it make a difference. 
 

I didn't mention it because of the capacity.  I mentioned it because it's likely dual-rank and that performs a little better than single-rank.  

[MageTank]

16 hours ago, nick name said:

I didn't mention it because of the capacity.  I mentioned it because it's likely dual-rank and that performs a little better than single-rank.  

That is true, however dual rank is also far more difficult to overclock due to the addition of dual rank tertiary timings that come into play. For someone experienced with overclocking ram, dual rank is almost always a better choice. For someone just starting out, I would recommend sticking to single rank and 1DPC. From there, upgrade to 2DPC, then if you still need capacity, move on to the larger dual rank DIMM's.

[nick name]

52 minutes ago, MageTank said:

That is true, however dual rank is also far more difficult to overclock due to the addition of dual rank tertiary timings that come into play. For someone experienced with overclocking ram, dual rank is almost always a better choice. For someone just starting out, I would recommend sticking to single rank and 1DPC. From there, upgrade to 2DPC, then if you still need capacity, move on to the larger dual rank DIMM's.

I also made mention of the difficulties more capacity can introduce when overclocking.  

[junzhi2002]

On 12/10/2019 at 4:49 AM, Otto_iii said:

32gb won't do anything at all, even if streaming and gaming, i've never seen more then about 10gb usage while running a game, even with multiple backround tasks and you only get performance degradation if ram is filled right up to or beyond max capacity.  Even at 8gb (2x4) you rarely see a performance improvement going to 16 at same speeds, only in a few newer titles that use slightly more then 8gb including backround processes (win10, steam,discord, chrome etc in backround) does it make a difference. 

32gb is for work not for gaming, and by the time games might actually use slightly more then 16gb we'll probably be on DDR5

Nice point, but I didn't even mentioned what would I do with my system though...... I wanted to do video editing, 32GB sticks are for "kinda" future proof. At least for 2~3 years from now. Since AM4 will use DDR4 for its life time.

 

On 12/10/2019 at 4:49 AM, Otto_iii said:

For overclocking the sweet spot is pairing speed of ram with highest speed you can get out of infinity fabric while stable, so if possible ram tweaked to 3733 speed and 1866 FCLK, or if lucky ram at 3800 and 1900 FCLK.  3600 C16 is the sweet spot in regards to buying affordable ram right now, plug and play no tweaking.

Do you already own the vipers?
if not see what and if they sell this for in your region, look up "F4-3600C16D-16GVKC"

Actually they are the cheapest B-die kit in my region, but not the cheapest 16GB (2 x 8GB) kit out there, but the kit is USD $2~$10 more expensive from cheapest equivalent 16GB (2 x 8GB) DDR4-3200, so I got it.

 

Edit: "F4-3600C16D-16GVKC" aren't B-die haha but good recommendation! But they are much more expensive then my current Vipers USD $10 difference. (I can return the Vipers and get Ripjaws but it isn't good shopping manners)

 

On 12/10/2019 at 5:15 AM, nick name said:

I didn't mention it because of the capacity.  I mentioned it because it's likely dual-rank and that performs a little better than single-rank.  

Nice point, but who cares about RAM ranks when extreme overclocking aren't involved XD 

 

18 hours ago, MageTank said:

For someone just starting out, I would recommend sticking to single rank and 1DPC. From there, upgrade to 2DPC, then if you still need capacity, move on to the larger dual rank DIMM's.

1DPC mobo is not available in my region, at least for my budget range, they are crazily expensive boards in my region

 

 

Thanks for all of the informations given.  

Edited December 11, 2019 by junzhi2002
Add informations

[MageTank]

5 hours ago, junzhi2002 said:

1DPC mobo is not available in my region, at least for my budget range, they are crazily expensive boards in my region

 

 

Thanks for all of the informations given.  

1DPC doesn't refer to a physical trait of the board, but rather using only a single DIMM in each memory channel. For example: If you have a board with 4 DIMM slots, only occupy 2 of them, with one DIMM in Channel A, the other in Channel B. Every board supports a 1DPC configuration, however in order to support a 2DPC configuration, you would need more than 2 DIMM slots, so most standard ITX boards are out of question.