DDR4 4400 C16 1T Gear1 vs DDR5 8000 C36 performance with 13900K and RTX 4090

bscool · Jan 17, 2023

Jimbodiah said:
You could test only the frequency aspect by using a kit set for 4000mhz and then clock it down to 3000/2000 and see what just the freq does to gaming performance. Then the same kit at 4000 with differenct latencies. I don't think I have ever seen such a test. The only comparissons on mem speeds always use different sticks, so none of the other timings are the same. Kind of making it moot in my eyes. Has GN ever done such a test? Would have seemed worth while.

Not exactly what your lookig for but 3600xmp vs tuned 4000 and 4300. PCBuilding

bscool · Jan 17, 2023

JoeRambo said:
PRO tip: always test with Large Pages enabled to reduce the noise and improve "stability" of results run to run.

What happens with 4KB pages -> CPU has 96 of DTLB entries and with 4KB pages it means when accesing more than 384kb, CPU needs to read page tables from RAM, reading random pages from RAM that are impacted by other OS level loads happening at same time.
View attachment 2594739

I just compared and it dropped it a lot using the Large Pages. 47ns vs 70ns. Might be run to run variance as in the past I was in the 64 or 66ns range when using Defaults.

bhav · Jan 17, 2023

Jimbodiah said:
Could DDR5 be hacked to run G1?

No. The 2 channels per dimm can only work in G2. It needs to run at double the IMC frequency to do its double calculations over DDR4.

Arni90 · Jan 17, 2023

Jimbodiah said:
Could DDR5 be hacked to run G1?

DDR5 could run in gear 1 if the IMC for DDR5 supported it.

Arni90 · Jan 17, 2023

bhav said:
No. The 2 channels per dimm can only work in G2. It needs to run at double the IMC frequency to do its double calculations over DDR4.

That's just wrong, it has nothing to do with 2 channels per DIMM.

bhav · Jan 17, 2023

Arni90 said:
That's just wrong, it has nothing to do with 2 channels per DIMM.

So you can set DDR5 to 4000 G1 then? I don't think you can, and it isnt due to IMC limitation.

Arni90 · Jan 17, 2023

bhav said:
So you can set DDR5 to 4000 G1 then? I don't think you can, and it isnt due to IMC limitation.

Ah yes, I forgot you are an expert, despite not showing the faintest shred of evidence of having even run a single benchmark more complex than AIDA 64

bhav · Jan 17, 2023

Arni90 said:
Ah yes, I forgot you are an expert, despite not showing the faintest shred of evidence of having even run a single benchmark more complex than AIDA 64

Thats funny given that I dont even post Aida scores, and you've clearly already seen my microbench results.

Always full of crap aren't you? DDR5 cannot work in G1 due to it being designed to work in G2, its a technical limitation, not an IMC limitation.

Anyhow:

Rectangle Font Parallel Screenshot Circle

Also you only started ranting about Aida after I told you that its useless in the first place.

X909 · Jan 17, 2023

G1 for DDR5 isnt implemented in the microcode as its practically useless. IMC is rated for 1600 Mhz, this translates to 3200 Mhz DDR speed. DDR5 3200 is not existtent.
G1 vs. G2 vs. G4 is just a clocking thing. The IMC runs at max. 1600 Mhz normally and can overclock to 2000-2100 on most CPUs. Thats useless for DDR5.

And DDR5 scales well... if you have bandwidth bound scenarios or heavy multithreading. You need to utilize the 4 channels. Otherwise, DDR4 can compete. 1 DDR4 channel (64 Bit) is faster than 1 DDR5 channel (32 Bit) in terms of latency and also in terms of throughput. DDR5 is made for parallelism...

bhav · Jan 17, 2023

X909 said:
G1 for DDR5 isnt implemented in the microcode as its practically useless. IMC is rated for 1600 Mhz, this translates to 3200 Mhz DDR speed. DDR5 3200 is not existtent.
G1 vs. G2 vs. G4 is just a clocking thing. The IMC runs at max. 1600 Mhz normally and can overclock to 2000-2100 on most CPUs. Thats useless for DDR5.

And DDR5 scales well... if you have bandwidth bound scenarios or heavy multithreading. You need to utilize the 4 channels. Otherwise, DDR4 can compete. 1 DDR4 channel (64 Bit) is faster than 1 DDR5 channel (32 Bit) in terms of latency and also in terms of throughput. DDR5 is made for parallelism...

Actually you raised a good point there, technically the specification for DDR4 is only supposed to go up to 3200, and for DDR5 the IMC equivalent for that is 6400.

Anything above that is ofc overclocking and running out of spec, though ofc thats what we all do here.

So regardless of whether G1 would work or not for DDR5 is irrelevant, all we can see is that all these bandwidth based improvements are having a minimal impact on performance, also if comparing strictly XMP, then the improvement with DDR5 is basically zero.

You only see any noticeable difference between DDR4 and DDR5 when tuning them to the maximum, and thats coming from the timings and not the frequency.

If the frequency mattered at all, then I shouldn't be beating 6800 XMP with my overclocked DDR4 at just 4300.

Arni90 · Jan 17, 2023

bhav said:
Thats funny given that I dont even post Aida scores, and you've clearly already seen my microbench results.

Always full of crap aren't you? DDR5 cannot work in G1 due to it being designed to work in G2, its a technical limitation, not an IMC limitation.

If DDR5 was unable to work in gear 1, why is AMD running their IMC at exactly the same frequency as the memory on AM5?

bhav · Jan 18, 2023

Arni90 said:
If DDR5 was unable to work in gear 1, why is AMD running their IMC at exactly the same frequency as the memory on AM5?

Might be because theres also FCLK which runs at half instead.

I don't think gears on Intel are comparable to however AMD handles their FCLK and IMC. If it was running in the equivalent of 1:1 on Intel, the 6000 DDR5 latency on AMD should be even lower than any DDR4 setup, from what I can find the memory latency on Ryzen 7000 is around 60+ns.

MisterSheikh · Jan 18, 2023

Goddamn, didn’t know I ignited a war in here 😂.

Arni, saw your results, pretty good findings. Yea I kinda ****ed up by not having a completely cpu bound scenario but I think I explained why that was the case.

Currently busy as hell with real life stuff so I’ll get back to you with the 3200 xmp vs 4400 c16 1T tuned profile findings.

I’ll reply in more detail later on regarding some of the points you mentioned. But to briefly summarize, not enough of a notable difference between fast tuned of ddr4 and ddr5 that were achievable by my specific 13900K and the motherboards I have purchased to justify a significant price premium for negligible performance differences, I would also be decreasing my memory capacity by half. I think your results showed fast tuned ddr4 being able to hang with fast tuned ddr5, being a few percentage off.

My IMC is capable of doing 4400 whereas yours struggles with 4200. In contrast, mine struggles with 8000 MHz ddr5. This test was simply to compare what performance I can get with my hardware using different setups in games that I play, and is there enough of a difference to justify a hefty price increase and a 32GB decrease in capacity, to which I concluded no.

Some here may pay hundreds more for minute performance increases and all power to them, but I’m not one of those individuals.

Arni90 · Jan 18, 2023

MisterSheikh said:
Goddamn, didn’t know I ignited a war in here 😂.

Arni, saw your results, pretty good findings. Yea I kinda ****ed up by not having a completely cpu bound scenario but I think I explained why that was the case.

Currently busy as hell with real life stuff so I’ll get back to you with the 3200 xmp vs 4400 c16 1T tuned profile findings.

I’ll reply in more detail later on regarding some of the points you mentioned. But to briefly summarize, not enough of a notable difference between fast tuned of ddr4 and ddr5 that were achievable by my specific 13900K and the motherboards I have purchased to justify a significant price premium for negligible performance differences, I would also be decreasing my memory capacity by half. I think your results showed fast tuned ddr4 being able to hang with fast tuned ddr5, being a few percentage off.

My IMC is capable of doing 4400 whereas yours struggles with 4200. In contrast, mine struggles with 8000 MHz ddr5. This test was simply to compare what performance I can get with my hardware using different setups in games that I play, and is there enough of a difference to justify a hefty price increase and a 32GB decrease in capacity, to which I concluded no.

Some here may pay hundreds more for minute performance increases and all power to them, but I’m not one of those individuals.

You could run the same Factorio bench as I did: FactorioBox Maps

Factorio is very memory-dependent, and extremely single-threaded, leading to at least B-die outperforming my current DDR5 setup. It also barely reacts to increasing clock speeds on the CPU.

I have no issue with people wanting to save money, DDR4 can after all get pretty close in most gaming scenarios. However, just because DDR4 can get close doesn't mean it's faster.

domdtxdissar · Jan 18, 2023

bhav said:
Might be because theres also FCLK which runs at half instead.

I don't think gears on Intel are comparable to however AMD handles their FCLK and IMC. If it was running in the equivalent of 1:1 on Intel, the 6000 DDR5 latency on AMD should be even lower than any DDR4 setup, from what I can find the memory latency on Ryzen 7000 is around 60+ns.

🤦‍♂️

AMD dont run FCLK at half the MCLK/UCLK, recommend ratio is 2:3:3
Memory controller run 1:1 with memory, what you would call "gear 1" on Intel platform.
And i dont know where you are getting this ~60ns from, if your talking about useless Aida numbers.. Should get around 55ns @ 6000MT/s 2000mhz FCLK.

At my 24/7 settings i'm getting 51.4ns fantasy points, which is not so bad considering offdie memory controller.. 7000's APU should offer very good memory performance/clocking..

Font Screenshot Technology Software Electronic device

bhav · Jan 18, 2023

Arni90 said:
You could run the same Factorio bench as I did: FactorioBox Maps

Factorio is very memory-dependent, and extremely single-threaded, leading to at least B-die outperforming my current DDR5 setup. It also barely reacts to increasing clock speeds on the CPU.

I have no issue with people wanting to save money, DDR4 can after all get pretty close in most gaming scenarios. However, just because DDR4 can get close doesn't mean it's faster.

You seem to clearly take huge issue with my use of Micron B die, and not once have I ever claimed it is faster, nor have I seen that claim made by anyone on this site.

bhav · Jan 18, 2023

domdtxdissar said:
🤦‍♂️

AMD dont run FCLK at half the MCLK/UCLK, recommend ratio is 2:3:3
Memory controller run 1:1 with memory, what you would call "gear 1" on Intel platform.
And i dont know where you are getting this ~60ns from, if your talking about useless Aida numbers.. Should get around 55ns @ 6000MT/s 2000mhz FCLK.

At my 24/7 settings i'm getting 51.4ns fantasy points, which is not so bad considering offdie memory controller.. 7000's APU should offer very good memory performance/clocking..
View attachment 2594952
View attachment 2594953

Well that is surprising, and you mean to tell me not one individual has yet tested 1:1 IMC:Ram to 2:1 on AMD with DDR5 to see how much a performance difference / loss there is running at such settings as G2?

Literally every single tech news source you find goes on and on about how DDR5 is designed for gear 2, gains no benefit from gear 1, needs to run 2:1 to the IMC.

The latency numbers I gave were taken from however many articles written about some Roberto guy from AMD's info on how to run DDR5 on Ryzen 7000 series.

I have never once posted anything positive about Aida on this forum and never even use it, the only one time I did use it was when I suspected IMLC was giving incorrect bandwidth reads and confirmed as such. Both are however simply a quick test to make sure your bandwidth and latency are not dropping when ram overclocking, which they were doing for me as it turned out due to the auto settings on tertiaries.

That however doesn't mean that latency isnt important, you can clearly see its advantage in a few titles like Factorio, Start Citizen, and Anno 1800, which as it turns out are the games I play / am interested in (Not paying pledge prices for star citizen, just no idea how long until it releases).

For the games I actually play, latency matters the most, and I still apparently have the best latency result in the microbench competition for Intel 12th / 13th gen on this forum.

Furthermore if latency didn't matter at all for gaming, 4000CL14 G1 would not be outperforming 4800CL17 G2 in Civ 6 turn times. With most things regarding ram, there is a 'sweet spot' for performance. To get to the sweet spot I require for latency I would need to be running 8000+ DDR5 tuned, and have no interest in wasting that much money on an overpriced Asus Apex and A die kit.

For other games, who even cares about 1080p low results when everything they do is going to be GPU limited? I run at 4K, everything maxed out or as close to maxed out as I can get on my 3080 Ti. Why waste double the money for a 1-2 FPS improvement at most?

bhav · Jan 18, 2023

Regarding 'sweet spots', there are very few reviews like this one that compare ram speeds starting at 2133 or even 2400 and up to DDR5 6000+:

https://www.tomshardware.com/features/ddr5-vs-ddr4-is-it-time-to-upgrade-your-ram

You can see in the gaming results, even at 1080p that going from 2133 to 3600 gives a much more noticable improvement than going anywhere above 3600.

An explanation for why that might be, I found this post on reddit:

Many system activities (such as gaming) are wholly GPU bound or otherwise not limited by interaction with memory. With dual channel 2933mhz memory at CL16, memory bandwidth and latecy speeds are wide open and the bottleneck in your system is somewhere else (observable by the relative parity in memory configurations at this spec and above, and the harsh performance drops below this spec as memory interaction begins to bottleneck the system). This wide open status was realized while using an i9-12900 and Radeon RX 6900XT which are top of the line components, ergo memory should be wide open at this spec for any consumer hardware configuration today. I wrote a little more about memory bottleneck here.

Now for strictly gaming performance, there is very little need to go above 3600CL16, people who overclock their ram simply do to squeeze out however much tiny extra performance it might give.

Why spend £320 on Samsung B die over £150 on Micron E die, or £600 or however much it was on launch for DDR5 6000+ (4800 was £400) over £180 for micron B die and simply overclock the cheaper ones?.

Likewise, why even spend £150 on hynix M die instead of £90 on Kingston DJR for a 2% at most performance difference? Also in any case where the higher frequency ram is better, its at most 2%.

In cases like the ones I mentioned where latecy matters, it can be something like 10-20% more performance with DDR4. Facotio I heard I straight up unplayable without overclocked DDR4, but an explanation for this was something to do with syncing to the server needing to match your FPS which is heavily latency dependant, so just playing it in single player doesn't have that issue.

In short, lower latency has a far larger impact when it matters than frequency ever has. Further and evident proof of this is when simply comparing XMP, or even JEDEC spec settings between DDR4 and DDR5.

As I mentioned, for the most part DDR5 still sits at a meme price for its tiny performance gains. A relatively cheap option with Hynix M die is ok, but you can still get a Hynix DJR setup for far less and not get anymore than 2% less performance even at 1080p for gaming.

If frequency mattered at all, surely even 4800 DDR5 at XMP should be beating my 4300G1 no? How about 6000? 6800? How come as it turns out at XMP vs XMP theres pretty much zero difference going above 3600 XMP DDR4?

Arni90 · Jan 18, 2023

bhav said:
Well that is surprising, and you mean to tell me not one individual has yet tested 1:1 IMC:Ram to 2:1 on AMD with DDR5 to see how much a performance difference / loss there is running at such settings as G2?

Literally every single tech news source you find goes on and on about how DDR5 is designed for gear 2, gains no benefit from gear 1, needs to run 2:1 to the IMC.

The latency numbers I gave were taken from however many articles written about some Roberto guy from AMD's info on how to run DDR5 on Ryzen 7000 series.

Has anyone ever told you that we have two ears and one mouth for a reason before now?
People have tested it.

The big Ryzen 7000 Memory and OC Tuning Guide - Infinity Fabric, EXPO, Dual-Rank, Samsung and Hynix DDR5 in Practice test with Benchmarks and Recommendations | Page 5 | igor´sLAB

AMD’s new Ryzen 7000 desktop CPUs, based on the Zen4 micro-architecture, still use the same chiplet design as their predecessors, with a few small but not negligible changes. Igor already gave details…

www.igorslab.de

bhav · Jan 18, 2023

Arni90 said:
Has anyone ever told you that we have two ears and one mouth for a reason before now?
People have tested it.

The big Ryzen 7000 Memory and OC Tuning Guide - Infinity Fabric, EXPO, Dual-Rank, Samsung and Hynix DDR5 in Practice test with Benchmarks and Recommendations | Page 5 | igor´sLAB

AMD’s new Ryzen 7000 desktop CPUs, based on the Zen4 micro-architecture, still use the same chiplet design as their predecessors, with a few small but not negligible changes. Igor already gave details…

www.igorslab.de

So you can see from that how much lower the setup on the MSI board with 1500 UCLK performs then, nice to know that. I actually already asked a while ago in one of the AMD threads if Ryzen 7000 runs DDR5 in 1:1 or 1:2 and no one could answer, so don't assume that how new platforms function is common knowledge.

So then, why would anyone specifically want to run DDR5 G2 over DDR4 G1 given that you get an improvement of at most 2%? And then what happens if you do want to Run Factorio, Star Citizen, Anno 1800 or any other game that scales with latency?

If frequency is more important than latency which is your claim, why doesn't 4800 G2 DDR5 outperform 4300 G1 DDR4? If Latency doesn't do anything, why does tuning the timings on ram increase performance much more than simply raising the frequency?

For bandwidth limited scenarios in gaming, your memory frequency only needs to be fast enough to not be the bottleneck, also referred to as its sweet spot, this is still more than achieved with 3600CL16 DDR4.

DDR4 4400 C16 1T Gear1 vs DDR5 8000 C36 performance with 13900K and RTX 4090

bscool

bscool

Attachments

bhav

Arni90

Arni90

bhav

Arni90

bhav

X909

bhav

Arni90

bhav

MisterSheikh

Arni90

domdtxdissar

bhav

bhav

bhav

Arni90

The big Ryzen 7000 Memory and OC Tuning Guide - Infinity Fabric, EXPO, Dual-Rank, Samsung and Hynix DDR5 in Practice test with Benchmarks and Recommendations | Page 5 | igor´sLAB

bhav

The big Ryzen 7000 Memory and OC Tuning Guide - Infinity Fabric, EXPO, Dual-Rank, Samsung and Hynix DDR5 in Practice test with Benchmarks and Recommendations | Page 5 | igor´sLAB

Top Contributors this Month

Recommended Communities