Higher mem clock speeds vs lower latency?

[KapiteinKoek007]

hello,

I have been wondering, which is better, higher clock speeds with loose timings or lower latency with tight timings. What do you think ?
I did some testing and benching on my own system results are in attachement, (excuse the format in which it is in)

 

[dgianstefani]

For productivity apps, bandwidth is better, for gaming latency is. But actual latency changes, so 4000/14 is lower latency than 3200/14. 3600/16 is lower latency than 2666/14 despite 2666/14 having tighter timings.

 

[KapiteinKoek007]

For productivity apps, bandwidth is better, for gaming latency is. But actual latency changes, so 4000/14 is lower latency than 3200/14. 3600/16 is lower latency than 2666/14 despite 2666/14 having tighter timings.

i think thats what im seeing in my testing, at 3200mhz/16 i average 60 ns on aida bench, you can clearly see the latency starts to drop when i increase clockspeeds and loosen timings, the latency is at 60ns again when i have 4600mhz/23

 

[ir_cow]

It generally goes in this order. Highest frequency you can go and while keeping 1:1 IMC ratio. Second lowest timings. Like everything, this is just a general rule of thumb. Some apps do like bandwidth over tighter timings. If you are GPU bound, memory isn't going to matter much either.

AIDA64 doesn't represent real world use. Just the latency and total bandwidth.

 

[eidairaman1]

Combination of both is best

 

[KapiteinKoek007]

It generally goes in this order. Highest frequency you can go and while keeping 1:1 IMC ratio. Second lowest timings. Like everything, this is just a general rule of thumb. Some apps do like bandwidth over tighter timings. If you are GPU bound, memory isn't going to matter much either.

AIDA64 doesn't represent real world use. Just the latency and total bandwidth.

Yeah ive tested the difference with gaming between 3200mhz default timings and 4600mhz with extremely loose timings and I saw no difference in fps. It did make a difference with a feature on my ssd, called momentum cache, when I use default xmp memory profile, my ssd read and write speeds are 6000/4000MB/s, and at 3600mhz xmp profile, I get over 10.000MB/s for both read and write speeds. So thats the only boost in performance im noticing.

Combination of both is best

Yeah, as you can see in the chart I made, when it reaches 4600mhz with extremely loose timings it had the same latency as with 3200mhz xmp profile, shame I cant seem to get it fully stable at 4600mhz, well not without increasing system agent voltage to over 1.35v which generates abit too much heat for my taste!

 

[eidairaman1]

Yeah ive tested the difference with gaming between 3200mhz default timings and 4600mhz with extremely loose timings and I saw no difference in fps. It did make a difference with a feature on my ssd, called momentum cache, when I use default xmp memory profile, my ssd read and write speeds are 6000/4000MB/s, and at 3600mhz xmp profile, I get over 10.000MB/s for both read and write speeds. So thats the only boost in performance im noticing.


Yeah, as you can see in the chart I made, when it reaches 4600mhz with extremely loose timings it had the same latency as with 3200mhz xmp profile, shame I cant seem to get it fully stable at 4600mhz, well not without increasing system agent voltage to over 1.35v which generates abit too much heat for my taste!

Well you've reached the max limit without it going Chernobyl or Fukushima on you.

 

[KapiteinKoek007]

Well you've reached the max limit without it going Chernobyl or Fukushima on you.

Hahaha if im really honest,im just suprised these mem sticks can even go that far with only basic adjustments of timings and voltages. Default is 2667mhz, 4500mhz is the highest speed I can get it stable, so thats still a massive overclock. Not sure if thats a ddr4 thing. My old ddr3 1866 sticks cant boot past 2133mhz, no matter what setting I adjust. so is overclocking improved alot with ddr4?

 

[ir_cow]

I think the easiest way to illustrate memory would be with lower graphical settings. The reason you are probably not seeing a change is because you are 100% GPU bound is my guess.

Spoiler: Benchmark Charts

 

[KapiteinKoek007]

I think the easiest way to illustrate memory would be with lower graphical settings. The reason you are probably not seeing a change is because you are 100% GPU bound is my guess.

Spoiler: Benchmark Charts

View attachment 256244View attachment 256245

Lowering graphical settings is ofcourse a solid way to make sure im not gpu bound, but in the end that would be pointless because I like my graphical setting maxed out in games. In some games my gpu usage is 100% mainly in First Person Shooters where I could settle for 60 fps because im still at 1080p and that would dramaticly lower my gpu usage but in battlefield 4 for example, 60 fps feels really laggy and slow, even with an average of 5ms input delay because im used to 150-200 fps. For all the other turn based and or rpg kind of games I settle for 60 fps, because i usually dont need these extra fps to give me a slight edge over my target when im playing Rome total war for example My cpu usage rarely goes above 30% when gaming, only in BF2042 the cpu usage peaks around 50%

Edit: In the end I think for my setups its pointless to overclock memory because everything above 3200mhz will cause the motherboard to change from gear 1 to gear 2, meaning the mem controller and memory ratio change from 1:1 to 1:2, which isn't helping in terms of raw performance.

 

[eidairaman1]

Hahaha if im really honest,im just suprised these mem sticks can even go that far with only basic adjustments of timings and voltages. Default is 2667mhz, 4500mhz is the highest speed I can get it stable, so thats still a massive overclock. Not sure if thats a ddr4 thing. My old ddr3 1866 sticks cant boot past 2133mhz, no matter what setting I adjust. so is overclocking improved alot with ddr4?

Nah, so many variables, My Sig Rig Kit is 2133, I have them with XMP 2400 but I tightened the primary timings as far as they can go (Cas10 is fastest for this 16GB Kit) . I havent messed with secondaries.

 

[KapiteinKoek007]

Lowering graphical settings is ofcourse a solid way to make sure im not gpu bound, but in the end that would be pointless because I like my graphical setting maxed out in games. In some games my gpu usage is 100% mainly in First Person Shooters where I could settle for 60 fps because im still at 1080p and that would dramaticly lower my gpu usage but in battlefield 4 for example, 60 fps feels really laggy and slow, even with an average of 5ms input delay because im used to 150-200 fps. For all the other turn based and or rpg kind of games I settle for 60 fps. My cpu usage rarely goes above 30% when gaming, only in BF2042 the cpu usage peaks around 50%

Nah, so many variables, My Sig Rig Kit is 2133, I have them with XMP 2400 but I tightened the primary timings as far as they can go (Cas10 is fastest for this 16GB Kit) . I havent messed with secondaries.

Think my old 1866 were maxed out on cas9 not sure tho. I dont understand 90% of the secondairy timings theres soo many as well, not used to soo many timings. My old z77 motherboard with these 1866 sticks had some secondairy options but this z590 board allows for MUCH more tweaking.

 

[JustBenching]

hello,

I have been wondering, which is better, higher clock speeds with loose timings or lower latency with tight timings. What do you think ?
I did some testing and benching on my own system results are in attachement, (excuse the format in which it is in)

Im assuming you are running gear 2 at 3600 and upwards, thats why there is a latency increase.

 

[KapiteinKoek007]

Im assuming you are running gear 2 at 3600 and upwards, thats why there is a latency increase.

I actually just found out that increasing system agent voltage to 1.41v and dram to 1.48 lets me get the memory to 4000mhz without going to gear 2, anything above 4000mhz is highly unstable in gear 1 it seems, can boot, but always crashes when loading windows. The PCH and VR are getting abit on the Hot side, 45 degrees celcius idle and 55-60 with heavy workloads. also dont like the voltage on the ram and system agent for 24/7 use. The extra heat and added wear on the motherboard and dram of running high voltage 24/7 is just not worth it.

 

[JustBenching]

I actually just found out that increasing system agent voltage to 1.41v and dram to 1.48 lets me get the memory to 4000mhz without going to gear 2, anything above 4000mhz is highly unstable in gear 1 it seems, can boot, but always crashes when loading windows. The PCH and VR are getting abit on the Hot side, 45 degrees celcius idle and 55-60 with heavy workloads. also dont like the voltage on the ram and system agent for 24/7 use. The extra heat and added wear on the motherboard and dram of running high voltage 24/7 is just not worth it.

Thats why i much preferred my 8700k over the 11600k. At stock yeah, the 11600k was obviously faster, but when both tubed the 8700k was a beast. I got unlucky as well, my 11600k didn't wanna boot in gear 1 above 3333 mhz

 

[KapiteinKoek007]

Thats why i much preferred my 8700k over the 11600k. At stock yeah, the 11600k was obviously faster, but when both tubed the 8700k was a beast. I got unlucky as well, my 11600k didn't wanna boot in gear 1 above 3333 mhz

Haha yeah i really struggled to not buy an 8700k when it released, but I still was rocking a 2600k at 5.1 ghz, that Sandy bridge cpu was honestly the best I ever had, it held its own against all the games I threw at it in 10 years time, up untill last year when it couldnt keep up with current AAA titels. I thanked it for its 10 year service and retired it. Its now resting in the hall of fame . I like my current 11600kf, but its not the same love as I had with my 2600k, intel took all the fun out of overclocking with 11th gen cpu because they saw they were missing a big piece of the market by factory overclocking their already outdated 14nm node to extend its life because 10nm node was giving intel headaches at that time to say the least.

 

[dgianstefani]

A good memory OC will increase 1% and 0.1% lows much more than the average FPS.

 

[KapiteinKoek007]

A good memory OC will increase 1% and 0.1% lows much more than the average FPS.

Can be usefull when my system is struggling to stay at 60 fps for example, but since im still at 1080p, thats not gonna happen for awhile. Most current AAA titles I play have an average of 100 to 150fps. And never drop below 60

 

[dgianstefani]

When I moved from 5950x to 5800x3d my lows went from 60 to over 144. Shows the difference cache and memory makes.

 

[bug]

hello,

I have been wondering, which is better, higher clock speeds with loose timings or lower latency with tight timings. What do you think ?
I did some testing and benching on my own system results are in attachement, (excuse the format in which it is in)

Long story short, ideally you want both. If you have to settle, that's highly workload dependent. If you don't absolutely need to squeeze every last drop of performance from your system, just buy whatever RAM is the sweet spot of the day and call it a day.

 

[KapiteinKoek007]

When I moved from 5950x to 5800x3d my lows went from 60 to over 144. Shows the difference cache and memory makes.

The 5800x3d does have double the L3 cache, so that would make sense. But these kind gains wouldnt be beneficial for me because my 0.1% and 1% rarely drop below 80 in most games and never below 60, so the added wear on the motherboard and dram because of heat and high voltage just to add a few fps in the area's where I dont notice the difference, is not worth it

 

[ir_cow]

Did you run any stability tests when increasing the frequency? Often you can boot into windows, but it sure isn't stable. AIDA64 or other software will give false results do to all the WHEA errors. Just something to think about for future testing.

 

[KapiteinKoek007]

Did you run any stability tests when increasing the frequency? Often you can boot into windows, but it sure isn't stable. AIDA64 or other software will give false results do to all the WHEA errors. Just something to think about for future testing.

For stability test I use memtest86 DOS version, when i use aida for benching I dont test stability because it only needs to be stable for the bench, if i then decide I want to use that overclock for 24/7 use then ill do stability testing.

 

[Wirko]

There's an older thread here with many results, details, and leaderboards:

How Low Can You Go? Memory Latency Competition - AIDA64

Leaderboards are located at the bottom of this post - *Updated 3/14 4:00am* I thought it would be cool to get a "pulse" on some of the newer systems and see how modern platforms perform in terms of memory latency. We can compare and contrast to older generations as well (DDR3 vs DDR4 vs dual...

www.techpowerup.com

@KapiteinKoek007 , the nanoseconds you're getting amount to 98 clock cycles at 3200 MT/s and 124-140 clock cycles above that. It's obvious that for AIDA64, other latencies matter more for the result than CL does.

Aida's 60-70 ns is probably the maximum latency, the one you get if you read minimum amounts of data (64 bytes) from completely random locations, with no sequential access at all. It's not wrong, RAM latency (and SSD latency alike) can be measured in many ways, but it's no surprise if the results are of little significance to you.

(Also, I never forget to mention that my home computer, of a very common type, had 250 ns DRAM latency back in 1984.)

 

[ir_cow]

Eh. Good enough. Memtest86 is a good baseline and catches most unstable memory setups. 1-4 test is Timings / frequency related, 5-6 is IMC and the rest is mixture. This goes out the window when your pushing the frequency and IMC to the max though. Say at 5000 MT/s, a error in test 3 could be a IMC voltage related issue or DRAM voltage / timings.