3200 or 3733 for upcoming Zen 2?

[dasa]

Latency vs bandwidth maters more in some games than others and I think F1 prefers bandwidth while games I typically play prefer latency.

In my testing 3000c12 vs 4000c17 with the same final latency in aida64 is only ~ 1-2% faster despite 20-30% increase in bandwidth.

While 3000c12 with tweaked subs was ~4-6% faster than 3200c14 XMP despite slightly less bandwidth due to the 7% latency difference.
https://www.overclock.net/forum/180...0c11-2133c9-ddr4-2133c15-3000c12-4000c17.html
Still every little bit helps and 4000 c17 was actually more stable requiring less v than 3000c12

3000\12=250
4000\17=235
just goes to show this rough calculation is only a guideline.

 

[Lepardi]

I wonder how the pricing will be in 2020 for Zen 3. DDR5 is of course going to be way too expensive unless you want super slow speeds, and the B-die supplies for DDR4 must have ran out by then as well.

 

[tangoseal]

I wonder how the pricing will be in 2020 for Zen 3. DDR5 is of course going to be way too expensive unless you want super slow speeds, and the B-die supplies for DDR4 must have ran out by then as well.

B die isn't being made anymore

To my understanding the alternate option is Hynix C die. Also Samsung is aking A die now?

But see I have no idea what this stuff translates too. Does the A mean grade A loke in eggs and B is grade B?

Is C hynix's best or is it B grade. So confusing. Why not just make an easy to understand nomenclature.

 

[tangoseal]

Let me ask this ...

Here is a SS of some Gskill I have and this on my son's 2400G APU

How would I bump the ram frequency higher than 3200 using this application? I mean if I wanted to try and squeeze (hypothetically) 3600mhz, besides throwing some more voltage at it and changing the multiplier of the ram what am I to do.

Does this software calculate the settings if say I change the 3200MT/s drop down in the application to something higher? Or is the software assuming that the ram I have is rated for said speed already and im just applying tighter timings?

For instance I am taking 3200 CAS 16 and using the ram calculator tool to get the above settings which is rock solid at cas 14 with the above parameters. What do I need to do if I want to change it say 3600mhz?

This is especially important as I am going to be getting two Zen2 CPUs next month.

 

[kirbyrj]

I don't think you bump the memory speed from the application. It just gives you numbers and you can plug them into your bios. it's really just a repository of testing that they have done with different modules.

Honestly I never had good luck with the numbers they gave me from that application.

 

[cyklondx]

Its worth to question the supplied slide from AMD, if this is only x570 or does this also take into account older mobo's with x470. (I have feeling they only account for x570.)
I think most x470 boards will top up at ~3466MHz (with 3600MHz 2x16GB CL19) with zen2, and some I assume will go higher.

I expect CL16 3200 vs CL16 3600 will be around 2-4FPS in most games, at best. Not mentioning that sticks with better density at lower timings will cost significantly more... from ~$150 2x16GB CL19 to ~$320 2x16 CL17 | Is it worth it?

 

[IdiotInCharge]

Its worth to question the supplied slide from AMD, if this is only x570 or does this also take into account older mobo's with x470.

This will need to be tested, including the case where motherboard manufacturers state that their boards can enable Ryzen 3000 CPUs to handle the faster memory.

 

[Lepardi]

Its worth to question the supplied slide from AMD, if this is only x570 or does this also take into account older mobo's with x470. (I have feeling they only account for x570.)
I think most x470 boards will top up at ~3466MHz (with 3600MHz 2x16GB CL19) with zen2, and some I assume will go higher.

I expect CL16 3200 vs CL16 3600 will be around 2-4FPS in most games, at best. Not mentioning that sticks with better density at lower timings will cost significantly more... from ~$150 2x16GB CL19 to ~$320 2x16 CL17 | Is it worth it?

Samsung has shifted from 8GB B-die onto 16GB A-Die production, wouldn't that reduce the cost of 2x16GB sets? If the A-Die proves to be decent

 

[cyklondx]

Samsung has shifted from 8GB B-die onto 16GB A-Die production, wouldn't that reduce the cost of 2x16GB sets? If the A-Die proves to be decent

in perfect world; but we both know in their eyes 2x the density 2x the price.

 

[chameleoneel]

Its worth to question the supplied slide from AMD, if this is only x570 or does this also take into account older mobo's with x470. (I have feeling they only account for x570.)
I think most x470 boards will top up at ~3466MHz (with 3600MHz 2x16GB CL19) with zen2, and some I assume will go higher.

I expect CL16 3200 vs CL16 3600 will be around 2-4FPS in most games, at best. Not mentioning that sticks with better density at lower timings will cost significantly more... from ~$150 2x16GB CL19 to ~$320 2x16 CL17 | Is it worth it?

No it is not worth it.

For DDR4 3600, CL17 Vs. CL19 is only about 1 nanosecond difference in total latency, at the same speed.

 

[dasa]

If you are buying a 3600 with a cheap MB and a cheap HSF no it is not worth it but if you are spending more on the CPU\MB\cooling with the intent to OC then I believe 3200c14\3600c16 b die is worth the extra.

 

[Zarathustra[H]]

https://www.techpowerup.com/reviews/AMD/Ryzen_Memory_Tweaking_Overclocking_Guide/

seems pretty good, probably what I'll be using as a baseline

I wonder if this will still be relevant with Zen2, as they have made substantial changes to the memory controller and how it links with the infinity fabric in this revision.

I've been having similar thoughts as OP as of late.

What kind of RAM to buy for Ryzen 3000?

The AMD CPU guy (Robert Hallock?) in the video in this news post stated that he thought the sweet spot was DDR4-3600. Last week I posted a news story suggesting that the memory controller would hit peak performance at DDR4-3733.

He was talking about 3600Mhz CL14, but in my searches, this RAM simply does not exist. Even CL15 is pretty rare. Most 3600Mhz DDR4 seems to be CL16 or above.

If the news story from last week is correct in that you maximize the infinity fabric clock when RAM is clocked at 3733, I wonder if we'd be better off buying 3600Mhz CL15 RAM (lowest timings I could find at 3600Mhz)) or 3733Mhz CL17 (lowest timings I could find at 3733Mhz) or if it even really matters, if they are essentially the same ram with lowered timings so they can hit a higher clock?

(Part of me wants to buy the higher rated RAM just to be guaranteed to hit the best clock rather than leave it up to the chance of overclocking though...)

I guess no one except AMD really knows how the memory controller in Ryzen 3000 will scale with clock speed and timings yet, but what are some of you who know RAM better than me thinking?

(I have to admit, RAM performance is my biggest gap in knowledge in this hobby)

 

[Zarathustra[H]]

Latency vs bandwidth maters more in some games than others and I think F1 prefers bandwidth while games I typically play prefer latency.

In my testing 3000c12 vs 4000c17 with the same final latency in aida64 is only ~ 1-2% faster despite 20-30% increase in bandwidth.

While 3000c12 with tweaked subs was ~4-6% faster than 3200c14 XMP despite slightly less bandwidth due to the 7% latency difference.
https://www.overclock.net/forum/180...0c11-2133c9-ddr4-2133c15-3000c12-4000c17.html
Still every little bit helps and 4000 c17 was actually more stable requiring less v than 3000c12

3000\12=250
4000\17=235
just goes to show this rough calculation is only a guideline.

We also have to keep in mind that the news last week suggested that the infinity fabric clock maxes out when the RAM is clocked at 3733. After that it scales back with a divider, so even if you hit 5200Mhz on the RAM (which is reportedly possible) it may be slower than 3733. Everything above 3733 may wind up being slower than 3733.

 

[extide]

We also have to keep in mind that the news last week suggested that the infinity fabric clock maxes out when the RAM is clocked at 3733. After that it scales back with a divider, so even if you hit 5200Mhz on the RAM (which is reportedly possible) it may be slower than 3733. Everything above 3733 may wind up being slower than 3733.

I wonder if you can control that divider manually. Say I wanted to try 3800 or something with 1:1 -- hopefully higher ends boards will have a manual option in the BIOS for that ratio instead of it being automatic only.

 

[Zarathustra[H]]

I just came across this which is a little troubling.

Tom's is quoting ASRock recommendations as anything 3200 or higher being achieved only using two slots, and the max dropping with more slots populated.

I had planned on going 4x 16GB at 3733Mhz.

I don't really NEED 64GB of RAM, but that is what I have in my current build (mostly because I had spare RAM from a server build I just popped in) and it just doesn't feel right to go backwards spec wise as part of an upgrade...

 

[extide]

These are the official specs, so definitely not the max. So I was watching the below video and they were talking about memory speeds. They said they got ~4400 in the labs and partners have gotten up to 5100! AMD even said the sweet spot would be about 3733 or 3600 with lowish cas (~16). DDR4 3600 is priced pretty good right now so that is probably what I will shoot for.



I am also going to be going with 64GB minimum but I think I will try out the new 32GB DIMM's from Samsung which apparently overclock to about that speed pretty well. (Plus that would put me at 2 sticks of Dual Rank instead of 4 sticks of Dual Rank)

 

[Zarathustra[H]]

I am also going to be going with 64GB minimum but I think I will try out the new 32GB DIMM's from Samsung which apparently overclock to about that speed pretty well. (Plus that would put me at 2 sticks of Dual Rank instead of 4 sticks of Dual Rank)

If those 32GB sticks work and aren't outrageously priced, that may be the way to go!

 

[extide]

If those 32GB sticks work and aren't outrageously priced, that may be the way to go!

They are quite fairly priced IMHO, check out this: https://www.anandtech.com/show/14524/samsung-32gb-non-ecc-memory

Currently $160 @ AVADirect

 

[Zarathustra[H]]

They are quite fairly priced IMHO, check out this: https://www.anandtech.com/show/14524/samsung-32gb-non-ecc-memory

Indeed,

I can't watch the video at work, but what were the final speeds and timings they wound up with?

I googled the part number and they appear to be CL19 2666Mhz out of the box. :/

 

[extide]

Indeed,

I can't watch the video at work, but what were the final speeds and timings they wound up with?

I googled the part number and they appear to be CL9 2666Mhz out of the box. :/

So that video is a Buildzoid video and he got them up to 3600 at cas 18 or 19 I think. Not too shabby honestly.

 

[Zarathustra[H]]

So that video is a Buildzoid video and he got them up to 3600 at cas 18 or 19 I think. Not too shabby honestly.

This does seem like an option. I've always had bad luck with RAM overclocking, I always seem to wind wup with instability when I try, so my inclination is to go with something that I don't HAVE to run out of spec in order to be happy, just in case it doesn't work out, but with this price and the potential shortcomings of running 4 sticks with Ryzen3000, this may just wind up being what has to be done.

I bet they are cheaper because they don't have those stupid heat spreaders and RGB lights.

Do 32Gb sticks "just work" in platforms not explicitly designed for them?

 

[IdiotInCharge]

We also have to keep in mind that the news last week suggested that the infinity fabric clock maxes out when the RAM is clocked at 3733. After that it scales back with a divider, so even if you hit 5200Mhz on the RAM (which is reportedly possible) it may be slower than 3733. Everything above 3733 may wind up being slower than 3733.

We need to see how data flows through the Inifnity Fabric on Zen 2. We don't really know how CAS latency will affect performance either given the bucketloads of cache that they added, and if everything has to flow through that cache we may see best performance overall with fastest memory just to keep the IF clock up.

We might also see the effects depend on workload, as they should, but even more so.

 

[dasa]

I wonder if you can control that divider manually. Say I wanted to try 3800 or something with 1:1 -- hopefully higher ends boards will have a manual option in the BIOS for that ratio instead of it being automatic only.

It is apparently a software change so I am sure at least some MB will make it available rather than just being a auto setting that switches after 3733 but apparently even anything beyond 3600 is often not stable so even people aiming for 3733 without tweaking v may be out of luck.

We don't really know how CAS latency will affect performance either given the bucketloads of cache that they added,

Unfortunately I think they are still way short of the a mount of cache required to satisfy a lot of games I think it needs to be closer to the 128MB L4 on the 5775c not 16MB per CCX.
A few games like crysis 3 already see minimal improvement from higher RAM speeds so it may go close to fitting into the cache and with a larger cache the number of games that benefit more from overclocking CPU than those that benefit greater from overclocking RAM may increase.

So just for kicks I thought I would see how much of a performance hit my 6700k takes if I try make the latency as bad as possible.
So here it is at 3200c21 any higher and it would fail to post.
This resulted in 59ns vs 42ns@3200c14 a 33% difference.

Cinbench 20 multi 2167 vs 2178 0.5% difference
Farcry5 83 vs 102FPS 20% difference
Arma III 53FPS vs 63FPS 17% difference <The reason the difference in Arma is lower than farcry is probably due to the lower FPS.

 

[Zarathustra[H]]

It is apparently a software change so I am sure at least some MB will make it available rather than just being a auto setting that switches after 3733 but apparently even anything beyond 3600 is often not stable so even people aiming for 3733 without tweaking v may be out of luck.


Unfortunately I think they are still way short of the a mount of cache required to satisfy a lot of games I think it needs to be closer to the 128MB L4 on the 5775c not 16MB per CCX.
A few games like crysis 3 already see minimal improvement from higher RAM speeds so it may go close to fitting into the cache and with a larger cache the number of games that benefit more from overclocking CPU than those that benefit greater from overclocking RAM may increase.

So just for kicks I thought I would see how much of a performance hit my 6700k takes if I try make the latency as bad as possible.
So here it is at 3200c21 any higher and it would fail to post.
This resulted in 59ns vs 42ns@3200c14 a 33% difference.

Cinbench 20 multi 2167 vs 2178 0.5% difference
Farcry5 83 vs 102FPS 20% difference
Arma III 53FPS vs 63FPS 17% difference <The reason the difference in Arma is lower than farcry is probably due to the lower FPS.

View attachment 171381

View attachment 171382

Crazy how much more RAM latency impacts games than it does rendering.

In retrospect it makes sense though, as the rendering threads are really just running the same algorithm over and over so that is probably almost entirely in the L3 cache.

 

[IdiotInCharge]

Crazy how much more RAM latency impacts games than it does rendering.

In retrospect it makes sense though, as the rendering threads are really just running the same algorithm over and over so that is probably almost entirely in the L3 cache.

Which means less branching. Every time the CPU needs to make a decision, it takes a series of chances- and one of those chances is whether the information that it needs is in cache. When the CPU loses on every opportunity to get stuff from one of its caches and must reach out to main memory, that's when various memory latencies come into play.

 

[chameleoneel]

Crazy how much more RAM latency impacts games than it does rendering.

In retrospect it makes sense though, as the rendering threads are really just running the same algorithm over and over so that is probably almost entirely in the L3 cache.

I've actually seen very little evidence that lower timings are meaningful in most games. There are a couple of games here or there, yes. But most RAM tests I have seen, show that increased RAM speed is better in most situations. Not just games. Even though timings get higher, increased speed keeps overall latency down. and you still have increased speed.

However, there are not many articles which have tested with really expensive, fast, tweakable RAM. However, I did recently find that Tomshardware did such an article for AMD and Intel:

https://www.tomshardware.com/reviews/best-ram-speed,5951-5.html

https://www.tomshardware.com/reviews/best-ram-speed-x470-pinnacle-ridge,6064-5.html


Regarding Zen 2/Ryzen 3000: I wonder if the information so far about the memory divider is a miscommunication or an actual performance cut. I mean, of course official marketing PR would say that a 1:1 situation is ideal. But does performance actually drop when the RAM speed scales higher? or is it just similar to Intel where you can just keep increasing RAM frequency and gaining benefits, even though you are well away from 1:1.

 

[dasa]

The thing that makes 1:1 faster with AMD is that it gains performance with the increased infinity fabric speed while Intel doesn't gain much from changes in it's bus speed.

I wouldn't be surprised if bandiwdth mattered more on a 8 core CPU and fully multithreaded game making use of 16 threads than it does on my quad core but the final latency will still be of high importance.

2133 vs 4000 will also make next to no difference with cinbench as unlike games, it seems it is a orderly process that mostly works mostly within the cache while games are rather random.
But other rendering tasks like a batch of photos in lightroom or photoshop can complete a lot quicker with higher RAM speeds perhaps this is just bad optimization?
File compression\decompression is also affected by RAM speed so things like 7zip or loading programs can actually complete faster in some circumstances with higher RAM speeds on a SATA SSD than with slow RAM speeds and a 970 EVO.

 

[IdiotInCharge]

The thing that makes 1:1 faster with AMD is that it gains performance with the increased infinity fabric speed while Intel doesn't gain much from changes in it's bus speed.

We know this applies strongly to Zen and Zen+, but we really don't know how Zen 2 will respond.

 

[extide]

We know this applies strongly to Zen and Zen+, but we really don't know how Zen 2 will respond.

Yeah I mean you have the IF link between the two CCX's in a single CCD (core complex die) and then the IF links between the CCD's and the IO die. I don't think I have seen any info on whether or not there is just a single IF clock or if those links run at different speeds. It may end up that ram speed and thus IF speed has even more of an impact, or perhaps less.

 

[IdiotInCharge]

or perhaps less.

I'm actually expecting less; I see it as a bit embarrassing for AMD that they tied their Zen / Zen+ CPUs to memory clock and then had trouble running memory at relatively pedestrian speeds.

 

[OnceSetThisCannotChange]

My problem is knowledge. Is there a beginner's guide to ram tweaking?

Not sure if anyone posted in the meantime - but here it is, also main article on TPU.

You can see some examples there with Shadow of the Tomb Raider having 20% improved avg FPS and 24% better minimums using Flare X optimized timings vs Flare X XMP, which is what reviewers typically use.

edit: on Zen/Zen+ , we do not know what is the potential on Zen 2, it's too early at this stage.

 

[chameleoneel]

Not sure if anyone posted in the meantime - but here it is, also main article on TPU.

You can see some examples there with Shadow of the Tomb Raider having 20% improved avg FPS and 24% better minimums using Flare X optimized timings vs Flare X XMP, which is what reviewers typically use.

edit: on Zen/Zen+ , we do not know what is the potential on Zen 2, it's too early at this stage.

Techpowerup has a decent little article for RAM scaling on Zen 2
https://www.techpowerup.com/review/amd-zen-2-memory-performance-scaling-benchmark/

Overall, there doesn't seem to be as much benefit, as usual. Seems Zen 2 is well optimized with its larger caches and new memory subsystem.

However, it looks like 3200mhz CL14 or 3600mhz CL16 are good targets.

 

[OnceSetThisCannotChange]

Thanks, there doesn't seem to be much, if anything, in it now.

 

[dasa]

I bet there will still be a decent improvement with 3600c14-c15\3733c16 with tweaked subs vs 17-19-19
Here is a review testing 3200c14 and 3733 unknown timings and PBO enabled at which point it is faster than the stock 9900K with 2666c16 in most games till you OC the 9900K.
It is ~17% faster in some games and PBO typically only makes the CPU ~3% faster in most tests where RAM speed doesn't matter like cinbench.
https://pclab.pl/art81116-23.html

 

[Nenu]

B die isn't being made anymore

To my understanding the alternate option is Hynix C die. Also Samsung is aking A die now?

But see I have no idea what this stuff translates too. Does the A mean grade A loke in eggs and B is grade B?

Is C hynix's best or is it B grade. So confusing. Why not just make an easy to understand nomenclature.

I use Samsung D die 3000MHz C15, its running at 3733MHz C17 1.46V.
It will use less voltage at higher CAS.
It is faster than 3733MHz at over 1.5V but I dont want to be a guinea pig. ie can this cause harm?
And higher than 3600MHz restricts my 6700K overclock by 100MHz.

I would like to know how well similar ram runs on AMD, especially Zen 2.
Will I get the same memory overclock and no impact on CPU overclock, for example?

 

[chameleoneel]

I bet there will still be a decent improvement with 3600c14-c15\3733c16 with tweaked subs vs 17-19-19
Here is a review testing 3200c14 and 3733 unknown timings and PBO enabled at which point it is faster than the stock 9900K with 2666c16 in most games till you OC the 9900K.
It is ~17% faster in some games and PBO typically only makes the CPU ~3% faster in most tests where RAM speed doesn't matter like cinbench.
https://pclab.pl/art81116-23.html

Honestly, I don't think Zen 3 gets big benefits from RAM improvements. In Techpowerup's RAM scaling article, the one game which showed a large improvement over really average DDR4 3000, was Sekiro. Considering that Zen 2 supports DDR4 3200 out of the box-------there isn't a lot of room there for improvement, with the way it scales.

*I think that's a good thing. It means you can likely carry over your existing RAM or just buy something affordable and get MOST of the performance. Within a few FPS.

 

[chameleoneel]

I use Samsung D die 3000MHz C15, its running at 3733MHz C17 1.46V.
It will use less voltage at higher CAS.
It is faster than 3733MHz at over 1.5V but I dont want to be a guinea pig. ie can this cause harm?
And higher than 3600MHz restricts my 6700K overclock by 100MHz.

I would like to know how well similar ram runs on AMD, especially Zen 2.
Will I get the same memory overclock and no impact on CPU overclock, for example?

https://www.techpowerup.com/review/amd-zen-2-memory-performance-scaling-benchmark/

 

[Nenu]

https://www.techpowerup.com/review/amd-zen-2-memory-performance-scaling-benchmark/

I have no doubt new ram rated at a certain speed will function as expected.
I wonder if my 4 year old ram will get the same overclock on Zen 2.

 

[dasa]

Honestly, I don't think Zen 3 gets big benefits from RAM improvements.

Then why does the review I linked suggest it does?
No doubt it will come down to the game and considering the performance gains in CSGO it probably fit's most of it's data into the larger cache now while there will be other games that are far from it.

Techpowerup shows 5.7% improvement in farcry 5 from 3000c16 to 3200c14 that is not bad

 

[chameleoneel]

Then why does the review I linked suggest it does?
No doubt it will come down to the game and considering the performance gains in CSGO it probably fit's most of it's data into the larger cache now while there will be other games that are far from it.

Techpowerup shows 5.7% improvement in farcry 5 from 3000c16 to 3200c14 that is not bad

Honestly, its difficult to use the data from that review. Because the 3733 RAM speed is also using Precision Boost Overclocking. And they are comparing that to the stock clocks, with 3200 RAM. So, RAM isn't the only thing changing performance.

In Techpowerup's Zen 2 memory scaling article, RAM is the only variable they changed.