DDR5 vs. DDR4 in gaming perf.

[phanbuey]

Been thinking about this comment for a few days now. Especially pertaining to DDR4 vs. DDR5 gaming performance. Chances are fairly decent DDR4 is performing nearly at it's peak and that DDR5 has not come close to realizing full potential. Also fair to say (Asus) some mobo focus on stability and usability over pushing limits except on their dedicated two DIMM OC'ing board.

Generally speaking, is there an established trend showing better in game performance after populating all four DIMM with one or both gen? Is it dependent on board or RAM channel (single/dual/quad)?

It depends on the memory controller, but essentially it's a question of ranks -- enhanced interleaving (and whatever AMD calls their version) allows the MC to access from multiple ranks faster than having to wait for two or 4 ranks to recover -- 2 stick ddr5 configurations can already have 8 ranks since they have 2 subchannels and each subchannel can have 2 ranks...

At the same speed, more ranks are faster (thanks to the interleaving) but put more stress on the memory controller -- which means you have to actually run the dimms slower to maintain stability -- so there's a tradeoff between the 10% that multiple ranks can net you, and running your dimms 10% slower to maintain stability at max ranks.

 

[nomdeplume]

It depends on the memory controller, but essentially it's a question of ranks -- enhanced interleaving (and whatever AMD calls their version) allows the MC to access from multiple ranks faster than having to wait for two or 4 ranks to recover -- 2 stick ddr5 configurations can already have 8 ranks since they have 2 subchannels and each subchannel can have 2 ranks...

At the same speed, more ranks are faster (thanks to the interleaving) but put more stress on the memory controller -- which means you have to actually run the dimms slower to maintain stability -- so there's a tradeoff between the 10% that multiple ranks can net you, and running your dimms 10% slower to maintain stability at max ranks.

Excellent synopsis of the technical limitations. Given the success Samsung SSD had implementing MC and their memory. It wasn't a complete shock B die RAM excelled.

I will still be interested to hear from quoted member on their real world experiences. Which are as considerably varied as builds across the membership here.

 

[80251]

It depends on the memory controller, but essentially it's a question of ranks -- enhanced interleaving (and whatever AMD calls their version) allows the MC to access from multiple ranks faster than having to wait for two or 4 ranks to recover -- 2 stick ddr5 configurations can already have 8 ranks since they have 2 subchannels and each subchannel can have 2 ranks...

At the same speed, more ranks are faster (thanks to the interleaving) but put more stress on the memory controller -- which means you have to actually run the dimms slower to maintain stability -- so there's a tradeoff between the 10% that multiple ranks can net you, and running your dimms 10% slower to maintain stability at max ranks.

Is the relationship between greater number of ranks and performance or frequency and performance that simple though? I read a whitepaper (on Intel westmere CPU's) when I was working on a HPC cluster on dual rank vs. single rank (DDR3 though) and the difference in performance was less than 3%. I'd figure with DDR5 it would be much more complicated too.

 

[phanbuey]

Is the relationship between greater number of ranks and performance or frequency and performance that simple though? I read a whitepaper (on Intel westmere CPU's) when I was working on a HPC cluster on dual rank vs. single rank (DDR3 though) and the difference in performance was less than 3%. I'd figure with DDR5 it would be much more complicated too.

no not at all - You're absolutely right -- it's pretty complicated and extremely memory controller/platform dependent... I like to oversimplify these because we're talking about a fairly narrow scope of hardware in context of gaming.

Zen 2 for instance, didn't really see benefits from dual ranks at all... where the 9900k did - so did all skylake/coffee lake/comet lake variants :
(2) 4x4GB vs. 2x8GB, Intel & AMD Dual-Channel Gaming Benchmark - YouTube

Zen 3 saw some of the biggest boosts with more ranks:
AMD Ryzen: 4 vs. 2 Sticks of RAM on R5 5600X for Up to 10% Better Performance - YouTube

Alder lake and intel chips still see DDR4 scaling at more sticks... (but lose frequency faster so not worth) --
2 dimms vs 4 dimms FPS BENCHMARKS - YouTube

Zen 4 does not really gain anything from additional sticks of DDR5
(2) Is 2 or 4 sticks of RAM better for Zen 4? - YouTube

Raptor lake does gain with 4 sticks of DDR5 - but YMMV and application dependent (warzone is 8% gain, but others less than 1%/margin of error)
DDR5 Dual Rank is ACTUALLY FASTER! - YouTube

I would even imagine motherboard topology has an impact on these results -- so YMMV with any given setup but if you're running zen 3 or any of the lakes generally more sticks at the same speed is slightly faster (by how much varies).

 

[Mussels]

If I use dlss I can make most games CPU-bound. Dunno about BW. Tips on how to test?

Changing the MT/s, i guess
If you had DDR5 8000, lower the MT/s by 1000 at a time and retest, then find the point of diminishing returns and so some inbetween tests
My guess is that without DLSS, it's going to be related to the bandwidth between RAM and VRAM (and maybe PCI-E, without knowing the bandwidth limits off the top of my head)

a 3060s VRAM is slower than a 3090's, so it should be happier with a lower RAM bandwidth - but more sensitive to latency, since when it does need data it needs it faster since it has less cached.

Is the relationship between greater number of ranks and performance or frequency and performance that simple though? I read a whitepaper (on Intel westmere CPU's) when I was working on a HPC cluster on dual rank vs. single rank (DDR3 though) and the difference in performance was less than 3%. I'd figure with DDR5 it would be much more complicated too.

On AM4 it's almost a free 10% performance boost to get four ranks - it changes between generations.
For whatever reason 2x2R is easier to run than 4x1R according to AMD's specs (power draw?) but the gains are real, i was really annoyed when my 2x16GB kit arrived and they were single rank for my ITX system

 

[80251]

Ranks are not the same as DIMMs per channel. For example if you have a single DPC with dual rank DIMMs it's not the same as having a single DPC with single rank DIMMs. I'm also not sure if you can have a page open on a DIMM that doesn't have its chip select activated (which would mean 4 single rank DIMMs across two channels are not the same as having two double rank DIMMs across two channels). They used to have quad rank DIMMs, but those might have only been available for servers and the performance difference of them over dual rank DIMMs was even more minimal according to the old whitepaper I once read: "WHITE PAPER FUJITSU PRIMERGY SERVERS MEMORY PERFORMANCE OF XEON 5600 (WESTMERE-EP) BASED SYSTEMS". I only read it because I was expected to brief the director on the best memory configurations for their HPC cluster nodes.

 

[phanbuey]

Ranks are not the same as DIMMs per channel. For example if you have a single DPC with dual rank DIMMs it's not the same as having a single DPC with single rank DIMMs. I'm also not sure if you can have a page open on a DIMM that doesn't have its chip select activated (which would mean 4 single rank DIMMs across two channels are not the same as having two double rank DIMMs across two channels). They used to have quad rank DIMMs, but those might have only been available for servers and the performance difference of them over dual rank DIMMs was even more minimal according to the old whitepaper I once read: "WHITE PAPER FUJITSU PRIMERGY SERVERS MEMORY PERFORMANCE OF XEON 5600 (WESTMERE-EP) BASED SYSTEMS". I only read it because I was expected to brief the director on the best memory configurations for their HPC cluster nodes.

So the performance was specific to ranks -- not dimms. In this case 2xDR = 4xSR I think HWunboxed did that comparison where they were much closer.

(1) GamersNexus / 4 sticks vs 2 sticks / Ranks : Amd (reddit.com)

 

[Mussels]

AM3+ had something like that with DDR3, where you could run higher capacity RAM that wouldnt work at all on the intel platforms of the era.

I only know the DDR4 ones off the top of my head - where TPU shows these values in their memory benchmark reviews.

You are correct with dimms per channel being one metric and ranks as the other but it seems getting four ranks from two dimms is the best middle ground, without as much difficulty reaching higher clock speeds. 2xDR is easier than 4xSR on the IMC, for whatever reason.


This being the most recent review, it's got the most comparisons too:
Acer Predator Vesta RGB DDR4-3600 2x 16 GB Review | TechPowerUp

All the 2x8 setups are single rank here, while the 2x16 and 2x32 are dual rank.
I've asked IR_cow to clarify the IF used on those insane DDR4 setups since not all could be 1:1 past 4000MT/s

The 2nd and third result are the same MT/s and timings at 3600 CL14, with an extra 7.7 %FPS just for having two extra ranks

This applies less on DDR5, but still exists.
These charts show how easy it is to make unfair comparisons in reviews and benchmarks - run two ranks on AM4 and you can make it look like absolute trash, where four ranks with slow latency or MT/s will easily overcome the 'faster' memory

 

[Cowboystrekk]

Changing the MT/s, i guess
If you had DDR5 8000, lower the MT/s by 1000 at a time and retest, then find the point of diminishing returns and so some inbetween tests
My guess is that without DLSS, it's going to be related to the bandwidth between RAM and VRAM (and maybe PCI-E, without knowing the bandwidth limits off the top of my head)

a 3060s VRAM is slower than a 3090's, so it should be happier with a lower RAM bandwidth - but more sensitive to latency, since when it does need data it needs it faster since it has less cached.


On AM4 it's almost a free 10% performance boost to get four ranks - it changes between generations.
For whatever reason 2x2R is easier to run than 4x1R according to AMD's specs (power draw?) but the gains are real, i was really annoyed when my 2x16GB kit arrived and they were single rank for my ITX system

I only have 5600X and 12400F with DDR4 atm

 

[80251]

So the performance was specific to ranks -- not dimms. In this case 2xDR = 4xSR I think HWunboxed did that comparison where they were much closer.

(1) GamersNexus / 4 sticks vs 2 sticks / Ranks : Amd (reddit.com)

Tomshardware also had a write-up like this but with more data:
https://www.tomshardware.com/reviews/amd-ryzen-3000-best-memory-timings,6310-2.html

I guess all the pages/ranks are held open even if the DIMM isn't having its chip select asserted.

Would Mussel's 2x16 GiB of single rank potentially overclock higher than 2x16GiB of dual rank? Higher frequencies reduce latency and avoid stalls when an open page/rank has to be closed to open another.

 

[chrcoluk]

Been thinking about this comment for a few days now. Especially pertaining to DDR4 vs. DDR5 gaming performance. Chances are fairly decent DDR4 is performing nearly at it's peak and that DDR5 has not come close to realizing full potential. Also fair to say (Asus) some mobo focus on stability and usability over pushing limits except on their dedicated two DIMM OC'ing board.

Generally speaking, is there an established trend showing better in game performance after populating all four DIMM with one or both gen? Is it dependent on board or RAM channel (single/dual/quad)?

I think why someone would have 4 dimms isnt for memory performance, perhaps its because it was more affordable for them to build the rig that way.

A fair amount of people start their build with less than ideal RAM to get it off the ground, then at a later date buy another 2 sticks to finish it off. A lot of my builds have often got that way.

So I know 4 dimms wont match 2 dimms in timings and clock speed ease, but its a bit of a shame that it has so little representation in the overclocking community.

I also do agree that DDR5 will improve over the coming years, but at the same time I think when a decision is made, the often its not worth buying DDR4 now as DDR5 has come in price, but I would expect the vast majority of people upgrading to already own DDR4, so its a $/£0 vs DDR5 cost, which makes it a much more difficult sell for jumping straight on DDR5. Its a shame there was no Z690/790 boards with mixed ddr4/5 dimm slots.

Ironically my best gaming performance came in a 3000CL12 configuration in a particurly stuttery game. 3000CL12 is a pretty aggressively low latency, lower than both 3200CL14 and 3600CL16. Knowing my RAM can handle that latency, I am curious if on the new system, I can do 3600CL14. If not then 3733 CL15,

Ultimately another factor in my decision is I would rather only buy DDR5 once, so buy it when its capabilities are peaked or at least much further forward than they are now.

 

[80251]

@chrcoluk
There is at least one person here who had their DDR4 running at 4133Mhz. at a command rate of 1 and CL14! I think it was a late model intel system.

I wouldn't want to have 4 DIMMs in my present build, I'm sure getting 4000+ Mhz. w/my coffeelake IMC would be impossible.

Mixed DIMM technology motherboards I think sacrificed a lot of performance to implement this.

 

[chrcoluk]

Well I ended up ordering the CPU today, Noctua I have ordered the brackets for the socket.

I dont know when the build will be done, but I will be happy to report in this thread if I try to tune the ram, and what I manage to achieve.

 

[80251]

@chrcoluk,
Did you ever try to squeeze more memory perf. out of your 9900k? I've seen 9900k's that can get 4400Mhz.+ out of their memory sticks with 32GiB but maybe that was only on z390 'boards.

 

[chrcoluk]

@chrcoluk,
Did you ever try to squeeze more memory perf. out of your 9900k? I've seen 9900k's that can get 4400Mhz.+ out of their memory sticks with 32GiB but maybe that was only on z390 'boards.

Cant remember.

So basically when I first got this z370 board (in my profile), it was with an 8600k. With 4 dimms I couldnt even hit 3200 despite everyone saying at the time intel can sleepwalk at 3200, thats what made me drop down to 3000, and then I tried to compensate and ended up running at 3000CL12.

I then installed the 9900k, and out of curiosity, tried XMP again, and suddenly 3200 was stable, so I guess either bios update improved it (as I had to update to be compatible with 9900k) or/and the 9900k has a better IMC.

At 3000CL12 I was running at over 1.4v, and I decided I wanted to go back to a lower dimm voltage, so ended up leaving it at XMP spec.

I guess I could try and do some tests prior to disconnecting, not sure if I want to push all the way to over 4000mhz though.

One of the reasons I scaled back the voltage is annoyingly my dimms dont have a temp sensor.

 

[Override17]

i have in mind this setup.
asus rog strix B760-A + i513600.
asus webpage says b760-A can run DDR4 at 5333mhz (OC).

do you recommend this 2x16gb ram kit from g-skil for mi setup?

--> https://www.amazon.com/-/es/gp/product/B08TQVC3HR/ref=ox_sc_act_title_2?smid=A3TOECTKC4OEBD&psc=1

 

[oxrufiioxo]

i have in mind this setup.
asus rog strix B760-A + i513600.
asus webpage says b760-A can run DDR4 at 5333mhz (OC).

do you recommend this 2x16gb ram kit from g-skil for mi setup?

--> https://www.amazon.com/-/es/gp/product/B08TQVC3HR/ref=ox_sc_act_title_2?smid=A3TOECTKC4OEBD&psc=1

Why would you not just go DDR5 you can get a 6000CL32 kit for that price or a 6000CL30 kit for like 8 bucks more that would be faster in gaming. You'd really need to spend 200+ on some high end DDR4 Bdie to match it. It really only makes sense to go DDR4 if you already have a decent kit and don't need to buy one.

 

[ir_cow]

This being the most recent review, it's got the most comparisons too:
Acer Predator Vesta RGB DDR4-3600 2x 16 GB Review | TechPowerUp

Most recent and probably the last DDR4 one for a while

I'm fairly confident in those results. 3 runs averaged together. Locked CPU clocks and semi controlled room temp (within a 5F). Dual-Rank pulls ahead when the game in question is bandwidth starved. BF5 is one of those games. Far Cry 5 is as well, but that game has like a 10 FPS swing in each direction per run, so I don't like using it for comparison much. Oh and this is just the average. I do not have the data for 1% lows on AM4.

 

[Shrek]

I still like the error correction with DDR5, not so much for gaming as normal work.

 

[Mussels]

Tomshardware also had a write-up like this but with more data:
https://www.tomshardware.com/reviews/amd-ryzen-3000-best-memory-timings,6310-2.html

I guess all the pages/ranks are held open even if the DIMM isn't having its chip select asserted.

Would Mussel's 2x16 GiB of single rank potentially overclock higher than 2x16GiB of dual rank? Higher frequencies reduce latency and avoid stalls when an open page/rank has to be closed to open another.

It's easier on the IF, so yes potentially it could clock higher
My 3700x says no. With my 5800x it might.

I've got three systems and rebalancing the parts is fun.
The other two are an x370 w/ GTX 1080 and a B550 w/ 3070
Theres a 5800x and 3700x to go between them. The 3700x is on the ITX B550 because of its 45W ECO mode being laughably easy to cool in ITX

(And until yesterday I only had a PCI-E 3.0 riser, so any time the BIOS was reset i had to GPU swap to get it to POST and lock it to 3.0, so tweaking RAM was a nightmare)

 

[chrcoluk]

I have already ordered my motherboard for upgrading (yes I am upgrading my platform).

I decided to stick with DDR4, time will tell if I regret it, I do have B die at least. Chose the DDR4 steel legends z690 board from asrock one of the few boards with decent connectivity.

The issue for me, is not that is it do I buy DDR4 or DDR5, but rather do I spend £0 on keeping DDR4, or buy DDR5, thats the issue, the fact I already own DDR4. The DDR5 steel legends board, z790 version had worse connectivity, unless one is bothered about a gen 5 m.2 slot.

Stock seems real low as well, I got the last board from amazon, and its amazon EU, none in UK at all. I will buy CPU bit later.

Guessing will also need to either get new cooler or beg noctua for new bits, as I got my existing noctua cooler before the socket came out. Also read about some contact frame's that improve temp's.

Just ordered a cheap arctic cooler so I can test the components, as its all still boxed up.

I am prepping a new windows ISO, backing up all my settings etc. so I can move to LTSC 2021 from 1809 before putting in the hardware, I am away a week from now, but my plan is to do the platform upgrade when I come back from that, but test the components this weekend, as its just been sitting in boxes, and if I find any of the parts dont work it will be harder to return if left too long, the board might even be out the easy return window for amazon now.

The test will be on probably the ddr4 I was using in my AMD rig (before I upgraded it to 2 x 32gigs), as those are cheap dimms, there wont be any experiments there, its just to see if it posts ok, might stick an OS on as well on a spare SSD.

 

[Mussels]

Run a live install of linux for the quick and dirty testing, straight from USB

Good way to ensure the basics work and run stress tests with no risk of OS corruption, even before storage drives arrive

 

[chrcoluk]

Run a live install of linux for the quick and dirty testing, straight from USB

Good way to ensure the basics work and run stress tests with no risk of OS corruption, even before storage drives arrive

Just ordered the thermalright contact bracket as well.

 

[chrcoluk]

This thread shouldnt be my diary so will try to keep it brief.

As mentioned before brought the components to test, they came late yesterday so waited until today, I wasnt sure about using the thermalright device, but when I undid the latch, and tested the tension to close it again, wow that was some serious pressure needed to close the thing, so decided on the spot to replace it.

It was also my first time ever doing a bios flashback, so new to that as well.

I done photos of before starting ILM replacement, and after, as well as showing the ILM parts detached.

Videos of bios flashback process.

It posted first time, and detects all 4 sticks of ram.

Given how neat and tidy the new cpu frame is, I can see intel taking note and changing this for future boards, as I think a cheap version of it could be implemented, keep the latch, but just have it close across all 4 sides of cpu, with less pressure needed as it doesnt need to exert the downward pressure outwards. The black version of it looks so neat against my board as well, shame its covered by the cooler. (they also changed 12 gen text to lga 1700 on it, to reflect is also 13 gen cpus).

But anyway, I cant really try out exotic memory configurations right now, as its using the ram I had in my second machine which I think is either hynix or micron, cant remember which but was cheap ram.

Will do some test including what mussel suggested just to make sure its working in out of the box config, and come back to this thread when the parts are in my main rig and feedback on what I achieve on B die.

 

[chrcoluk]

I have finally started on this, had to deal with bios quirks like weird vcore issues, and dram voltage droop. but currently testing 3600 on this B die, and will see whats possible on my raptor lake setup, this is probably mostly for science as I probably wont run an extreme memory profile daily. Will up the clocks after every successful ram test, and then see where I can get latency, afterwards will save profile on bios so can go back to it if people want specific benches doing.