Wow! Is ddr5 very overclockable or did I just get lucky?

[Wirko]

ECC Regisited server memory has a parody bit

No matter what it's called, let's hope it works...........

 

[tomcug]

Is this normal? Is it safe? Should I try pushing it even more?

If you consider yourself lucky, how would you call the owner of the FURY Beast 5600 MHz kit that can push 8000+ MHz due to A-die chips used? I happen to have such a kit

As for your case, I'd focus on optimising primary timings first and secondary/tertiary later. This is where you can gain performance with DDR5, not with clock that doesn't help too much on this platform.

 

[bug]

Anyway I've just never had ram that could so easily run above it advertised clock speed. Thats awesome.

Far from awesome, this is actually expected behavior. You pushed the kit 400MHz which may sound like a lot, but it's actually 7% above base speed. I'm amazed you can feel a difference. But hey, faster is faster, so congrats.

 

[maxfly]

Memory... the final frontier.

Once manufacturers figure out a way to squash memory ocing. We'll be doomed to a slow agonizing death by ocing boredom.

Sounds like those fury beast 5600s are the hidden gems everyone wants to stumble across! Better grab a set before word gets out that their A-die and clock like...beasts hehe.

 

[Wirko]

Is DDR5 so overclockable because it is madly error correcting? if so I would prefer to run at a speed that has few, if any, errors.

That can't be. With memory overclocking, you risk errors during transfer, and those are not detected by the on-die ECC.

So you need some flavour of MemTest to determine the maximum speed. "Few" errors are too many, actually a single error a month is a lot. Also, the probability of an error may be temperature-dependent, so you'd ideally test both at low and elevated temperatures.

The errors will still show up snd I would think things would be slower due to having the data discarded and resent.

Nothing is ever re-sent, even with server memory. I looked at some documents that discuss the details, such as this one by Xilinx/AMD, but repeated sending just isn't mentioned anywhere. Radiation-hardened systems probably do things like that but those are on another level, they can also repeat computing operations and so on.

 

[bug]

Nothing is ever re-sent, even with server memory. I looked at some documents that discuss the details, such as this one by Xilinx/AMD, but repeated sending just isn't mentioned anywhere. Radiation-hardened systems probably do things like that but those are on another level, they can also repeat computing operations and so on.

I believe the point is to properly signal to apps they got invalid data, instead of just presenting them with a bit-flipped byte.

 

[Speedyblupi]

ECC Regisited server memory has a parody bit while sending data I believe.

Parity.

 

[iameatingjam]

Far from awesome, this is actually expected behavior. You pushed the kit 400MHz which may sound like a lot, but it's actually 7% above base speed. I'm amazed you can feel a difference. But hey, faster is faster, so congrats.

Alright one guy says its fishy that my ram runs like this, another says its normal behavior. idk. All I know is with my ddr4 kits, an overclock like this, while also reducing timings, was like a 100% chance of either a crash or at the very least, some weird behavior.

No I don't notice the difference in everyday computing, not yet anyway. But in benchmarks the difference was crazy stark. I wish I kept the original score but I do remember it was in the 88th percentile. Went from that to 98th. Also brought my cpu percentile up from 94 I think it was to 96. And just reducing the timings brought the score up over a 1000 points ( on passmark) To me thats pretty sweet. Since I have a 4090 I was kind of wondering if I made the wrong decision getting a 13600k so I'm happy to get any extra cpu performance I can. But I mean if thats just so unremarkable to you, okay then. I'm just new to this platform and was pleasantly surprised to find out how flexible ddr5 was.

If you consider yourself lucky, how would you call the owner of the FURY Beast 5600 MHz kit that can push 8000+ MHz due to A-die chips used? I happen to have such a kit

As for your case, I'd focus on optimising primary timings first and secondary/tertiary later. This is where you can gain performance with DDR5, not with clock that doesn't help too much on this platform.

Way ahead of you, I went to cl34 which worked but didn't seem to increase performance so I went to cl32 and copied the other timings from some other kingstom kit and that did gain me a lot of performance. I probably could have pushed it further but I'm pretty happy with where its at now. Perfoming singificantly better than when I got it, performing better than my old ddr4 3200 cl16 on passmark ( whereas before it didn't) and its singificantly increased my cpu score. Now with my 13600k I'm getting over double the score I got with my 11600k (~19500 vs 41000 and change). I know benchmarks aren't everything but still, to me, thats awesome and more than I was expecting. Hoping that will help my cpu keep up with the 4090 when I test out some demanding games. I've seen online that 400mhz difference in ddr5 ram speed actually does make a significant fps difference in games when you're on a 4090.

Anyway, sorry to everybody that I have bored with my excitment over this minor incident. Just because of my experience with ddr4, I didn't know whether it was out of the ordinary, or not.

 

[Wirko]

I believe the point is to properly signal to apps they got invalid data, instead of just presenting them with a bit-flipped byte.

Yeah, and it's called kill signal.
Actually, the entire system must lock up if an unrecoverable error occurs. That's by design. I don't know if there exist any variations to that, for example, killing only the process whose memory space is affected.
Correctable errors are corrected on the fly, using the parity bits.

 

[maxfly]

Alright one guy says its fishy that my ram runs like this, another says its normal behavior. idk. All I know is with my ddr4 kits, an overclock like this, while also reducing timings, was like a 100% chance of either a crash or at the very least, some weird behavior.

No I don't notice the difference in everyday computing, not yet anyway. But in benchmarks the difference was crazy stark. I wish I kept the original score but I do remember it was in the 88th percentile. Went from that to 98th. Also brought my cpu percentile up from 94 I think it was to 96. And just reducing the timings brought the score up over a 1000 points ( on passmark) To me thats pretty sweet. Since I have a 4090 I was kind of wondering if I made the wrong decision getting a 13600k so I'm happy to get any extra cpu performance I can. But I mean if thats just so unremarkable to you, okay then. I'm just new to this platform and was pleasantly surprised to find out how flexible ddr5 was.



Way ahead of you, I went to cl34 which worked but didn't seem to increase performance so I went to cl32 and copied the other timings from some other kingstom kit and that did gain me a lot of performance. I probably could have pushed it further but I'm pretty happy with where its at now. Perfoming singificantly better than when I got it, performing better than my old ddr4 3200 cl16 on passmark ( whereas before it didn't) and its singificantly increased my cpu score. Now with my 13600k I'm getting over double the score I got with my 11600k (~19500 vs 41000 and change). I know benchmarks aren't everything but still, to me, thats awesome and more than I was expecting. Hoping that will help my cpu keep up with the 4090 when I test out some demanding games. I've seen online that 400mhz difference in ddr5 ram speed actually does make a significant fps difference in games when you're on a 4090.

Anyway, sorry to everybody that I have bored with my excitment over this minor incident. Just because of my experience with ddr4, I didn't know whether it was out of the ordinary, or not.

Rock on man, don't let jealous George and Georgina shit on your accomplishment. We need all the mem tweakers we can get. If those are in fact A-die ics, you've got plenty o headroom left to play around with.
You paid for ddr5 5600s and your running 6000...free performance. That's what ocing is all about, 400mhz or 4gigs. Doesn't matter. Its free and you did it yourself

 

[bug]

Alright one guy says its fishy that my ram runs like this, another says its normal behavior. idk. All I know is with my ddr4 kits, an overclock like this, while also reducing timings, was like a 100% chance of either a crash or at the very least, some weird behavior.

If your DDR4 kit was already overclocked (i.e. ran outside JEDEC specs), then yes, you wouldn't have had much headroom left. But 5600 for DDR5 is on the tame side, I would expect it to go faster than that.

But like I already said, faster is faster, I'm glad you could squeeze more juice out of it. gg

 

[AnotherReader]

DDR5 features on-die error correction, so it's more resilient to crashes. You need to keep in mind.

This is a misconception. With DDR5, capacity has been pushed high enough that errors have become more common without countermeasures. The countermeasure of choice is on-die ECC. Synopys explains:

With the higher capacity and speed coupled with the smaller process technology, the likelihood of single-bit errors increases on the DRAM memory arrays

 

[Dr. Dro]

This is a misconception. With DDR5, capacity has been pushed high enough that errors have become more common without countermeasures. The countermeasure of choice is on-die ECC. Synopys explains:

Cow already went into detail, it only applies to resting bits, but even then: having on-die ECC errors is bad, countermeasures or not

 

[AnotherReader]

Cow already went into detail, it only applies to resting bits, but even then: having on-die ECC errors is bad, countermeasures or not

Yes, it doesn't apply to link or transmission errors, but you have to look at the evolution of the technology to understand that they were forced to do this. Commodity DRAM is averse to increasing die size and they have always chosen lower die size over improvements in technology that could have yielded slightly higher die size, but more bandwidth or lower latency. They had to choose on-die ECC, because Rowhammer demonstrated even in the DDR3 days that they had pushed the density of DRAM high enough that reliability was suffering.

 

[JustBenching]

Alright one guy says its fishy that my ram runs like this, another says its normal behavior. idk. All I know is with my ddr4 kits, an overclock like this, while also reducing timings, was like a 100% chance of either a crash or at the very least, some weird behavior.

No I don't notice the difference in everyday computing, not yet anyway. But in benchmarks the difference was crazy stark. I wish I kept the original score but I do remember it was in the 88th percentile. Went from that to 98th. Also brought my cpu percentile up from 94 I think it was to 96. And just reducing the timings brought the score up over a 1000 points ( on passmark) To me thats pretty sweet. Since I have a 4090 I was kind of wondering if I made the wrong decision getting a 13600k so I'm happy to get any extra cpu performance I can. But I mean if thats just so unremarkable to you, okay then. I'm just new to this platform and was pleasantly surprised to find out how flexible ddr5 was.



Way ahead of you, I went to cl34 which worked but didn't seem to increase performance so I went to cl32 and copied the other timings from some other kingstom kit and that did gain me a lot of performance. I probably could have pushed it further but I'm pretty happy with where its at now. Perfoming singificantly better than when I got it, performing better than my old ddr4 3200 cl16 on passmark ( whereas before it didn't) and its singificantly increased my cpu score. Now with my 13600k I'm getting over double the score I got with my 11600k (~19500 vs 41000 and change). I know benchmarks aren't everything but still, to me, thats awesome and more than I was expecting. Hoping that will help my cpu keep up with the 4090 when I test out some demanding games. I've seen online that 400mhz difference in ddr5 ram speed actually does make a significant fps difference in games when you're on a 4090.

Anyway, sorry to everybody that I have bored with my excitment over this minor incident. Just because of my experience with ddr4, I didn't know whether it was out of the ordinary, or not.

The behavior is normal both for ddr4 ans ddr5. Just because you got a bad luck of the draw with your ddr4 doesn't mean much in its own.

The best way to go about overclocking your memory is to leave timings alone and just focus on frequency. Set the max voltage you are comfortable with ruining 24/7 and try to find the highest frequency you are stable with. After that focus on tightening down the latencies.

Also, set TREFI to 65535 and forget about it. Huge latency reduction

 

[ir_cow]

Memory... the final frontier.

Once manufacturers figure out a way to squash memory ocing. We'll be doomed to a slow agonizing death by ocing boredom.

That's easy. Lock the PMIC voltage. They can do it right now to prevent going above 1.43V. the next step would be to lock it to the XMP profile. I don't doubt that is coming in the coming years.

Nothing is ever re-sent, even with server memory. I looked at some documents that discuss the details, such as this one by Xilinx/AMD, but repeated sending just isn't mentioned anywhere. Radiation-hardened systems probably do things like that but those are on another level, they can also repeat computing operations and so on.

Hmmm. I guess I'm confusing it with GDDR or RAMBUS.

This is a misconception. With DDR5, capacity has been pushed high enough that errors have become more common without countermeasures. The countermeasure of choice is on-die ECC. Synopys explains:

This is much easier to read compared to the JEDEC white papers

 

[iameatingjam]

I'd focus on optimising primary timings first and secondary/tertiary later. This is where you can gain performance with DDR5, not with clock that doesn't help too much on this platform.

The best way to go about overclocking your memory is to leave timings alone and just focus on frequency.

The one thing I've learned from these forums is... there is no consensus on anything. Lol.

 

[AlwaysHope]

I used memtest64. Do you think thats adequate or should I run something else?

Like has been mentioned before in this thread, that app is too old & not routinely updated by the author.
Unless one likes correcting windows boot errors &/or registry corruption, it is highly advisable to test RAM outside of the OS environment as a first step before commencing booting into the OS.
I've always run memtest86, if all clear then move on... next step then run Windows memory diagnostic tool as an added secondary step but you have to boot into Windows first to tell it that (afaik there might be a way to get that to run before windows loads but I don't know) & as a bonus the tool is free!
If all is clear after the default run of that app, then throw other memory testing apps at it. RAM OC & testing is a long & tedious process, don't go there unless you have the patience.

 

[iameatingjam]

Like has been mentioned before in this thread, that app is too old & not routinely updated by the author.
Unless one likes correcting windows boot errors &/or registry corruption, it is highly advisable to test RAM outside of the OS environment as a first step before commencing booting into the OS.
I've always run memtest86, if all clear then move on... next step then run Windows memory diagnostic tool as an added secondary step but you have to boot into Windows first to tell it that (afaik there might be a way to get that to run before windows loads but I don't know) & as a bonus the tool is free!
If all is clear after the default run of that app, then throw other memory testing apps at it. RAM OC & testing is a long & tedious process, don't go there unless you have the patience.

I see. Thanks for the insight. I did run testmem5 extreme and used the windows diagonistic tool. Though I never booted into memtest86, I'll make sure to do that next time. Though I don't intend to push this memory any further.

 

[iameatingjam]

Though I don't intend to push this memory any further.

Okay I lied. Turns out that 6400mhz cl38 is more stable than the 6000mhz cl32 even though latency in nanoseconds is the same ( according to my benchmark anyway). The former can handle the anta777 extreme test and the latter cannot. Which is weird because I swore I ran it before. But its possible I accidentally picked the wrong profile.

Or its possible I modified some secondary timings. But even after making sure everything was back at defaults ( except the cl32 and 6000, and even after boosting voltage from 1.25 to 1.3, I still had errors).

Anyway to get to 6400 I increased cl to 40 ( and later decreased to 38, but can't go any lower than that) and increased all my voltages to 1.4 which makes me a little nervous with all the melting cpus lately - but that only happened on AMD, right? Should I try reducing the cpu side voltages? I just don't want my pc to not boot. I know I can fix it by popping out the battery but always freaks me out when it doesn't boot.

Or is 1.4 safe?

Edit: to be more specific, this is what I changed:

 

[ir_cow]

Anyway to get to 6400 I increased cl to 40 ( and later decreased to 38, but can't go any lower than that) and increased all my voltages to 1.4 which makes me a little nervous with all the melting cpus lately - but that only happened on AMD, right? Should I try reducing the cpu side voltages? I just don't want my pc to not boot. I know I can fix it by popping out the battery but always freaks me out when it doesn't boot.

Or is 1.4 safe?

Firs off, you are on a Intel platform, so the SoC CPU voltages meltdowns will not affect you. Second, 1.4 for VDD2 and VDDQ_TX is a bit high for such a lower DRAM frequency I'm doing those numbers for DDR5-8200 lol. You may want to look into the DRAM first and getting that up.

 

[iameatingjam]

Firs off, you are on a Intel platform, so the SoC CPU voltages meltdowns will not affect you. Second, 1.4 for VDD2 and VDDQ_TX is a bit high for such a lower DRAM frequency I'm doing those numbers for DDR5-8200 lol. You may want to look into the DRAM first and getting that up.

You mean the... dram voltages? Yeah they are at 1.4 too. When I first tried 6400 my pc didn't boot at all so kind of panicky I just put everything to 1.4 and loosened the timings and one of those things worked, just not sure which lol.

Tried lowing the cpu-dram side of things to 1.3. Seems to boot fine. And no immediate errors with memtest5 like I was getting with the other configuration. I'll test more thoroughly tonight, but thanks!

 

[ir_cow]

For DDR5-6000 you probably are fine with SA 1.2 / VDD2 1.25 / VDDQ_TX 1.25. You can always start higher and work your way down.

Did you figure out what type of ram it is? Hynix A / B , Samsung or Micron? Well can't be Micron.. probably not going to get 6000 out of that 16GB DIMMs at least.

Edit: Assuming the DRAM isn't the actual issue here, take notes of the voltages you are using. Make sure it passes Y-Cruncher first and run Memtest5 after. Can't pass yc, not point in running the memory test. I'm currently working on DDR5-8400. Managed to make some adjustments so I can run lower voltages vs before. If you don't keep notes, it gets confusing to remember what was actually stable to begin with. Here I already know for a fact the ram isn't the problem at 8200 and below, so it makes dialing in the voltages much easier.

Cold CPUs and Ram can trick you into thinking its stable. Once it warms up, it will fail the same tests. I would first start with DDR5-5600 stock. Figure out the lowest CPU voltages you can use as a reference / base. Instead of blindly applying values, how do you know what the floor is?

 

[iameatingjam]

For DDR5-6000 you probably are fine with SA 1.2 / VDD2 1.25 / VDDQ_TX 1.25. You can always start higher and work your way down.

Did you figure out what type of ram it is? Hynix A / B , Samsung or Micron? Well can't be Micron.. probably not going to get 6000 out of that 16GB DIMMs at least.

All I know is its hynix. Do you know how I can find out the exact die?

And I'm just learning now, it doesn't mind being clocked up. It just doesn't like the timings being brought down. So I brought the timings way up CL 42 and now I'm seeing how high it will go. I'm at 6666mhz now.

 

[ir_cow]

All I know is its hynix. Do you know how I can find out the exact die?

Well since your not going to the moon. Either M or A can do 6400 CL34-38-38-84 with 1.5V with brute force overclocking and probably down to CL30-36-36-50. A-Die can't go as low in timings, but it can go much higher.

Another trick is to type in all loose values with high voltage, overclock to the desired freq and work your way down. All Hynix can do 40-40-40-77-117 + tRFC 840 / tRFC2 640 / tWR 96 @ 6400 1.45V. Leave the rest of the values on auto.

This is why I always suggest buying the bin you want. overclocking is very tedious because of the many unknowns. You are shooting in the dark and hoping something sticks. You don't know the maximum freq of the MB or DRAM and also need to figure out the correct CPU voltages. It's a tall order without prior experience.