pwn wrote: ↑21 Jun 2020, 07:24
memory latency is measured in ns .. Draw your own conclusions ..
I think it will be impossible to notice the difference between 70 and 50ns ..
And as said before, CPU core works at 5GHz(5 billion hertz a second) so, thats a clock cycle of 0.2ns 1/5000000000, if it takes ~ 50-70ns for your CPU to fetch data from RAM for every single read or write it does to memory, compare it to 1 cpu clock cycle, 100 times longer. All modern CPUs have 3 levels of caching to reduce that latency, surely theres a reason billion dollar companies on both sides focus so much on this latency to put 3 levels of reduction on all of their CPUs. And not to mention those caches are very small which means more than 90% of memory usage still is done from RAM. Oh, i forgot to say you have 8 of those cores, so 5Ghz * 8 all doing reads and writes from RAM.
Somehow people interested in input lag fail to realize that they use a system which doesnt run in their mind so their perception of nanoseconds/microseconds/milliseconds is irrelevant, yet its the most common flawed argument, to say such small numbers dont matter because they compare it against their perception of time. Once u scale the time to human perspective from CPU perspective it ends up being minutes, months and years.
Also, your CPU does millions of reads and writes a second, so multiply that latency by same amount and now you see why it matters so much.
Me and others posted more in
viewtopic.php?f=10&t=7033#p52618 worth reading!
deama wrote: ↑21 Jun 2020, 05:24
So let's say I had 16GBs with 14-14-14-28 timings. Let's say there was a precise way to measure input lag and I measured total input lag at 50ms.
If I got another ram set, but this time the timings were at 12-12-12-24, how much would the total input lag reduce by? 5ms? 1ms?
It depends, theres more timings than those 4, like 100+ more, there also some other things a lot of motherboard BIOS developers hide, such as RAM power down mode which adds 5-7ns to RAM latency, and there's also things like tRFC, tCR (Command rate- set it to 1, almost all RAM runs at 2 for bandwidth, not latency), tREFI(refresh interval - increase as much as u can) and many more.
RAM latency affects everything, every single memory operation your OS does, everything your game does, everything your devices do with RAM, so it affects input, ofc.
However tuning ram latency wont do much if you have bottlenecks in your latency pipeline. It doesnt matter if you run to the bus stop or walk to it if bus arrives in 3 hours (say CPU low clock\bloated OS, instead of 5GHz constant + cleaned OS)
Latency reduction of RAM scales with its command rate, power down mode and most important factor being which CPU you have, on AMD you cant be helped much, on Intel its a different story, 5GHz + low latency ram scales far better than 4.5GHz or 4.8GHz. Higher CPU clock on intel + lower RAM latency help eachother out in terms of responsivness/smoothness.
However by my testing i'd say on 4.8GHz+ Intel CPU(constant clocks, tuned bios, no C states or downclocking) 1ns of RAM latency is worth ~ 30-100 FPS of input lag reduction, thats total, meaning both in terms of responsivness/smoothness and actual pipeline latency reduction
Sparky wrote: ↑21 Jun 2020, 09:07
only so far as it has an impact on framerate. So, negligible.
Wrong, read above. Even if your framerate stays the same, RAM latency is the most impactful thing in entire PC after CPU clock/ucore speed that affects input lag, yet so many ignore it. To put this in simple perspective, RAM latency is equivalent of time it takes for you to remember something.
Also anyone reading this in the future, adjusting your timings wont do much in terms of responsivness if you dont disable RAM power saving/DRAM power down mode which causes it to be turned on and off to save power and adds huge latency penalty.
Not to mention most RAM overclocking is done for higher clocks and bandwidth, while optimizing for latency is opposite of bandwidth, so you need to aim for lowest RAM latency, use this calculator
https://notkyon.moe/ram-latency.htm