3200 or 3733 for upcoming Zen 2?

tangoseal

[H]F Junkie
Joined
Dec 18, 2010
Messages
9,743
I've never ran faster than 3200.

And with all the marketing jazz AMD is putting out concerning the upcoming Ryzen 3000 series and its support of high clocked memory, what do you guys that know memory very well think?

Is 3733 going to make a noticeable difference over say 3200?

According to AMD the lowest memory latency for the newly redesigned IMC on Ryzen 3000 is going to be around 3733 mhz ram.Im not talking about CAS latency, but this:

d24hymdbbm331.png


So any opinions, would it make a difference in gaming/productivity/video work etc... over 3200?

The price is substantially more.
 
Look there'll be a benefit in 3733C17 vs 3600c16. AMD recommends the latter, but you'll get a small bump with the former. 133mhz when you're talking 3600mhz is a very small amount.
 
Does 3200C16 memory in general overclock to 3600C16 by just raising the frequency?

It's ridiculous how a 32GB 3200C16 set costs about 160e, while 32GB 3600C16 costs you 370e
 
^ nooo..that kit blowsss and is literally unstable after 3200mhz...

You need something like the Trident Z kit and specifically state that it IS capable of some OC headroom etc.
 
I have two 3200c14 G.Skill kits capable of 3866c16 with 1.45v so if you are willing to take the time to OC you don't have to spend extra for speeds over 3200c14 unless you want high binned kits to push further.

As for the difference it will improve CPU performance probably as much as overclocking the CPU itself.
But only in programs that are sensitive to memory speed so not cinbench but most games and photoshop\lightrrom
 
Last edited:
Manually tweaking timings could net you up to 20% above default XMP on Zen+ with a few games and on average the gain was around 10%, highly likely that 3733 with tight timings will be worth it for latency sensitive programs.
 
Let's be realistic. Could you really tell the difference anyway? I mean it's not like we're talking about 2133Mhz DIMMs here. I mean RAM speed is probably the worst use of extra upgrade money.

Personally, I'm looking at the $160 DDR4 3600 2x16GB G.Skill kit and then trying to tighten the timings manually.
 
If you plan to manually dial in the best timings at the highest 1:1 clocks possible, the ideal kit to buy is 3600C15 SR. Not cheap, but some of the lowest latency b-die out there. As far as I know there is only one Gskill SKU for this. There are some technically lower latency 4000+ b-die kits but they cost the moon and given AMD's info so far going for >3733 speeds is only for bragging rights, not actual performance.

3200C14 dual rank (16GB) might be pretty good too as that is one of the lowest latency DR kits to be had, supposedly Zen2 will be more friendly to DR and more populated configurations this time around. At equal speeds and timings DR slightly outperforms SR.

While there are othernon-samsung based kits out there that will clock up high, b-die is still best because nothing touches it on latency. So if you care, once again, stick to samsung :)
 
Last edited:
Isn't the difference between 3200 and 3600 400mhz effective, so wouldn't that imply another 5 or 6% uptick in gaming performance?
 
Let's be realistic. Could you really tell the difference anyway? I mean it's not like we're talking about 2133Mhz DIMMs here. I mean RAM speed is probably the worst use of extra upgrade money.
For me tweaking 3200c14 to up 3866c16 yielded up to 7% increase in CPU performance for games that were running around 70-80FPS
Combined with the 6.4% I gained from overclocking the 6700k to 4.7GHz added up to a 13.4% increase in CPU performance for games.

Since Ryzen memory interface has latency problems it spends more time waiting on the RAM than Intel 1151 combined with the gains from overclocking the infinity fabric and it all adds up.

But spending cash on CPU\RAM\MB\Cooling while skimping on the GPU for a high res monitor is not wise.
 
For me tweaking 3200c14 to up 3866c16 yielded up to 7% increase in CPU performance for games that were running around 70-80FPS
Combined with the 6.4% I gained from overclocking the 6700k to 4.7GHz added up to a 13.4% increase in CPU performance for games.

Since Ryzen memory interface has latency problems it spends more time waiting on the RAM than Intel 1151 combined with the gains from overclocking the infinity fabric and it all adds up.

But spending cash on CPU\RAM\MB\Cooling while skimping on the GPU for a high res monitor is not wise.

Can you share a lamens method of such a tweak? I.e. set X to Y and then try A to B etc...

About your post ... I think the latency issues are now history with zen 2 hence the huge IPC gain of 15% give or take. Well see.
 
It is a slow process but here was the result for the tuned timings on my system vs the stock.
It also required increased v to the RAM and IMC.
Basically start with frequency and primary timings and once you have that stable you can probably halve your secondary timings and hope that your MB does a decent job of handling the tertiary timings.
aida64%203866c15.jpg

86cabae0_6700k204.720xmp-3200c14.jpg


One of the most important things is a good stability test there is some good info on that here
https://www.overclock.net/forum/10-...-ryzen-ddr4-24-7-memory-stability-thread.html
and here for Intel
https://www.overclock.net/forum/5-i...-intel-ddr4-24-7-memory-stability-thread.html

Unstable RAM can corrupt your OS so it is good to run Linux form a USB boot and use GSAT for preliminary stability testing.
I found GSAT to find RAM related errors quicker than anything else but it didn't make the IMC work very hard so it could still be unstable which is where HCI comes in it is great for finding IMC errors and the DRAM calculator has made it a lot easier to launch multiple instances for testing.
 

As good as Zen / Zen+ guides can be, Zen 2 is going to be a completely different animal.

tangoseal, we're going to have to wait for in-depth looks to see how Ryzen 3000-series CPUs are affected by memory. Higher clockspeeds may not be as useful as tighter timings, and neither may have the impact that we've seen in previous generations, if AMD has implemented the improvements they've announced.
 
It should be really interesting seeing how this plays out in a month. The price difference between 3000/3200 and 3600+ ram is pretty big. I can get 32gb of 3000 15/16/16 for $130 vs 16gb of 3600 16/18/18 for $137, it would be nice if the slower stuff oc'd to within range of the faster (and if ryzen2 plays nice with 4 DIMMs).
 
Last edited:
dasa

Thanks a million for your write up. I appreciate it.

To the thread, can you confirm if this ram would be considered Samsung B-Die or this probably Hyinx C-Die or some other manufacturer?

I was thinking of this ram but I have no idea.

https://www.newegg.com/p/N82E168202...Product&RandomID=8438798251619120190613205633
I would guess hynix but I'm no expert.

I've been hearing more about Micron e-die recently, and they set an ln2 world record a few weeks back. No frag harder discos but its 3600 with what looks like better timings

https://www.newegg.com/ballistix-32gb-288-pin-ddr4-sdram/p/N82E16820164150
 
3600c15 and c16 are b die and maybe c17 on some 32gb kits but anything over that is unlikely.
3600c19 probably has higher final latency 3200c16 and may be using the same chips.
 
I have a metric shitton of samsung b-die kits: both the branded gskill and team stuff with bling, and a bunch of unbuffered ECC as well. Ever since Ryzen came out and shit on all the random Micron/Hynix/Whatever spare kits I had, I haven't bought any other brands of IC. All combined, I'm over a terabyte for sure.

Going to be a launch X570 builder because hobby is fun, I intend to put some of these kits through their paces. 3600C15 SR 2x8 & 4x8 and 3200C14 DR 2x16 are on the top of my list to try out for best possible settings. If Zen 2 is DR friendly I will see if 4x16 is still doable when high & tight, if the stars align then 64GB at 3733C14 or similar bleeding edge would be pretty damn neat.
 
I have a metric shitton of samsung b-die kits: both the branded gskill and team stuff with bling, and a bunch of unbuffered ECC as well. Ever since Ryzen came out and shit on all the random Micron/Hynix/Whatever spare kits I had, I haven't bought any other brands of IC. All combined, I'm over a terabyte for sure.

Going to be a launch X570 builder because hobby is fun, I intend to put some of these kits through their paces. 3600C15 SR 2x8 & 4x8 and 3200C14 DR 2x16 are on the top of my list to try out for best possible settings. If Zen 2 is DR friendly I will see if 4x16 is still doable when high & tight, if the stars align then 64GB at 3733C14 or similar bleeding edge would be pretty damn neat.

I just used the Ryzen Ram Tool program or whatever it is called and just for fun on my 2400g ... took my gskill C-die Hynix from 16 to 14 cas and gained like 20+FPS or more in Games using a 1070ti. It really helped the little 4 core feed the GPU more, alot more in fact.
 
IdiotInCharge

I'm going to build a 100% gaming rig with a 12 core that can do productivity on the side. So I'm all about extracting every ounce of performance.
I’m springing for the 16 core beast to do the same thing and up to now have never thought about manually doing anything other than XMP. This thread has got me thinking.
 
I just used the Ryzen Ram Tool program or whatever it is called and just for fun on my 2400g ... took my gskill C-die Hynix from 16 to 14 cas and gained like 20+FPS or more in Games using a 1070ti. It really helped the little 4 core feed the GPU more, alot more in fact.
This is the tool Tangoseal used:
https://www.techpowerup.com/download/ryzen-dram-calculator/

If you decrease a bottleneck then you will get results. But it is harder to say that the same metric applies to cpu with more cores because that bottleneck is not there. That does not mean you won't get better results from lower latency ram (tighter timings) but there much less noticeable. That is why people tend to use certain benchmarks to display numbers rather then real life game play.
 
Last edited:
I've never ran faster than 3200.

And with all the marketing jazz AMD is putting out concerning the upcoming Ryzen 3000 series and its support of high clocked memory, what do you guys that know memory very well think?

Is 3733 going to make a noticeable difference over say 3200?

According to AMD the lowest memory latency for the newly redesigned IMC on Ryzen 3000 is going to be around 3733 mhz ram.Im not talking about CAS latency, but this:

View attachment 167320

So any opinions, would it make a difference in gaming/productivity/video work etc... over 3200?

The price is substantially more.

This is on my mind as well. Are the mid range cpu skus going to be held back by 3200?
 
I've never ran faster than 3200.

And with all the marketing jazz AMD is putting out concerning the upcoming Ryzen 3000 series and its support of high clocked memory, what do you guys that know memory very well think?

Is 3733 going to make a noticeable difference over say 3200?

According to AMD the lowest memory latency for the newly redesigned IMC on Ryzen 3000 is going to be around 3733 mhz ram.Im not talking about CAS latency, but this:

View attachment 167320

So any opinions, would it make a difference in gaming/productivity/video work etc... over 3200?

The price is substantially more.
DDR4 3600 or 3733, absolutely is faster than DDR 3200. The true latency of your RAM is not simply the timings. Its multiplication of the RAM's tCK (time clock) Vs. the timings. RAM speed matters....more than timings. For 3200 to maybe outperform 3600 or 3733 at CL 18 or 19, you'd have to get the timings way down. I mean like around 12 or something. Stupid low. But You'd still be running slower clock speed. the relationship of timings Vs. clock rate isn't clear. But at the same total latency, higher clockrate basically always wins.

You can buy decent CL 18/19 3600 right now, for about $80 per 16GB. It should be a no-brainer. And prices may go lower, according to recent news on DRAM pricing. Spend a few bucks more, if you want some tighter timings. But, its not necessary.
 
Last edited:
DDR4 3600 or 3733, absolutely is faster than DDR 3200. The true latency of your RAM is not simply the timings. Its multiplication of the RAM's tCK (time clock) Vs. the timings. RAM speed matters....more than timings. For 3200 to maybe outperform 3600 or 3733 at cas18 or 19, you'd have to get the timings way down. I mean like around12 or something. Stupid low. and that would only be to about match overall latency. You'd still be running slower speed.

You can buy decent cas 18/19 3600 right now, for about $80 per 16GB. It should be a no-brainer. And prices may go lower, according to recent news on DRAM pricing. Spend a few bucks more, if you want some tighter timings. But, its not necessary.

I have some Corsair 3000mhz Hynix C Die @ cas 15, 32 GB of the stuff.

I hear it is actually VERY nice ram chips. I was wondering if overclocking the Corsair to 3600 and running looser timings like 17 or 18 cas would be as fast as buying 3600 cas 18 out of the box?

If I used the AMD Ram calculator tool to get all the subtimings that I need. I wonder if overclocking the multiplier is as easy as just changing it to 36 instead of 3000 with lower timings.

Or is there more to it than that.?Pardon me but I am not a ram overclocker when it comes to technical prowess.
 
I have some Corsair 3000mhz Hynix C Die @ cas 15, 32 GB of the stuff.

I hear it is actually VERY nice ram chips. I was wondering if overclocking the Corsair to 3600 and running looser timings like 17 or 18 cas would be as fast as buying 3600 cas 18 out of the box?
Shouldn't be any difference in performance. You'll just be running hotter, if you have to increase the RAM voltage.

The industry standard for total latency of DDR4 RAM, is about 14 nanoseconds. DDR4 3200 hits 14 nanoseconds, at CL 22.
 
Last edited:
Shouldn't be any difference in performance. You'll just be running hotter, if you have to increase the RAM voltage.

The industry standard for total latency of DDR4 RAM, is about 14 nanoseconds. DDR4 3200 hits 14 nanoseconds, at CL 22.

So your saying just leave it at 3000mhz cas 15? It's going to be about as fast as 3600 at 18 or 19?
 
So your saying just leave it at 3000mhz cas 15? It's going to be about as fast as 3600 at 18 or 19?
Total latency would be similar. But, 3600 would still have a 600mhz faster clock rate. At similar total latency, that matters. You still benefit from the faster speed. Let me use a more extreme example to illustrate the point. DDR2 800mhz at CL 5 is about 14 nanoseconds of total latency. But trying to say that's as fast as DDR4 3200 at 14 nanoseconds of total latency, would be ridiculous. At similar total latency, the faster speed matters.

Its unclear about how low you'd have to tighten timings, how much less total latency you'd need, before a lower speed/mhz RAM would generally perform better than one with faster speed/higher MHZ. I've never seen a test/review which really tried to quantify it. And I myself do not have super high mhz RAM, in order to test it. My current RAM is DDR4 3000 CL15, which I have overclocked to 3333 CL17.
 
Last edited:
Total latency would be similar. But, 3600 would still have a 600mhz faster clock rate. At similar total latency, that matters. You still benefit from the faster speed. Let me use a more extreme example to illustrate the point. DDR2 800mhz at CL 5 is about 14 nanoseconds of total latency. But trying to say that's as fast as DDR4 3200 at 14 nanoseconds of total latency, would be ridiculous. At similar total latency, the faster speed matters.

Its unclear about how low you'd have to tighten timings, how much less total latency you'd need, before a lower speed/mhz RAM would generally perform better than one with faster speed/higher MHZ. I've never seen a test/review which really tried to quantify it. And I myself do not have super high mhz RAM, in order to test it. My current RAM is DDR4 3000 CL15, which I have overclocked to 3333 CL17.

Thanks Chameloneel

so using the formula: ( CAS/Ram Freq ) x 2000 which I got from Linus Tech Tips

My Corsair (edit sorry had gskill) 3000 cas 15 would be ---> (15/3000mhz)*2000 = 10ns

DDR4-3600@ 18ns would be ----> (18/3600)*2000 = 10ns as well

So based on these numbers then I would derive no performance benefit from 3600@18ns over my 3000@15ns in raw access times such as gaming. But of course transfer bandwidth is significantly higher on the 3600 mhz.

hrrm so lets say a real super high end like Corsair Platinum 4800mhz @ 750.00 USD for 16GB ouch

would be (18/4800)*2000 = 7.5ns

Is this accurate or am I total out in space somewhere?
 
Last edited:
If you plan to manually dial in the best timings at the highest 1:1 clocks possible, the ideal kit to buy is 3600C15 SR. Not cheap, but some of the lowest latency b-die out there. As far as I know there is only one Gskill SKU for this. There are some technically lower latency 4000+ b-die kits but they cost the moon and given AMD's info so far going for >3733 speeds is only for bragging rights, not actual performance.

3200C14 dual rank (16GB) might be pretty good too as that is one of the lowest latency DR kits to be had, supposedly Zen2 will be more friendly to DR and more populated configurations this time around. At equal speeds and timings DR slightly outperforms SR.

While there are othernon-samsung based kits out there that will clock up high, b-die is still best because nothing touches it on latency. So if you care, once again, stick to samsung :)
Some hynix cjr can beat b-die sub timing. Not every timing but some. Cheap as hell too and clocks to the magic number.
 
Thanks Chameloneel

so using the formula: ( CAS/Ram Freq ) x 2000 which I got from Linus Tech Tips

My Gskill 3000 cas 15 would be ---> (15/3000mhz)*2000 = 10ns

DDR4-3600@ 18ns would be ----> (18/3600)*2000 = 10ns as well

hrrm so lets say a real super high end like Corsair Platinum 4800mhz @ 750.00 USD for 16GB ouch

would be (18/4800)*2000 = 7.5ns

Is this accurate or am I total out in space somewhere?
oh that's a cool forumula. Yes that seems accurate. My DD4 3000 has a tCK of 0.667 and an XMP CL of 15.

0.667 x 15 = 10 nano seconds. So that formula looks correct.
 
oh that's a cool forumula. Yes that seems accurate. My DD4 3000 has a tCK of 0.667 and an XMP CL of 15.

0.667 x 15 = 10 nano seconds. So that formula looks correct.

SOrry I had to correct my post you quouted. Its not gskill 15 cas 3000 its Corsair Vengeance Pro. A typo on my part.
 
So based on these numbers then I would derive no performance benefit from 3600@18ns over my 3000@15ns in raw access times such as gaming.

3600@18CL? (cas latency)

Do some tests.

Virtually all reviews I have seen, show that higher MHZ RAM is beneficial for gaming. I think people put a lot more emphasis on timings than they should. As I said before, at similar total latency, the extra speed matters.

I think that DDR4 3600 would outperform DDR4 3000 in most cases. And probably never really lose. Especially considering we are talking about a new CPU which is built for faster RAM.
 
Last edited:
3600@18CL? (cas latency)

Do some tests.

Virtually all reviews I have seen, show that higher MHZ RAM is beneficial for gaming. I think people put a lot more emphasis on timings than they should. As I said before, at similar total latency, the extra speed matters.

I think that DDR4 3600 would outperform DDR4 3000 in most cases. And probably never really lose. Especially considering we are talking about a new CPU which is built for faster RAM.

Ok I will just wait for the legitimate reviews to come out and go with the ram that gives the best performance without getting too epeen.
 
Think about memory latency the same way we measure hard drives.

It's not just the initial response time, but also how long it takes to transfer targets of various sizes. This is why higher clocked RAM wins; at the same real latency, slower clocked memory will start a transfer at the same time as the faster clocked memory, but as the transfer size grows, the faster clocked memory will finish the whole transfer sooner due to higher bandwidth.

This is also why it might be advantageous in some situations to sacrifice real latency for clockspeed.
 
Back
Top