Best chance to get 3200 Mhz ram or more with 4X8 GB kit

[SexyTerrorisT]

Hello to fellow rog enthusiasts.

For TL;DR (Too long ; Don't read) jump to last paragraph. Thanks in advance

Seeing how ram bandwidth really drives performance on some games and other latency sensitive application i was thinking what would be the optimal decision to make regarding ram purchase.

I have currently ordered a Gskill F4-3200C14Q-32GTZR ram which as its name suggest is a 4X8 GB RGB kit from gskill rated for 3200 Mhz for intel platform.
I know AMD does not support officially support that kind of speed for 4 dimm single ranked modules but a lot seem to have managed to reach those speed with various BIOS by overclocking.

Demand for kits at 3200 Mhz has been very high and its very hard to find them is stock. I am considering dropping the ram i currently have ordered and getting higher rated kits at 3600 Mhz or even 3866 Mhz (for intel platforms) which are in stock

Instead of waiting for the F4-3200C14Q-32GTZR
https://www.gskill.com/en/product/f4-3200c14q-32gtzr

to arrive I am specifically eyeing the following (yea i am in love with the looks for this line of RBG) since those even though more expensive are in stock:
- F4-3600C17Q-32GTZR
https://www.gskill.com/en/product/f4-3600c17q-32gtzr
- F4-3600C16Q-32GTZR
https://www.gskill.com/en/product/f4-3600c16q-32gtzr
- F4-3866C18Q-32GTZR
https://www.gskill.com/en/product/f4-3866c18q-32gtzr

From what i gathered all those kits must be Samsung B die chips inside. Best memory manufacturinng process or ram atm. Is that assumption correct? Any chance any of those have the undredly hynix ones?

I know getting above 3200 mhz is pretty much a wet dream now and probably in the near future with a Ryzen 7 1800X and a 370X chipset but my train of thought is that the next iteration of ryzen should also run on the crosshair VI hero and most likely will be made on a better and more mature manufacturing process. This process then should support better speeds.

Some reading of the crosshair VI hero overclocking thread of overclock.net i found some intresting data.
http://www.overclock.net/t/1624603/rog-crosshair-vi-overclocking-thread

Both the ram i ordered F4-3200C14Q-32GTZR and the F4-3866C18Q-32GTZR have managed to get 3200 Mhz stable with 32 GB.
The F4-3866C18Q-32GTZR has managed to do so with slight higher voltage. Then again this might be due to varition of the IC in memory controller

DIMM Part number DRAM IC DRAM Configuration BCLK Frequency DRAM Ratio DRAM Frequency Timings Voltage CPU SOC Voltage BIOS version
F4-3200C14Q-32GTZR Samsung B 4x8GB 100 3200 3200 MHz 18-16-16-38-75-2T 1.35 1.1 Bios:0083
F4-3866C18Q-32GTZR Samsung B 4x8GB 100 3200 3200 MHz 18-16-16-38-2T 1.4 1.15 0902

My understanding is the 3866 Mhz kits since is better ic ram can run at 3200 14-14-14-36-2T but the F4-3200C14Q-32GTZR most likely wont be able to reach 3866 at the more loose timing of 18-19-19-39-2N unless you are really lucky. Ryzen 7 seems to care more about speed than timings.

Should i get the more expensive but faster in stock F4-3866C18Q-32GTZR kit or wait for the out of stock F4-3200C14Q-32GTZR but cheaper kit to increases chances of getting 3200Mhz with 32GB(regarless platform considered)? Maybe the F4-3600C17Q-32GTZR or F4-3600C16Q-32GTZR.

Thanks in advance.

[BaneSilvermoon]

On a related note, I've been benchmarking RAM speed results all day today and have a bunch of numbers collected. Might start a thread with it. But there's definitely the potential for large performance increase from memory speed bumps. I'm running a 4x8 2666 kit at 3200 right now, with very, VERY loose timings. And all of my test results are pretty much equal or better, despite going from 16,17,17,36 to 21,22,22,44.

I was contemplating possibly picking up higher speed memory, but now I'm starting to wonder if I should even bother. Still need to do full stability stress testing though.

https://browser.geekbench.com/v4/cpu/compare/2562829?baseline=2561881

[SexyTerrorisT]

BaneSilvermoon wrote:
On a related note, I've been benchmarking RAM speed results all day today and have a bunch of numbers collected. Might start a thread with it. But there's definitely the potential for large performance increase from memory speed bumps. I'm running a 4x8 2666 kit at 3200 right now, with very, VERY loose timings. And all of my test results are pretty much equal or better, despite going from 16,17,17,36 to 21,22,22,44.

I was contemplating possibly picking up higher speed memory, but now I'm starting to wonder if I should even bother. Still need to do full stability stress testing though.

https://browser.geekbench.com/v4/cpu/compare/2562829?baseline=2561881

You are indeed confirming what independent tester and even AMD admitted. I ma however perplexed by your link? 3600 _2666 3600_3200? Running at 3600 Mhz with a 2666 strap vs a 3200 strap?

[BaneSilvermoon]

The 3600 is CPU, It's all stock speeds. Had to bump the core voltage a bit to get the RAM to post consistently at 3200 though. Was reading up on the Ryzen Infinity Fabric a lot last night, and with that it makes sense that my under-volted processor might not like increased RAM speeds.
The infinity fabric stuff is what got me wanting to test this though.

This is the kit I'm using. 4 sticks.
https://www.newegg.com/Product/Product.aspx?Item=N82E16820226836

2666 @1.25v 16,17,17,36
3200 @1.35v 20,21,21,44

For what it's worth, it'll survive numerous benchmarks and 20 minutes of Prime, but it's definitely not stable. Locks up Ghost Recon Wildlands after a few minutes, and the longest memtest I managed was 7 minutes. I've gone back to stock settings, but it's got me thinking about buying some 3200 memory now.

[SexyTerrorisT]

BaneSilvermoon wrote:
The 3600 is CPU, It's all stock speeds. Had to bump the core voltage a bit to get the RAM to post consistently at 3200 though. Was reading up on the Ryzen Infinity Fabric a lot last night, and with that it makes sense that my under-volted processor might not like increased RAM speeds.
The infinity fabric stuff is what got me wanting to test this though.

This is the kit I'm using. 4 sticks.
https://www.newegg.com/Product/Product.aspx?Item=N82E16820226836

2666 @1.25v 16,17,17,36
3200 @1.35v 20,21,21,44

For what it's worth, it'll survive numerous benchmarks and 20 minutes of Prime, but it's definitely not stable. Locks up Ghost Recon Wildlands after a few minutes, and the longest memtest I managed was 7 minutes. I've gone back to stock settings, but it's got me thinking about buying some 3200 memory now.

Interesting i nice gain all across besides Face recognition and latency of course
2666 cas 16 would be: 6 nanoseconds
3200 cas 20 woulf be: 6.25 nanoseconds
The calculation is a bit more evolved than that only considered Cas
A slight hit on system responsiveness for some bandwidth gain

20 mins of prime is not good enough. 24 houes mimum is needed 😛

Prime95 will barely test all features of modern cpu and does not heat it that much or hard compared to intel burn test.
Try using the bundled asus tweaker software for stability. Has a nice range of how hard you want to test it.

I know that i wont be satisfied with my cpu and ram oc until everything run perfect for 48 hours straight. The errors that will crop up without doing so are going to haunt you. At least me 😛 OS system corruption , files corruption, sofwtare crashes and the good old BSOD 😛

[BaneSilvermoon]

Yeah I was just letting it run stress tests when I stepped away to do other stuff to see if any of them caused instant failures. I'm fairly certain it was never even remotely stable, but it was nice to see real world gains on my own system, rather than just looking at youtube benchmarks, in almost everything I tested across the board. Cinebench is the only thing I used that didn't show any improvement. Ghost Recon Wildlands and For honor benchmarks, Geekbench, and 3DMark Timespy and Firestrike tests all showed, at least minor, consistent gains. Often all the way around, from physics tests to video frame rates.

What really interested me though was that a few of those tests had saved results from when I was playing with overclocking my processor, and had runs saved at 4ghz. Comparing against those, the 4ghz CPU, 2666 DRAM results were generally nearly identical to the 3.6ghz CPU, 3200 DRAM runs.

https://browser.geekbench.com/v4/cpu/compare/2563526?baseline=2522789
Was enough to convince me that I chose the wrong RAM getting 2666, and I just ordered new kits.

[Theagg]

SexyTerrorisT wrote:
Interesting i nice gain all across besides Face recognition and latency of course
2666 cas 16 would be: 6 nanoseconds
3200 cas 20 woulf be: 6.25 nanoseconds
The calculation is a bit more evolved than that only considered Cas
A slight hit on system responsiveness for some bandwidth gain

20 mins of prime is not good enough. 24 houes mimum is needed 😛

Prime95 will barely test all features of modern cpu and does not heat it that much or hard compared to intel burn test.
Try using the bundled asus tweaker software for stability. Has a nice range of how hard you want to test it.

I know that i wont be satisfied with my cpu and ram oc until everything run perfect for 48 hours straight. The errors that will crop up without doing so are going to haunt you. At least me 😛 OS system corruption , files corruption, sofwtare crashes and the good old BSOD 😛

When it comes to stress tests, the number of 'required' hours often has me wondering, why that many hours ?

So, why is a minimum of 24 hours of Prime95 needed ? What does 24 hours tell you that say 16 hours doesn't ? And would 48 hours tell you something that 24 hours doesn't.

Heck and if 48 hours straight is required to 'prove' stability, then following the pattern, perhaps actually 72 hours straight is needed. Because then 72 hours might reveal something 48 hours didn't. Or, well, maybe just run the stress test solidly for a week, or maybe a month ?! Because well, maybe 72 hours won't actually be enough....and so on.

So yes, what objectively does one, just one clean straight run of 48 hours prove over a single, clean straight run of 24 hours, or 12, or 6 hours. Probably nothing 🙂 Because, of course, when it comes to science, you need more than just one single reading. You need a whole repeated series to get the average from. Because then, one of those '48 hour runs' might tip over at the 13 hour mark, throwing out the results !

Now, the question becomes, how many '48 hour runs' are needed to prove the 48 hour stability claim...! And so the dog continues to chase its tail...:rolleyes::D

[Praz]

Theagg wrote:
When it comes to stress tests, the number of 'required' hours often has me wondering, why that many hours ?

So, why is a minimum of 24 hours of Prime95 needed ? What does 24 hours tell you that say 16 hours doesn't ? And would 48 hours tell you something that 24 hours doesn't.

Heck and if 48 hours straight is required to 'prove' stability, then following the pattern, perhaps actually 72 hours straight is needed. Because then 72 hours might reveal something 48 hours didn't. Or, well, maybe just run the stress test solidly for a week, or maybe a month ?! Because well, maybe 72 hours won't actually be enough....and so on.

So yes, what objectively does one, just one clean straight run of 48 hours prove over a single, clean straight run of 24 hours, or 12, or 6 hours. Probably nothing 🙂 Because, of course, when it comes to science, you need more than just one single reading. You need a whole repeated series to get the average from. Because then, one of those '48 hour runs' might tip over at the 13 hour mark, throwing out the results !

Now, the question becomes, how many '48 hour runs' are needed to prove the 48 hour stability claim...! And so the dog continues to chase its tail...:rolleyes::D

Hello

When using Prime type programs each individual will need to determine how long is a satisfactory length of time to cook their CPU. If memory stability testing is the goal a better option is either HCI or GSAT. Either of these programs will find memory errors that the shake and bake utilities miss.

[Theagg]

Praz wrote:
Hello

When using Prime type programs each individual will need to determine how long is a satisfactory length of time to cook their CPU. If memory stability testing is the goal a better option is either HCI or GSAT. Either of these programs will find memory errors that the shake and bake utilities miss.

I was, of course, being a little sarcastic, but yes, its a subjective determination when it comes to quoting a specific number of hours as being 'right'. I will go take a peek at those two you mentioned.

[SexyTerrorisT]

Theagg wrote:
When it comes to stress tests, the number of 'required' hours often has me wondering, why that many hours ?

So, why is a minimum of 24 hours of Prime95 needed ? What does 24 hours tell you that say 16 hours doesn't ? And would 48 hours tell you something that 24 hours doesn't.

Heck and if 48 hours straight is required to 'prove' stability, then following the pattern, perhaps actually 72 hours straight is needed. Because then 72 hours might reveal something 48 hours didn't. Or, well, maybe just run the stress test solidly for a week, or maybe a month ?! Because well, maybe 72 hours won't actually be enough....and so on.

So yes, what objectively does one, just one clean straight run of 48 hours prove over a single, clean straight run of 24 hours, or 12, or 6 hours. Probably nothing 🙂 Because, of course, when it comes to science, you need more than just one single reading. You need a whole repeated series to get the average from. Because then, one of those '48 hour runs' might tip over at the 13 hour mark, throwing out the results !

Now, the question becomes, how many '48 hour runs' are needed to prove the 48 hour stability claim...! And so the dog continues to chase its tail...:rolleyes::D

I see were you are getting with this. 🙂 We can argue that stability is quite relative depending on your needs and the type of workload you'll be running and for how long and under what temps. To put things into perspective nothing last forever. The more you zoom in the faster things transform and change state. From the lifespan of stars counted in billion of years to the wind that erodes montains in thousands of years to the electrons that rush through circuits at manometer scale in modern electronics.

There are some industry standards set in by manufacturers that differ whether you are a consumer or a professional. When you buy a cpu you expect to work flawlessly until its MTBF (mean time between failure) 24/7 provided you run it within manufacturer defined environment. Prosumers have more critical application with higher needs that is why they use ECC memory to correct errors when they occur. You can get memory errors because of cosmological radiation..

On a more practical level it all depends what you are going to run for how long and the rate of errors you are comfortable with. Is getting a BSOD everyday for you ok? A week ? A month? Maybe a year. I personally when overclocking i found errors to appear consistently after 24 hours that i did not manage to get even by running multiple shorter sessions. 48h for me is an arbitrary time i found from experience that i could not remember when the prior crash or bsod occured. Its a length of a period that has proven me good enough extrapolation under my usage pattern. Maybe i get a couple a year. More often during hot periods or when dust has overwhelmed my cooling system. In my line of work as a big data dev, the systems i use are designed with expecting every single part in a cluster to fail and to recover without interrupting the computation. And all the parts in those clusters are high quality xeons with ECC ram not overclocked. I really can not afford as a developer to waste time on errors I am not responsible or that i can't consistently recreate. Up to you to find what your needs are and how long you want your hardware to live and how well to operate.

I think it's a funny that people just run Prime95 for some hours or 1 day really when this program algorithm is meant to search for huge Mesrene prime numbers that require massive computation power to run for years... Also as i stated before does not use all the cpu instruction set. Fid
If you are intrested a couple easy vids to grasp the concept. Computation algorith is fancier that what you see in 2nd and 3rd video.
The definition of Mesrene Primes:
https://www.youtube.com/watch?v=T0xKHwQH-4I&t=330s
The biggest Mesrene Prime yet:
https://www.youtube.com/watch?annotation_id=annotation_3661090753&feature=iv&src_vid=T0xKHwQH-4I&v=Q...
Using Mesrene Primes to find the biggest prime number:
https://www.youtube.com/watch?v=lEvXcTYqtKU