DDR5-6000 Memory is the Sweet Spot For AMD Ryzen 7000 Zen 4 CPUs

Zen3 can do 2000Mhz if you have 4000Mhz ram and you don't try to overclock any lesser ram.(5900X with 4000Mhz ram using atm)
Zen3 found in APUs can do 2500Mhz Infinity Fabric (DDR4-5000 RAM is possible on a 5700G).
Zen3+ (Ryzen 6000) does 2400Mhz Infinity Fabric (DDR5-4800, when we get faster SODDIMs we should know if can be pushed). We don't know the Infinity Fabric speed when LPDDR5-6400Mhz is been used. (is supported).

Zen2 could do 1900-1933Mhz Infinity Fabric (my 3900X used the old 3600C16 kit at 3800C14)
Zen1 could do 1800mhz Infinity Fabric. (my 1800X run with 3600C16 ram on 1:1 since April 2017)

 

I'm really interested in the Unreal Engine performance problems, do you have any link to a video or article were i can learn more about it?

 

modern cpus have so much cache that the performance penalty for the extra latency is really minimal if at all present, There are edge cases where it can matter, but generally it doesn't, DDR5 does more simultaneous memory accesses than ddr4 (more channels and more bank groups), which helps mitigate the increased latencies, since you can access a different memory while memory is busy. the maximum delay can actually be lower in some cases, even if the minimum delay is larger.

 

Memory Latency (ns) and total bandwidth matter only.

Gskill DDR5-6400 CL32 is probably the fastest ram we will see for a long time and has roughly the same latency with 3600C16


(It can OC to 7235 CL32 which means 117Gb/s Read, 112GB/s Write and 47.6ns. But that is pushing a 12900K to it's limits.)

 

What UE perf problems you talking about?

 

Memory latency matters even at 4K but rarely if ever will you see a gaming benchmark that charts memory based on some achieved latency verified by something like AIDA64. A lot of these higher speed memory kits with low latency on DDR4 had high tRCF values that impacted actual latencies fairly significantly.

 

Exactly. Also calculated latency vs actual latency gets overlooked by most of us. There are things like really loose tRFC values that will make actual latencies suck but can make a lesser chip hit a lower CAS which looks good on paper and the vendor can charge more. This is actually a very common thing memory vendors do.

 

Wauw like Intel i have Intel 12 Gen with DDR5 5200 mhz cas 36-36-36-84 OC to 6000 mhz with cas 34-34-34-75 T1..

 

U got an extremely good 1800x my 1600 wouldn't even go over 3133C16 with a C14 flarex kit and my 2400g couldn't get it stable past 3333C16 on flarex as well

 

There are few benchmarks around but the ones I've seen show that AMD isn't bottlenecked at all by average RAM. On Intel, you need to be pushing min framerates of well over 100 in order to start seeing any difference with super fast DDR4. Even then, 5% is all you can expect to acheive by spending twice what some 3600 CL18 will cost you.

eg https://www.pcgamer.com/does-ram-speed-matter-gaming-amd-intel/

 

running in 1:2 (gear 2) mode

 

We have to consider that DDR5 is at the start. Speed and timings are very likely to improve in the next few years.

That is a really bad test.
Tweaking memory on zen is not just about clock speed. Tweaking timmings is as important as increasing clocks.
And why are they testing at 4K? Everyone know that at 4K, the GPU is the major bottleneck.
And testing with IF and memory not at 1:1 ratio? Seriously.

 

I've not used AMD for years but they test both frequency and latency. They also test both at 1080p and 4K.

 

I thought Zen 2 sweet spot was 3733MHz - 3800MHz RAM/1900MHz IF was outside spec and considered an OC?
Same for Zen 3, being 3800MHz - 4000MHz RAM/2000MHz IF also outside spec and considered an OC

 

As you could see I'm asking to Zeblote because he said "especially in Unreal Engine games that spend lots of time chasing pointers.", I have heard before that Unreal engine have performance problems (stutters and stuff, nothing i have experienced) and I work a lot with Unreal so I think it could be interesting to learn what kind of performance issues does Unreal could have.

 

Ive read a few places that the CPUs are ready to go now so id imagine there is ample time for the RAM and MOBO manufactures to deconflict.

 

I guess this also means that REALLY high memory clocks will be doable on zen4 in 1:2 mode, >10000mt/s potentially.

 

What was the actual Ram kit DOOP? Mine was 3600C16 and worked perfectly fine. If you tried to use a lesser kit and overclock it wasn't working.

 

You can run Zen 2 and Zen 3 on async mode. On Zen 2 it will add 6ns latency but you can use 5000Mhz or higher DDR4 ram.

A 5700g with 4733 14-14-14-28-1T RAM has 44ns latency wth 3:71 ratio and almost 70GB/s read. Which is insane. Thats Intel territory and is asynchronous.

 

When I say 1:2 I refer to the fact that zen 2/3 will drop the uclk(memory controller speed) to 1/2 rate in particular circumstances, this happens whenever the fabric speed does not match the memclk, or when the memclk exceeds the predefined limit (3600mt/s on zen2/3 and 4000-4400 mt/s on zen 3 apus depending on the board), you can of course force it to stay in 1:1 mode , but this is the default behaviour.

In async mode with the 1/2 uclk , you can greatly increase the memory speed (>5000mt/s), but it doesn't get you much because the data fabric is limited by its 256bit width,Ie if you set an fclk of 2000, it nets you a theoretical max bandwidth of 64gb/s and no matter how high you set the memory speed it will never exceed this.

This is where zen4 presumably differs, Since it supports AVX-512 and they have a nice die shrink, there is a pretty good chance that they have widened the data path to 512bit, Which means the required frequency to achieve the same bandwidth is half, this is what intel has been doing since icelake(maybe technically cannonlake), and is why they can achieve >100gb/s despite having a low ring clock or mesh speed.

The implication is that if the data path is 512bit, and the max fclk is 3000mhz, that would net you a max achievable bandwidth of 192gb/s ! , This means that even if the uclk is 1/2 the memory speed (thus enabling very high memory speeds), You would still get bandwidth increases, Unlike zen2/3. essentially Zen 4 may potentially scale with ddr5 memory all the way up to 12000mt/s with a minor latency penalty.(assuming the motherboard can even reach those speeds).

This is all conjecture of course until further details are revealed, but it certainly Looks promising.