Zen 2 "3700x" and high speed low latency RAM

[Synedisis]

Alright boys I've been thinking about this all day and I just wanna do a quick hypothetical discussion to get some different ideas and opinions.

Money doesn't matter and because this is hypothetical there is no waiting to see what it actually is.

Let's say that we're running the new Zen 2 3700x that's coming out soon and we want to put 32gigs of 4000mhz CL17 RAM in it.

What are the minimum specs that the new processor would need to run that RAM stably?

What would be the maximum performance unstable?

Will the boards coming out be able to run that? (I have heard next to nothing about them other than wifi 6)

Is this realistically possible? (rumormill that Zen 2 will be capable of 5000mhz)

Feel free to reply however you want I personally don't mind a wall of text.

 

 

 

[GoldenLag]

5 minutes ago, Synedisis said:

What are the minimum specs that the new processor would need to run that RAM stably?

depends on if you win the silicon lottery with the memmory controller. 

6 minutes ago, Synedisis said:

Let's say that we're running the new Zen 2 3700x that's coming out soon and we want to put 32gigs of 4000mhz CL17 RAM in it.

why would you want that ram. its got worse timings. just get 3200mhz kit if you are feeling spicy or just a 3000mhz kit.  anything above 3600mhz is more or less a waste of money.

 

(based on what we know from testing done today, stuff might be different for Zen 2, tho i doubt it)

8 minutes ago, Synedisis said:

Is this realistically possible? (rumormill that Zen 2 will be capable of 5000mhz)

daydream at best. not happening. 

 

 

 

Zen 2 is rather close right now, its just better to wait and see

[MeatFeastMan]

Forget 5000mhz. Not happening. This biostar leak confirms that 4000mhz is realistic:

https://wccftech.com/biostar-x570-racing-gt8-motherboard-amd-ryzen-3000-cpu-leak/

 

But that doesn't mean 32gb 4000mhz. I've decided to choose a corsair b die kit running 2x16gb at 3466 cl16 for zen2. That's much more reasonable for the boards to handle on dual rank 32gb. You've got to remember that when it says 4000mhz, it's talking about the absolute maximum possible but mainly using 8gb and 16gb of ram, not 32.

 

If we go off the 4000mhz leak, then we can assume the following.. Boards from zen+ went up to 3600mhz ram if you knew what you were doing and you got lucky and had B Die. So Zen2 has gained 400mhz on the memory (Although we are comparing the biostar board to a board like the crosshair, so 400mhz may be on the low side). On dual rank 32gb kits, with B Die you can currently just about get away with 3200 cl14 stable. So for 32gb, you're looking at 3600mhz maximum on zen2. I would do what I've done and instead go for a 3466 cl16 kit. It's a safe bet for 2x16gb and is very very likely to work on x570. 4000mhz cl17 32gb ain't happening.

[ch3w2oy]

I agree, 5000 isn't happening. And just because a motherboard gives you the option to go as high as 4000mhz, didn't mean you will be able to..

 

I can set my CPU ratio to 52.00, doesn't mean it's going to run at 5.2ghz..

 

I got 3600 c16 RAM for my Zen 2 build just to be on the safe side..

[porina]

7 minutes ago, GoldenLag said:

why would you want that ram. its got worse timings. just get 3200mhz kit if you are feeling spicy or just a 3000mhz kit.  anything above 3600mhz is more or less a waste of money.

 

(based on what we know from testing done today, stuff might be different for Zen 2, tho i doubt it)

In latency, 4000C17 is actually slightly faster than 3200C14. Overclockers could get 3200 to C12 or maybe even lower, but I don't recall ever seeing that on retail kits.

 

The increased bandwidth is desperately needed, especially with AMD's "more cores" direction. In my use cases, dual channel can be inadequate for 4 Intel cores or 8 Zen cores due to the lower per-core FPU in Zen. But Zen 2 will catch up to consumer Intel levels and that's going to really hit the ram hard unless either ram gets much faster, more channels are added, or there is sufficient cache to mitigate it. I'm kinda expecting the 3rd will be the actual direction but will look forward to seeing actual details.

 

I'd further add that DDR4 isn't JEDEC specified beyond 3200, so I wouldn't expect AMD official support to go past that. Anything running above that is effectively a manufacturer overclock, like it is today.

[BigRom]

24 minutes ago, ch3w2oy said:

I agree, 5000 isn't happening. And just because a motherboard gives you the option to go as high as 4000mhz, didn't mean you will be able to..

 

I can set my CPU ratio to 52.00, doesn't mean it's going to run at 5.2ghz..

 

I got 3600 c16 RAM for my Zen 2 build just to be on the safe side..

Its a classic case of "just because you can doesn't mean you should"

[GoldenLag]

1 hour ago, porina said:

The increased bandwidth is desperately needed, especially with AMD's "more cores" direction.

With the exception of certain tasks. There hasnt been a lack of memmory bandwidth.

 

Case in point, 2990wx which only really suffered due to w10 sqedualler.

 

2990wx managed fine with the bandwidth provided even through leach dies. The reason why higher speed memmory was creat was due to the locked clockspeed with IF. (Edit: which was latency related)

[porina]

22 minutes ago, GoldenLag said:

With the exception of certain tasks. There hasnt been a lack of memmory bandwidth.

It isn't all or nothing. As the total CPU potential increases, then without a corresponding increase in ram bandwidth there will be ever more scenarios where it will impact performance. Right now it might only be a little bit here and there, but the gap can only grow if cores continue to accumulate faster than the rest of the system. Cache can mitigate it to a degree but in itself isn't a long term solution.

 

IF speed being tied to ram is just an additional complication on AMD side. I've yet to see some good data separating IF speed and ram bandwidth e.g. by running single channel high speed ram vs dual channel lower speed ram. That should give clues if IF or ram bandwidth is dominant in a given scenario. I'd do it myself if I ever have time, but I need to prioritise and complete IPC and power efficiency testing before Zen 2 arrives so I can quickly compare it to current solutions.

 

The Windows thing with 2990WX is not really relevant when considering ram speed. I'm a believer in getting the right tool for the job. It might scale well for the low intensity tasks but I wouldn't touch it for my uses. 

[GoldenLag]

2 minutes ago, porina said:

It isn't all or nothing. As the total CPU potential increases, then without a corresponding increase in ram bandwidth there will be ever more scenarios where it will impact performance. Right now it might only be a little bit here and there, but the gap can only grow if cores continue to accumulate faster than the rest of the system. Cache can mitigate it to a degree but in itself isn't a long term solution.

it was almost only latency related and not bandwhidt related with the exception of very memmory intensive tasks. 

 

there is enough Ram-bandwidth to feed 32 cores in 4 channel with 16 of the cores being on leach-dies. 

4 minutes ago, porina said:

IF speed being tied to ram is just an additional complication on AMD side. I've yet to see some good data separating IF speed and ram bandwidth e.g. by running single channel high speed ram vs dual channel lower speed ram

singlechannel Ram has shown to be detrimental to performance in certain tasks. but the IF being locked to Ramspeed is more in relation with latency than actual bandwhidth. the IF has enough Bandwhidt to feed the cores perfectly fine. 

5 minutes ago, porina said:

The Windows thing with 2990WX is not really relevant when considering ram speed.

it isnt, but Ram and leach dies were blamed for a lot of the issues that originated in Windows not knowing how to handle the 2990wx. and in linux environment it was not bandwhidt that was the issue, it was latency due to leach dies. 

 

overclocking Ram in Ryzen brings improvement due to latency, not so much in terms of bandwhidt. 

7 minutes ago, porina said:

It might scale well for the low intensity tasks but I wouldn't touch it for my uses. 

2990wx was kinda a flop due to it being lackluster in windows due to the sqedualler. 

 

8 minutes ago, porina said:

That should give clues if IF or ram bandwidth is dominant in a given scenario.

we know singlechannel isnt prefferable. though IF brings what is needed, the issue is latency.

 

with the 2990wx running just fine with 4 channels of memmory, it should not be an issue runnin 2 channels 16 cores with no leach dies. 

[porina]

18 minutes ago, GoldenLag said:

it was almost only latency related and not bandwhidt related with the exception of very memmory intensive tasks. 

My position is the same, I haven't seen proof of that. If not bandwidth related, would tighter timings show much of an improvement at otherwise same configuration?

 

18 minutes ago, GoldenLag said:

there is enough Ram-bandwidth to feed 32 cores in 4 channel with 16 of the cores being on leach-dies. 

There is already not enough bandwidth from dual channel to feed 8 core consumer Ryzens. It wont apply in every scenario, but it is a problem that is getting worse not better.

[GoldenLag]

2 minutes ago, porina said:

My position is the same, I haven't seen proof of that. If not bandwidth related, would tighter timings show much of an improvement at otherwise same configuration?

tighter timings have been prefferable to higher frequency to some degree. though Ramspeed is key to get IF latency down. 

3 minutes ago, porina said:

There is already not enough bandwidth from dual channel to feed 8 core consumer Ryzens. It wont apply in every scenario, but it is a problem that is getting worse not better.

what usecases? because afaik that is IF latency related and not bandwhidt related. if it was the case then 8 core threadripper and 4 core Ryzen show improvement, though i havent seen much of that. 

 

there are offcourse workloads where there isnt enough bandwhidt, which is the case with most CPUs. most in regards to AVX

[porina]

37 minutes ago, GoldenLag said:

what usecases? because afaik that is IF latency related and not bandwhidt related. if it was the case then 8 core threadripper and 4 core Ryzen show improvement, though i havent seen much of that. 

 

there are offcourse workloads where there isnt enough bandwhidt, which is the case with most CPUs. most in regards to AVX

In my use cases, they are heavy AVX usage. That's why Zen 2 both excites and scares me at the same time. More cores. More AVX. No more ram bandwidth... I hope they pile on the cache to offset it. On that note, do we know the L3 cache per core of Rome? I don't recall seeing that.

[GoldenLag]

45 minutes ago, porina said:

In my use cases, they are heavy AVX usage. That's why Zen 2 both excites and scares me at the same time. More cores. More AVX. No more ram bandwidth... I hope they pile on the cache to offset it. On that note, do we know the L3 cache per core of Rome? I don't recall seeing that.

Wasnt there a doubling of L3?

 

Unshure. I would have to doublecheck that one.

 

If you want to look yourself. Check Tum Apisak on twitter and look at the ES

[porina]

1 hour ago, GoldenLag said:

Wasnt there a doubling of L3?

 

Unshure. I would have to doublecheck that one.

 

If you want to look yourself. Check Tum Apisak on twitter and look at the ES

I thought I heard noises suggesting an increase but nothing very solid. Couldn't find anything on that twitter that was of use. With the hopeful reveal just over a week away it wont be long before we hopefully get some meat on the details.

[GoldenLag]

10 minutes ago, porina said:

I thought I heard noises suggesting an increase but nothing very solid. Couldn't find anything on that twitter that was of use. With the hopeful reveal just over a week away it wont be long before we hopefully get some meat on the details.

Think there was some early talk through codename decoding. 

 

Which i think where more cache was suggested. 

 

Those codenames probably came from Apisak

[Jurrunio]

6 hours ago, Synedisis said:

What are the minimum specs that the new processor would need to run that RAM stably?

I dont totally understand. If you're talking about IMC quality, it would need to be a lot better than 2nd gen Ryzen. Right between Intel 14nm consumer platform and 2nd gen Ryzen to be exact.

 

6 hours ago, Synedisis said:

What would be the maximum performance unstable?

maximum unstable performace? That you will have to try for yourself. Too many variables in the equation

 

6 hours ago, Synedisis said:

Will the boards coming out be able to run that? (I have heard next to nothing about them other than wifi 6)

If the memory controller of the CPUs are any good, yes

 

6 hours ago, Synedisis said:

Is this realistically possible? (rumormill that Zen 2 will be capable of 5000mhz)

I give that a miniscule chance with stupid loose timings and hence terrible performance other than high frequency. Keep in mind that even an Intel 14nm CPU (8700k in this case) barely did that (5066MHz) on an mITX board. Talking about binning CPUs for their memory controllers rather than their cores...

 

@porina @GoldenLag what makes speculation harder is the IF = MEMCLK/2 option. Just no idea what will happen if memory speeds stay the same while IF frequency cuts in half. It's possible that we're still stuck with 3600MHz at the end of the day just to not drop IF frequency for memory speed.

[porina]

7 minutes ago, Jurrunio said:

what makes speculation harder is the IF = MEMCLK/2 option. Just no idea what will happen if memory speeds stay the same while IF frequency cuts in half. It's possible that we're still stuck with 3600MHz at the end of the day just to not drop IF frequency for memory speed.

Another thing we don't know yet is how they configured IF in terms of width, not just clock. If they go wider (than previously) then as high a clock isn't so important.