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

This isn’t about single channel or dual channel, it’s about memory ranks.
Wait, I didn't think he swapped the modules, just pulled two out. That won't change the memory's ranks.

Unless somehow it doesn't like more ranks in more channels; that could explain the difference.
 
Wait, I didn't think he swapped the modules, just pulled two out. That won't change the memory's ranks.

Unless somehow it doesn't like more ranks in more channels; that could explain the difference.
2 single rank vs 4 single rank.
 
Wait, I didn't think he swapped the modules, just pulled two out. That won't change the memory's ranks.

Unless somehow it doesn't like more ranks in more channels; that could explain the difference.
2 single rank dimms, 1 dimm per channel = single rank. 4 single rank dimms, 2 dimms per channel = dual rank. He removed 1 dimm from each channel, making it single rank.
 
When I built my 2700x system it was known that dual rank was faster at the same speed but the fastest dual rank kits were rated for much lower speed(2400Mhz iirc) so they were slower overall.

I was wondering about this when I was looking at memory recently and saw dual rank 32GB kits that were rated for high fairly speeds, it almost seems like it was an obvious conclusion that got lost in the shuffle of changes.
 
When I built my 2700x system it was known that dual rank was faster at the same speed but the fastest dual rank kits were rated for much lower speed(2400Mhz iirc) so they were slower overall.

I was wondering about this when I was looking at memory recently and saw dual rank 32GB kits that were rated for high fairly speeds, it almost seems like it was an obvious conclusion that got lost in the shuffle of changes.
Really? I built a 1600 system at launch with a 2x 16gb 3000 dual rank kit
 
When I built my 2700x system it was known that dual rank was faster at the same speed but the fastest dual rank kits were rated for much lower speed(2400Mhz iirc) so they were slower overall.

I was wondering about this when I was looking at memory recently and saw dual rank 32GB kits that were rated for high fairly speeds, it almost seems like it was an obvious conclusion that got lost in the shuffle of changes.
I bought a 2x16GB rated at 3200 CL16. I have it running at 3433 CL14. I don’t think I’ve seen any kits that low, the lowest I saw was 3000 CL15.
 
People talking about multiple things changed. ONE thing was changed, the quantity of memory sticks i.e the number of populated memory channels. The ranks do not ever change in a dimm, the ranks stayed the same, and the timings stayed the same.

Testing 2x16 vs 4x8 would be very tedious to find chips with the exact same timings. And you would want all of the chips to be single ranked. Then you could test the impact of more channels, which is the reason for the better performance.
An interesting test would be 2 dual rank 16's vs 4 single rank 8's, where the speeds and timings can be made exact. In this case, I still believe the 4 sticks will perform better.

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.
No. Ranks are how each stick is engineered. Yes, 2 duals are used by the cpu in the same way as 4 singles. But only 1 rank inside a dimm can be accessed at a time, so as you go to dual and quad rank dimms, your overall memory performance may be affected, especially with fewer channels to access them.

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.
Dimms are single, dual, or quad ranked.
2 single rank dimms is dual channel single rank.
4 single rank dimms is quad channel single rank.

More channels = more performance
More ranks =! more performance, it equals more RAM capacity.
 
Dimms are single, dual, or quad ranked.
2 single rank dimms is dual channel single rank.
4 single rank dimms is quad channel single rank.

More channels = more performance
More ranks =! more performance, it equals more RAM capacity.
assuming you have a two channel memory controller like ryzen on the am4

2 single rank dimms is dual channel single rank.
4 single rank dimms is dual channel dual rank.
 
AMD is limited to dual channel? That seriously sucks... Intel has had quad channel since x99, and my i7-920 was triple channel and that was what, 2009?

Triple channel = triple the bandwidth.

Ahh, only Threadrippers support quad channel.. https://en.wikipedia.org/wiki/Multi-channel_memory_architecture
So yea, that kinda sucks for desktop cpu's.
It's the same for both AMD and Intel. The desktop lines get dual channel and the HEDT lines get triple or better.
 
i7-920 wasn't high end desktop as far as I remember.
x99 was workstation class chipset, which is on my asus rampage V edition 10 mobo that I am still running. It supports quad channel. This 6 core 12 thread cpu was ~$550 new back in fall of 2016.

Can't think of any real reason AMD couldn't be doing quad channel memory controller on their cpu's in these price ranges.
 
Can't think of any real reason AMD couldn't be doing quad channel memory controller on their cpu's in these price ranges.
Same can be said for Intel.

It probably isn't worth it, and it would also probably require a different socket.
 
i7-920 wasn't high end desktop as far as I remember.
x99 was workstation class chipset, which is on my asus rampage V edition 10 mobo that I am still running. It supports quad channel. This 6 core 12 thread cpu was ~$550 new back in fall of 2016.

Can't think of any real reason AMD couldn't be doing quad channel memory controller on their cpu's in these price ranges.
x58 was basically the beginning of HEDT and they had dual channel boards alongside x58.
Also a X99 platform may theoretically have double the RAM bandwidth of a 5950x platform, but does it perform better in any workload besides a RAM bandwidth benchmark?

Also the Threadripper Pro series supports 8 Channels..... https://www.amd.com/en/processors/ryzen-threadripper-pro
 
Can't think of any real reason AMD couldn't be doing quad channel memory controller on their cpu's in these price ranges.
Cost, complexity of I/O die, and differentiation from HEDT (Threadripper) are the most likely reasons. For most consumer usage scenarios quad channel memory isn't needed as most common consumer usage scenarios are not main memory bandwidth-limited. Case in point: Ryzen 3 5000 series of CPUs. They are utilizing the same I/O die as Zen 2 which means the same dual-channel memory controller, yet saw 15%+ gains in performance in many applications compared to Zen 2. The limitations of Zen 2 were obviously not because of main memory bandwidth or its use of dual channel.
And your X99 platform system isn't much different in terms of speed in most common consumer software usage scenarios like games compared to Intel's dual channel memory systems. It's not like quad channel suddenly made dual channel obsolete. Sure, you can demonstrate and bench differences, but in the real world common applications it's just not that different and oftentimes the cost difference between a quad channel system and a dual channel system isn't worth the increased performance.
Heck, Ryzen's memory write performance is half of its read performance and even that doesn't affect most usage scenarios in a meaningful way. (It sure surprised me when I learned about that)
 
Yeah, not planning to upgrade for at least a year and a half.

I'm sure whatever I get will be faster, but if it's going down to dual channel ram it still feels like a downgrade.
 
People talking about multiple things changed. ONE thing was changed, the quantity of memory sticks i.e the number of populated memory channels. The ranks do not ever change in a dimm, the ranks stayed the same, and the timings stayed the same.

Testing 2x16 vs 4x8 would be very tedious to find chips with the exact same timings. And you would want all of the chips to be single ranked. Then you could test the impact of more channels, which is the reason for the better performance.
An interesting test would be 2 dual rank 16's vs 4 single rank 8's, where the speeds and timings can be made exact. In this case, I still believe the 4 sticks will perform better.


No. Ranks are how each stick is engineered. Yes, 2 duals are used by the cpu in the same way as 4 singles. But only 1 rank inside a dimm can be accessed at a time, so as you go to dual and quad rank dimms, your overall memory performance may be affected, especially with fewer channels to access them.


Dimms are single, dual, or quad ranked.
2 single rank dimms is dual channel single rank.
4 single rank dimms is quad channel single rank.

More channels = more performance
More ranks =! more performance, it equals more RAM capacity.


No, the number of memory ranks relating to more performance in this case is because of bank interleaving.
 
AMD has indicated? Flat-out confirmed? that Zen 3 uses the same memory controller as Zen 2. So this should be easy for lots of people to confirm.

The reason why I mentioned single-channel versus dual-channel, not rank, is that it amounts to about a 10 percent performance difference. If ram use isn't in play; capacity isn't a factor, then dual-channel memory will outperform single-channel memory by about 10 percent.
 
AMD has indicated? Flat-out confirmed? that Zen 3 uses the same memory controller as Zen 2. So this should be easy for lots of people to confirm.

The reason why I mentioned single-channel versus dual-channel, not rank, is that it amounts to about a 10 percent performance difference. If ram use isn't in play; capacity isn't a factor, then dual-channel memory will outperform single-channel memory by about 10 percent.


that isnt what was tested in the video. both setups were dual channel. the difference was 2x8gb (dual channel, single rank) and 4x8gb (dual channel, dual rank).
 
I can visit this topic myself if I can ever get a 5000 series CPU in. The MSI MEG X570 GODLIKE I'm running can do 4x16GB at DDR4 3600MHz. I can literally remove two modules or go up to four using the same exact modules. Nothing I'd run would use the extra physical RAM, and even two modules would still be 32GB.
 
That's because you value epeen more than actual performance.
LOL.

Yeah I go around bragging about having quad channel memory on my 4 year old system...

Try actually contributing to the thread.
 
i7-920 wasn't high end desktop as far as I remember.

It was. Socket 1156 was the mainstream consumer equivalent, with dual-channel DDR3, and CPUs such Lynnfield.

Yeah, not planning to upgrade for at least a year and a half.

I'm sure whatever I get will be faster, but if it's going down to dual channel ram it still feels like a downgrade.

I went from X99+5820K to X570+3900X and it was definitely not a downgrade, in any respect. Though since I had my 5820k @ 4.5Ghz, the 3900X didn't really feel like much of an improvement in terms of single-core IPC. I probably would have stuck with the X99 setup for a bit longer if it had not started to develop some quirky issues. The best part of the upgrade was doubling the number of cores while simultaneously halving my power consumption, heat-output, and average AIO fan noise. At 4.5Ghz the 5820k is a furnace.

Do you have a timestamp? Because I never saw that in the video.

If he was running 2 dimms and they were in the correct slots, why wouldn't it be running dual channel? X570 is a dual-channel platform.

It feels like a lot of people in this thread are confusing the discussion about single-rank and dual-rank with single-channel and dual-channel.
 
If he was running 2 dimms and they were in the correct slots, why wouldn't it be running dual channel? X570 is a dual-channel platform.
I'm suggesting he was running them in the incorrect slots. That would explain a difference, since from what we know so far, the memory controller is the same.
 
I'm suggesting he was running them in the incorrect slots. That would explain a difference, since from what we know so far, the memory controller is the same.

im sure tech Jesus had them in the right slots

like what was said above, this is a single/dual rank testing. Not single/dual channel test.
 
Or they bought a Dell or HP and didn't know any better, which, as a very recent LTT video shows, still happens, because those idiots are still shipping systems that way by default.
Ugggghhhh. I have given up trying to let the higher ups know that we should be ordering systems with 2 sticks of RAM.

I just order more RAM after the computers arrive and upgrade them.

Idiots at Dell.... and people in charge of what computers we order not having a clue.
 
Ugggghhhh. I have given up trying to let the higher ups know that we should be ordering systems with 2 sticks of RAM.

I just order more RAM after the computers arrive and upgrade them.

Idiots at Dell.... and people in charge of what computers we order not having a clue.
Yeah, me too.

On the positive side, I can usually drop in a second stick for like half of whatever Dell's upgrade price is. I do the same thing with storage.
 
very true. But with the core to core latency on ryzen it has a larger impact on performance.
Bingo. It's because Ryzen is much more sensitive to memory performance that you are seeing a larger difference.
 
Bingo. It's because Ryzen is much more sensitive to memory performance that you are seeing a larger difference.

Ryzen is "more sensitive to memory performance" because the IF clock has been tied to memory speed. Increasing the IF clock decreases the ping time between the cores in different CCXs because they use the infinity fabric to communicate. This doesn't have anything to do with the IF clock. Bank interleaving gives a small increase to performance, and the rest is accounted for by the larger memory size in this poorly-designed test. The memory controller in Vermeer is identical to that used with Matisse, and this scenario has already been investigated by multiple outlets. There's nothing special here about Vermeer and four sticks of RAM over previous generations.

Here's a link to another example of this testing: https://www.techspot.com/article/1971-more-ram-modules-better-for-gaming/ (Note that the 9900K used in these tests also benefits, often to a larger degree than the 3900X.)
 
I bought a 2x16GB rated at 3200 CL16. I have it running at 3433 CL14. I don’t think I’ve seen any kits that low, the lowest I saw was 3000 CL15.
Lol, I guess I don't recall the specifics correctly. I know there wasn't anything as fast as the 3600 c15 kit I got though and there was memory rated for at least 4000 Mhz(not that you could hit that speed on Ryzen at the time).
 
My guess is total RAM is more of a factor than RAM density.
No!
No one is getting these variables in going from 16 to 32 GB ram. If that was all we would have endless examples and everyone would be screaming for bigger kits as an enthusiast must have. I thought it was common knowledge all ready that 2 sticks dual rank was superior. 4 sticks single will get you there but might not get you the clocks.
GN is just rehashing old news because Zen3.
 
No!
No one is getting these variables in going from 16 to 32 GB ram. If that was all we would have endless examples and everyone would be screaming for bigger kits as an enthusiast must have. I thought it was common knowledge all ready that 2 sticks dual rank was superior. 4 sticks single will get you there but might not get you the clocks.
GN is just rehashing old news because Zen3.

This was common knowledge, but GN like to create 25 minutes videos to regurgitate known facts. Or they make a video just to declare everyone else is wrong over pedantic issues, or complain about RGB.

This is one of the worst videos GN has put out. I easily figured out 2x16 dual rank sticks was better than 2x8 or 4x8 last year when I built my PC. The info was all over the place.
 
This was common knowledge, but GN like to create 25 minutes videos to regurgitate known facts. Or they make a video just to declare everyone else is wrong over pedantic issues, or complain about RGB.

This is one of the worst videos GN has put out. I easily figured out 2x16 dual rank sticks was better than 2x8 or 4x8 last year when I built my PC. The info was all over the place.
Sure, but #'s never have shown 10% in the past. Now the falicey that it was all 16 to 32 GB's doing is what I am face palming. A dislike for GN style? Sure. But let's not ride the idiot bus to stupid town.
 
Back
Top