DDR5 RAM Speeds and the future

[Panther_Seraphin]

DDR6 seems to be scaling in both directions. Higher speeds but also more memory channels per module. I suspect moving from one channel per module to two has lowered the need for pure bandwidth/latency in DDR5 and in theory going to four will reduce it further.

 

[Bill_Bright]

For 4 years now AMD has used cache to compensate for any potential memory-side weaknesses

It may totally be a case of semantics. If so, I apologize. But for me, I would not categorize the slower speeds on the memory side as "weaknesses".

If one device (processors) has an integrated feature considered a "strength", that does not suggest a totally different device (RAM modules) without that integrated feature has a "weakness". CPUs and RAM are apples and oranges, not two kinds of apples.

CPU speeds are, and always have been faster than RAM (and motherboard bus) speeds. That is not a "weakness" of the RAM. Those are just basic characteristics of the different technologies between a CPU die and RAM. For this reason, for years (much longer than 4 years), processors typically include multiple levels of on-die cache to temporarily stuff (cache) high-priority data for quick (super quick) access.

So to me, processor makers have used on-die cache to "take advantage" of the faster speeds of the on-die capabilities. This is not to "compensate" for the RAM, but to "take advantage" of that extra die space.

Again, hopefully not arguing semantics. Just trying to clarify the issue.

 

[Dr. Dro]

The sweet spot between cost and performance is a memory kit at ~DDR5-6000 to 6400 at a latency of 32 to 34, appropriate subtimings. 6400 C30 is already well in the range of "optimal" but may be a little more difficult to get or stabilize. The very best first-generation Hynix A-die kits (such as the one I have, see on specs) on Z690 will likely max out here for any realistic usage pattern, and will require 1.475 to 1.5V to do so. -6800 on a 4 dimmer Z690 I've only seen done on a ROG Maximus Z690 Extreme and the guy had plenty of luck I didn't share with my MSI MEG Z690 ACE.

Is a DDR5-8000 kit fast and awesome? Yeah
Do you need a DDR5-8000 kit to get the best out of your machine? Probably not
Can your platform take a kit that fast without manual configuration (well beyond just enable XMP and call it a day)? I'd say very bloody unlikely!

If you're on Ryzen, basically give up just now. Whether due to early AGESA bugs or the having to run quarter-rate FCLK above DDR5-6000, it's not worth it.
If you're on Intel Z690 (doesn't matter if it's Alder or Raptor), there's like two motherboards that are gonna do it, basically the EVGA Dark and the Aorus Tachyon
If you're on Intel Z790, you might, might not. It'll depend on whether your IMC plays ball and if your motherboard's traces are good. 1DPC and boards such as the ROG Apex are still heavily preferred

 

[jxdking]

The sweet spot between cost and performance is a memory kit at ~DDR5-6000 to 6400 at a latency of 32 to 34, appropriate subtimings. 6400 C30 is already well in the range of "optimal" but may be a little more difficult to get or stabilize. The very best first-generation Hynix A-die kits (such as the one I have, see on specs) on Z690 will likely max out here for any realistic usage pattern, and will require 1.475 to 1.5V to do so. -6800 on a 4 dimmer Z690 I've only seen done on a ROG Maximus Z690 Extreme and the guy had plenty of luck I didn't share with my MSI MEG Z690 ACE.

Is a DDR5-8000 kit fast and awesome? Yeah
Do you need a DDR5-8000 kit to get the best out of your machine? Probably not
Can your platform take a kit that fast without manual configuration (well beyond just enable XMP and call it a day)? I'd say very bloody unlikely!

If you're on Ryzen, basically give up just now. Whether due to early AGESA bugs or the having to run quarter-rate FCLK above DDR5-6000, it's not worth it.
If you're on Intel Z690 (doesn't matter if it's Alder or Raptor), there's like two motherboards that are gonna do it, basically the EVGA Dark and the Aorus Tachyon
If you're on Intel Z790, you might, might not. It'll depend on whether your IMC plays ball and if your motherboard's traces are good. 1DPC and boards such as the ROG Apex are still heavily preferred

It is amazing that 1st gen DDR5 AMD cpus can do 8000 with a proper 2-dimm board.

For the future, the sweet spot should be around 7200 and up. The major change comparing DDR4 is that DDR5 halves the length of Command and Address bus. I would expect it will run double the speed of DDR4.

 

[ir_cow]

I can hardly wait to show you guys the results I got. Doing the final checks right now. Lots of data to cover so its taking a while. Don't want to spoil things yet

 

[RootinTootinPootin]

Is a DDR5-8000 kit fast and awesome? Yeah

for those who can reach the speed, YES

Can your platform take a kit that fast without manual configuration

Technically some do.

-6800 on a 4 dimmer Z690 I've only seen done on a ROG Maximus Z690 Extreme

There's someone from my other Forum who did 8000mhz on the Z690 Extreme. Runs stable. Even my Z790 Maximus Extreme does 8000mhz (Brute forced, same goes with the Maximus Hero) most newer revision boards do 8000mhz easy peasy with a god bin IMC 13th gen and 14th gen., sometimes its a skill issue, sometimes its silicon lottery, most of the time both.

 

[Panther_Seraphin]

The major change comparing DDR4 is that DDR5 halves the length of Command and Address bus. I would expect it will run double the speed of DDR4.

True but now you have 2 buses per channel vs 1 big one.

 

[ir_cow]

There's someone from my other Forum who did 8000mhz on the Z690 Extreme. Runs stable. Even my Z790 Maximus Extreme does 8000mhz (Brute forced, same goes with the Maximus Hero) most newer revision boards do 8000mhz easy peasy with a god bin IMC 13th gen and 14th gen., sometimes its a skill issue, sometimes its silicon lottery, most of the time both.

More incline to believe that any 4-slot Intel MB running 8000+ either is applying a stupid about of auto voltage and or has a good binned CPU. Don't forget most people think booting into windows is enough to be considered stable...

I don't personally have any of the Z790 14th Gen refreshes, but a number of OC reviewers show it needs 1.5V on the VDD2 for 8000 to work. Kinda stupidly high. I only need 1.25V for the same thing on the Apex.

 

[jxdking]

I don't personally have any of the Z790 14th Gen refreshes, but a number of OC reviewers show it needs 1.5V on the VDD2 for 8000 to work. Kinda stupidly high. I only need 1.25V for the same thing on the Apex.

I don't have any LGA1700. In the past, I rely on ASUS color code to indicate if a voltage is safe.
White is for looser.
Yellow is too low.
Red is just right.
Purple is too high.

 

[Bill_Bright]

Don't forget most people think booting into windows is enough to be considered stable...

I've come across folks like that too and just don't understand it. I'm not talking about "normal" users who simply believe their computer is a tool or communications appliance in their home (and nothing wrong with that). I'm talking about folks who consider themselves enthusiasts.

There is the other extreme who are almost as bad and those are the folks who are only convinced their system is stable if they can run the most demanding stress tests at 3 times normal clocks and voltages for 10 hours straight without exceeding 30°C. Nevermind the fact they run the most demanding games all day long with record breaking FPSs. It stuttered with Prime95, there must be something wrong.

 

[ir_cow]

I don't have any LGA1700. In the past, I rely on ASUS color code to indicate if a voltage is safe.

Yes.... That is not the same for all vendors and it is sometimes scary what is considered "safe", with documented cases of that killing a CPU. Last time I trusted the color codes was x79 and it killed my CPU .

One example is you can set the LLC to max and a safe "value" can kill the CPU.

. Nevermind the fact they run the most demanding games all day long with record breaking FPSs. It stuttered with Prime95, there must be something wrong.

Prime95 without AVX offset is a real challenge. Also the fact that it runs entirely on the CPU cache now causes heat problems.

 

[Bill_Bright]

Prime95 without AVX offset is a real challenge. Also the fact that it runs entirely on the CPU cache now causes heat problems.

I understand that. And it, in part, illustrates my point. My point is, stress tests are designed to take the hardware to the breaking point. They do not simulate or represent real-world scenarios.

For example, how many of the latest games, CAD/CAE program, or other really demanding programs run entirely in the cache?

If the goal in building a machine is to earn top scores and bragging rights on these stress tests, that is perfectly fine! If that is the intended purpose for the build, I have no problem with that.

But if the goal in building a machine is to be entertained playing the latest games, designing a nuclear submarine, creating CGI content, then build the machine for those tasks and be happy.

You don't build a drag racer then take it out on formula 1 circuits expecting it to compete, then be unhappy it won't run at 200mph for 24 hours straight.

Now for sure, coding is way out of my wheelhouse. That's by choice and I am not pretending to be an expert at this area. Instead, I've been a certified electronics tech for 50 years and much prefer to have a meter probe in one hand and a soldering iron in another. I've had my own shop for many years. I am just saying I have just seen too many times where folks put too much emphasis on stress tests and benchmark scores - to the point where they have convinced themselves there is something wrong with their hardware when, in fact, there is nothing wrong. They want me to fix something that is not broken, or they feel they must spend $1000 on a new graphics card to play solitaire because their hardware did not pass some extreme stress test.

My neighbor's BMW will do 150MPH. My F150 won't. Does that mean my truck is flawed?
My truck will haul a 9,000 pound trailer at 70mph. His BMW won't. Does that mean his Beamer is flawed? It means my knuckles are white, my heart is pounding at 150bpm and I'm coming through - so get out of my way!

To me as a hardware tech, it make sense with a new build to set everything at the default settings, then run a benchmark program to establish a baseline. Then, in a year or two if you get a funny feeling something might be wrong, run the benchmark program again and compare it to the original baseline. Odds are the new benchmark will show all is good and it is the user that has changed expectations - likely from reading something on the internet and believing it therefore, must be true.

 

[ir_cow]

I understand that. And it, in part, illustrates my point. My point is, stress tests are designed to take the hardware to the breaking point. They do not simulate or represent real-world scenarios.

I understand this point of view, but when do we decide a program is not a good representation of a stressful environment?

Or maybe we should accept neither Intel or AMD top CPU SKU can turbo to their max clocks and "pass" prime95. Base clocks sure, but turbo? But are we now just accepting false advertising or is Prime95 not a good representation of a heavy stress workload?

 

[Bill_Bright]

I understand this point of view, but when do we decide a program is not a good representation of a stressful environment?

It is not about representing stressful environments. It is about testing environments representing "real-world" environments. They rarely do.

I am NOT saying stress testing is always bad or misleading. On the contrary, when comparing competing/comparable products in a laboratory environment, in order to make a purchasing decision, it is probably one of the best. It is just not ideal and for sure, I don't know what is.

The best comparison testing occurs with a large number of sample units conducting a large variety of real world tests over a long period of time. But who has the resources to do that? Consumer Reports, maybe, but not tech support sites like TPU.

The problem I see is users failing to do their homework, then failing to understand their own expectations and the limits of their own hardware. If the computer is providing great game play, no stuttering, freezes, sudden reboots, or crashes, its working fine. Yet some run these stress tests, do not get the same results as they read on line, then think something is wrong.

Or maybe we should accept neither Intel or AMD top CPU SKU can turbo to their max clocks and "pass" prime95.

No doubt some of this special tweaking happens. The company marketing weenies get their grubby fingers in there and force changes just to peak out with such and such test so it can be used for marketing fodder - even though those tweaks often compromise other just as, if not more important practical areas.

It is just like when Windows Defender first came out. Microsoft announced they would be coding Defender to thwart today's real-world threats - not score well on the "simulated real-world" tests done in those testing labs. Defender being free and already in there, Microsoft does not need the marketing fodder of high test scores. Do they decided not to waste resources defending against 10,000 bits of malware from 10-20 years ago that no longer exists out in the wild.

And what happened? All the 3rd party programs, as well as the anti-MS media slammed Defender for doing poorly on the simulated tests while the 3rd party programs boasted about their scores.

But what was (and still is) the "real" real-world impact? Users of Windows Defender were NOT getting infected at higher rates than any of the programs that scored tops in the "simulated" real-world tests. Surprise surprise.

Point being, does a higher score in Prime95 automatically indicate better game play in the latest Halo or whatever? Nope.

I think one of the bigger issues is with gamers who believe the PC industry centers around them. That is not true by any means. Gaming is certainly important and for sure is a driving force in PC technology advancement. But gaming by gaming PC enthusiasts is actually a small segment of the entire PC industry.

 

[Dr. Dro]

for those who can reach the speed, YES

Technically some do.

There's someone from my other Forum who did 8000mhz on the Z690 Extreme. Runs stable. Even my Z790 Maximus Extreme does 8000mhz (Brute forced, same goes with the Maximus Hero) most newer revision boards do 8000mhz easy peasy with a god bin IMC 13th gen and 14th gen., sometimes its a skill issue, sometimes its silicon lottery, most of the time both.

8000 on the Z690 Extreme's 100% done with exotic cooling or for quick benchmarking purpose, none of the 4-dimmers will do that stable for 24/7. I mean "technically" they actually added a validated DDR5-10000 mode on my Z690 Ace's BIOS, I just don't think it'll even boot, regardless of IMC quality or the memory kit used.

Z790 was a big improvement over Z690 for memory speeds, unless you have one of the specific XOC Z690's like those I mentioned, you aren't getting good clocks on most of them, even the high-end boards.

 

[ir_cow]

6:29 it starts

It's almost like all these QVL 8000+ validations was done in a lab

 

[RootinTootinPootin]

More incline to believe that any 4-slot Intel MB running 8000+ either is applying a stupid about of auto voltage and or has a good binned CPU. Don't forget most people think booting into windows is enough to be considered stable...

philipp park, the korean overclocker, does 8400 CL34 on his Gigabyte Aorus Pro X, insane peeps on OC.net as well too (Asrock Z790 Nova, Maximus Z690 Extreme). on Chiphell, you would find peeps there doing 8000mhz on their Maximus Z790 Hero. those peeps don't just boot into windows and call it a day, those junkies do Y-Cruncher VTT/VST runs for a whole night. call me crazy with them.

 

[ir_cow]

those peeps don't just boot into windows and call it a day, those junkies do Y-Cruncher VTT/VST runs for a whole night. call me crazy with them.

And what are the voltages used? I'm willing to bet it isn't a low value.

 

[Keullo-e]

Doesn't the highest JEDEC speeds almost always get ~50% more than the first speeds?

 

[RootinTootinPootin]

And what are the voltages used? I'm willing to bet it isn't a low value

I did say they're crazy people right??

 

[ir_cow]

Doesn't the highest JEDEC speeds almost always get ~50% more than the first speeds?

Like the first JEDEC of each generation, as in DDR4-1600 to the last DDR4-3200? DDR5-8800 JEDEC is on it's way. Micron did a press release about it not long ago.

 

[jxdking]

philipp park, the korean overclocker, does 8400 CL34 on his Gigabyte Aorus Pro X, insane peeps on OC.net as well too (Asrock Z790 Nova, Maximus Z690 Extreme). on Chiphell, you would find peeps there doing 8000mhz on their Maximus Z790 Hero. those peeps don't just boot into windows and call it a day, those junkies do Y-Cruncher VTT/VST runs for a whole night. call me crazy with them.

Intel chip scales with temperature. The lower the temperature, the higher memory clock that you can go. If you delid it and put a chiller on it. It can go 8400+ stable. Also silicone lottery is a real thing. It gives advantage to someone has relationships with retailers or manufacturers that can access chips.

 

[Keullo-e]

Like the first JEDEC of each generation, as in DDR4-1600 to the last DDR4-3200? DDR5-8800 JEDEC is on it's way. Micron did a press release about it not long ago.

Whoa, didn't even know about DDR4-1600, I thought that 2133 is the slowest.

 

[RootinTootinPootin]

Intel chip scales with temperature. The lower the temperature, the higher memory clock that you can go. If you delid it and put a chiller on it. It can go 8400+ stable. Also silicone lottery is a real thing. It gives advantage to someone has relationships with retailers or manufacturers that can access chips.

I'm running 8200 CL36, I have the TForce Xtreem's and the Patriot Viper's, these sticks are well padded/cooled enought, I also have the G.Skill 7600 CL36 kit that does 8400 as well but I haven't dissected it as well, I also 2 more 6400 A-Die kits that do astounding well at 8400mhz CL38 as well.

TLDR, I just top them off with a 120mm fan..modern solution for modern problems, right?

 

[Panther_Seraphin]

TLDR, I just top them off with a 120mm fan..modern solution for modern problems, right?

These coolers may start making a comeback