10% more performance on Zen 3 with 4 sticks of ram?! - G Nexus

They tested 2x8 GB vs. 4x8 GB. A better test would have been 2x16 GB vs. 4x8 GB. I'm betting the advantage would disappear. Whatever they are testing obviously can take advantage of the extra RAM.

And IIRC the issue with four stick has never been about speed but stability at speed, and 3200 isn't that difficult a target.
 
Last edited:
Albeit on on Intel platforms, but i'v used full slots on most of my PCs the last 10 years , honestly it was always less stable that one dimm per channel.
I hope it not the same with Zen 3 as its shaping up to be my next build
 
BlueLineSwinger said:
They tested 2x8 GB vs. 4x8 GB. A better test would have been 2x16 GB vs. 4x8 GB. I'm betting the advantage would disappear. Whatever they are testing obviously can take advantage of the extra RAM.

And IIRC the issue with four stick has never been about speed but stability at speed, and 3200 isn't that difficult a target.
Click to expand...
He kept suggesting the capacity difference was not the key variable.
 
BlueLineSwinger said:
They tested 2x8 GB vs. 4x8 GB. A better test would have been 2x16 GB vs. 4x8 GB. I'm betting the advantage would disappear. Whatever they are testing obviously can take advantage of the extra RAM.

And IIRC the issue with four stick has never been about speed but stability at speed, and 3200 isn't that difficult a target.
Click to expand...
He points out towards the end of the video Wendell said 2x16GB single rank performed best of all. I think he didnt go into detail to save it for a separate video. I agree about stability at speed being the real issue.
 
tybert7 said:
He kept suggesting the capacity difference was not the key variable.
Click to expand...

Huh? It's the only variable, unless he screwed up and missed something. The setup is otherwise identical, yet is faster with 32 GB. Whatever they're testing with obviously makes use of and benefits from that extra RAM. That it was spread over four sticks is irrelevant.


pAiNkIlLaHvX said:
He points out towards the end of the video Wendell said 2x16GB single rank performed best of all. I think he didnt go into detail to save it for a separate video. I agree about stability at speed being the real issue.
Click to expand...

I didn't get that far into the video, checked out when it appeared the test was flawed (varying RAM amounts). Was the 2x16 GB test at the same speeds as the others?
 
BlueLineSwinger said:
I didn't get that far into the video, checked out when it appeared the test was flawed (varying RAM amounts). Was the 2x16 GB test at the same speeds as the others?
Click to expand...
Haha I guess we will have to wait and see. He is putting that in a different video. His working thesis is Zen 3 performs better with four sticks rather than two. There was no evidence of 2x16GB performance in this video because he said it didn't make a difference but sounds like he will post findings in the next video about 2x16GB SR performance.
 
BlueLineSwinger said:
Huh? It's the only variable, unless he screwed up and missed something. The setup is otherwise identical, yet is faster with 32 GB. Whatever they're testing with obviously makes use of and benefits from that extra RAM. That it was spread over four sticks is irrelevant.
Click to expand...

He said that no game even used close to 16GB, much less 32.
 
pAiNkIlLaHvX said:
Haha I guess we will have to wait and see. He is putting that in a different video. His working thesis is Zen 3 performs better with four sticks rather than two. There was no evidence of 2x16GB performance in this video because he said it didn't make a difference but sounds like he will post findings in the next video about 2x16GB SR performance.
Click to expand...

His thesis appears flawed, and so is his testing to this point. Another stating that 2x16 GB performed best would seem to bear this out (again assuming the speeds were the same). The only way I can see four sticks inherently performing better is if the CPU had a quad-channel controller. However, like every modern mainstream desktop/laptop CPU, Zen 3 only dual-channel.

That he seeming rambled on for almost a half hour and couldn't find time to properly test 2x16 GB makes it seem like he's trying to push for views and follows.


Derangel said:
He said that no game even used close to 16GB, much less 32.
Click to expand...

What was the OS using? Other apps? Were the GPU drivers caching anything in RAM? What was the total RAM used? Is more being paged out if less RAM is installed?
 
Nafensoriel said:
Cities skylines with mods would like to have a word with him....
Or galciv3...
There are actually several games out now that can easily push the 16gb limit.
Click to expand...

OBVIOUSLY he was solely talking about the games they tested.

BlueLineSwinger said:
His thesis appears flawed, and so is his testing to this point. Another stating that 2x16 GB performed best would seem to bear this out (again assuming the speeds were the same). The only way I can see four sticks inherently performing better is if the CPU had a quad-channel controller. However, like every modern mainstream desktop/laptop CPU, Zen 3 only dual-channel.

That he seeming rambled on for almost a half hour and couldn't find time to properly test 2x16 GB makes it seem like he's trying to push for views and follows.




What was the OS using? Other apps? Were the GPU drivers caching anything in RAM? What was the total RAM used? Is more being paged out if less RAM is installed?
Click to expand...

Unless, of course, Zen 3 simply manages memory better when more sticks are installed. Perhaps something with the way the infinity cache works with the memory might make it marginally more efficient with more RAM? Generally board memory topology only matters for overclocking, but perhaps that could play a part?

Steve just rambles, that’s the way he talks. He likes to take time to explain things and go into some kind of detail. Long videos are kind of GN’s thing. Seems like the reason they didn’t test 2x16 is because they don’t really have much of it laying around. I’d imagine any they have is in use in their production systems. He wants to talk to Wendell more after the video is up and try and dig into it further. So, I assume we’ll see some kind of update down the line.

Pretty sure they use Win 10, not sure which version. As for the rest, I imagine someone like Wendell would have brought that up when discussing the findings with Steve. Wendell knows his shit when it comes to memory and even he has seen this behavior with Zen 3. Also, repeating the tests on an Intel CPU (CPU and mobo being the only differences) produced no notable differences. This seems to point to something besides pure capacity being at play.
 
Nafensoriel said:
Cities skylines with mods would like to have a word with him....
Or galciv3...
There are actually several games out now that can easily push the 16gb limit.
Click to expand...
There are tons of applications in general that can use all the RAM you have. Gamers may not "need" more than 16GB but you could use all of that in video projects, VFX like After Effects or inside of Resolve, 3D modelling/Rendering, amonst others.

Unless Zen 3 has quad channel, it's hard to posit 4 sticks of RAM be faster as your thesis statement over simply "more RAM" being the relevant measure.
 
Derangel said:
He said that no game even used close to 16GB, much less 32.
Click to expand...
That's not the end of the story. RAM fragmentation can also play a role. Your game might only use 8GB at any given time, but if suddenly there is a request to allocate new memory it matters wheter it can be done on completely free memory, or it has to use already used bits filled with cached assets.
 
BlueLineSwinger said:
His thesis appears flawed, and so is his testing to this point. Another stating that 2x16 GB performed best would seem to bear this out (again assuming the speeds were the same). The only way I can see four sticks inherently performing better is if the CPU had a quad-channel controller. However, like every modern mainstream desktop/laptop CPU, Zen 3 only dual-channel.
Click to expand...

Back in the day with single channel controllers, more sticks of ram improved performance with interleaving enabled.
 
I think in either case we need to do yet another deep dive in memory speeds and cas latency along with this quad/dual setups and various SR and DR memory kits.

Ryzen has always been weird in its certification and speeds.

https://www.tomshardware.com/news/amd-ryzen-3000-cpus-memory-recommendations-asrock,39756.html

With AMD you have to make a lot of memory considerations where on an Intel platform it’s shows no noticeable difference regardless of all of the different variables.

I’d have never imagined I would have had to worry about SR vs DR sticks and wether the platform supported its speed if I’m using 4 or 2 sticks, etc.
 
Dual rank vs single rank. I'm surprised he missed that, considering it's been like that for a while now. 2x8GB usually just clocked higher so it was enough to overcome the benefits of the lower clocked 2x16GB/4x8GB kits.
 
-Strelok- said:
Dual rank vs single rank. I'm surprised he missed that, considering it's been like that for a while now. 2x8GB usually just clocked higher so it was enough to overcome the benefits of the lower clocked 2x16GB/4x8GB kits.
Click to expand...


Steve used the same kit for testing, just pulled two sticks.
 
alxlwson said:
Steve used the same kit for testing, just pulled two sticks.
Click to expand...
Yes, which would turn it from dual rank to single rank. All things equal, dual rank configurations will outperform single rank. The performance difference varies, but it has been the case for a very long time. You would achieve the same performance with 2 dual rank dimms as 4 single rank dimms if the frequency and timings are equal.
 
Last edited:
Pulsar said:
Yes, which would turn it from dual rank to single rank. All things equal, dual rank configurations will outperform single rank. The performance difference varies, but it has been the case for a very long time. You would achieve the same performance with 2 dual rank dimms as 4 single rank dimms if the frequency and timings are equal.
Click to expand...

Yeah but never a 10% uplift right? I mean Ive never seen it
 
tangoseal said:
Yeah but never a 10% uplift right? I mean Ive never seen it
Click to expand...
Depends on settings, but probably not that much. Keep in mind once you increase the resolution and settings to something you would actually use, the performance difference becomes even smaller.
 
tangoseal said:
Yeah but never a 10% uplift right? I mean Ive never seen it
Click to expand...
A 10 percent gain is in-line with going from single-channel to dual-channel, going all the way back. I doubt this has anything to do with the RAM capacity, just the RAM arrangement.
 
The dual rank issue has been around for a long time...

https://www.techpowerup.com/197145/amd-a-series-kaveri-performs-better-with-dual-rank-dimms-report

"AMD's A-Series "Kaveri" APUs perform tangibly better on system with dual-rank memory modules, according to an investigative report by ComputerBase.de. According to the report, the chip yields its best memory bandwidth when both its memory channels are populated with 2-rank DIMMs each. Either that, or four modules with one rank each."
 
Luckily, these questions will all b answered well-before we can buy a 5950x on Amazon
 
-Strelok- said:
Dual rank vs single rank. I'm surprised he missed that, considering it's been like that for a while now. 2x8GB usually just clocked higher so it was enough to overcome the benefits of the lower clocked 2x16GB/4x8GB kits.
Click to expand...
1604867519556.png
 
BlueLineSwinger said:
They tested 2x8 GB vs. 4x8 GB. A better test would have been 2x16 GB vs. 4x8 GB. I'm betting the advantage would disappear. Whatever they are testing obviously can take advantage of the extra RAM.

And IIRC the issue with four stick has never been about speed but stability at speed, and 3200 isn't that difficult a target.
Click to expand...
This is SO SLOPPY! The FIRST rule when doing comparison tests is to have only ONE variable. Now in complex systems that can be difficult or even impossible so you do the best you can. Matching the total RAM used in both tests is EASY! If you do not have 2x16 GB you wait! No one was dying to see this test. I am very disappointed in their methodology.
 
rinaldo00 said:
This is SO SLOPPY! The FIRST rule when doing comparison tests is to have only ONE variable. Now in complex systems that can be difficult or even impossible so you do the best you can. Matching the total RAM used in both tests is EASY! If you do not have 2x16 GB you wait! No one was dying to see this test. I am very disappointed in their methodology.
Click to expand...
You will change ram density by stick there if do that no ?

If you are certain to use scenario that do not really use more than 8 gig, maybe the closest comparable is same sticks different number of the,.

You cannot have just one variable that change you have too chose which variable you change, the sticks used or the amount of memory used.

Ideally you then have to do multiple test where more than one variable change and trying to do some "regression analysis?" to isolate the one that interest you.
 
rinaldo00 said:
This is SO SLOPPY! The FIRST rule when doing comparison tests is to have only ONE variable. Now in complex systems that can be difficult or even impossible so you do the best you can. Matching the total RAM used in both tests is EASY! If you do not have 2x16 GB you wait! No one was dying to see this test. I am very disappointed in their methodology.
Click to expand...

The "ONE variable" was the amount of RAM sticks. I don't think you quite realize how much difference there is between RAM kits.
 
LukeTbk said:
You will change ram density by stick there if do that no ?

If you are certain to use scenario that do not really use more than 8 gig, maybe the closest comparable is same sticks different number of the,.

You cannot have just one variable that change you have too chose which variable you change, the sticks used or the amount of memory used.

Ideally you then have to do multiple test where more than one variable change and trying to do some "regression analysis?" to isolate the one that interest you.
Click to expand...
My guess is total RAM is more of a factor than RAM density.
 
Derangel said:
The "ONE variable" was the amount of RAM sticks. I don't think you quite realize how much difference there is between RAM kits.
Click to expand...
OK fair enough, but he should also test with total RAM the same to see if that makes a difference. It looks like he is going to do it.
 
Agreed with concerns above.

You don't change multiple factors but then call out one factor in results. As I read it - the various factors (ranks of memory, size of memory) were not isolated. It may be the end result ends up the same, but I personally want to see specific contributor isolation when we publish results and conclusions.
 
I wonder if the size of the dimms matters? If bank A is 2x8 and Bank B is 2x16 - this could change how the interleaving works.
 
In all this time, I never realized there were different "ranks" of memory (i.e. single vs dual...). Maybe it was never or rarely focused on before because it did not matter much for games.

Some people above were suggesting that dual rank tends to yield better performance, but then Steve said Wendell was suggesting that 2x16gb single rank performed best on ryzen 5000...


So which is it? I guess we'll find out in coming weeks.



Edit: according to this thread, they think Steve misspoke and was talking about 2 dual rank 16gb sticks


https://www.reddit.com/r/Amd/comments/jq5u1a/comment/gbk8eht
 
tybert7 said:
Maybe it was never or rarely focused on before because it did not matter much for games.
Click to expand...
It matters about 10 percent since it was invented. But the reason why it's kinda less known today is that nobody runs single-channel unless they have to, on like, a laptop with a single slot for RAM or soldered RAM.

I'm 99 percent certain he just pulled the wrong sticks and ran all the tests, now he's in CYA mode since it's a pretty dumb mistake to make.
 
Axman said:
It matters about 10 percent since it was invented. But the reason why it's kinda less known today is that nobody runs single-channel unless they have to, on like, a laptop with a single slot for RAM or soldered RAM.

I'm 99 percent certain he just pulled the wrong sticks and ran all the tests, now he's in CYA mode since it's a pretty dumb mistake to make.
Click to expand...
Just to clarify, I've heard about single vs dual channel memory for years. Its the single vs dual rank memory that is new to me.


https://www.hardwaretimes.com/single-rank-vs-dual-rank-ram-vs-dual-channel-memory/


I could fill oceans with basic stuff I don't know.
 
Axman said:
It matters about 10 percent since it was invented. But the reason why it's kinda less known today is that nobody runs single-channel unless they have to, on like, a laptop with a single slot for RAM or soldered RAM.

I'm 99 percent certain he just pulled the wrong sticks and ran all the tests, now he's in CYA mode since it's a pretty dumb mistake to make.
Click to expand...

This isn’t about single channel or dual channel, it’s about memory ranks. Dimms are single or dual rank. Run 2 single rank dimms and you’ll have dual channel dual rank. Run 4 single rank dimms and you’ll have dual channel quad rank. Ryzen runs better with more ranks. Again it has been this way since zen2. This isn’t new. That’s the reason I went 4x8 in my current system.
 
pAiNkIlLaHvX said:
He points out towards the end of the video Wendell said 2x16GB single rank performed best of all. I think he didnt go into detail to save it for a separate video. I agree about stability at speed being the real issue.
Click to expand...
yay ..that's what I run!
 
You must log in or register to reply here.
Back
Top