specs:
i7-8700k delidded
32GB of 4x8GB (XMP is 2666 at C16) - two different kits so poor overclock ability, motherboard is t-topology not daisy chain, one kit is micron b-die the other samsung c-die
Windows 10 20H2 (19042.685)
game is on public beta running 4.0 beta 4
Game and OS are installed on a gen3 NVMe (trying to reduce bottlenecks, unlikely to matter in this case)
RTX 3090 (worthless for this test)
for CPU:
Stock = 4.3Ghz, cache 4.1Ghz (auto voltage, default bios)
OC = 5.0Ghz, cache 4.7Ghz (35% improvement, unlocked cache/uncore multiplier set to core multiplier minus 3)
1080p windowed, low preset, 4.0 beta 4 unless specified
all FPS ranges were decided by waiting for the load to stabilize for roughly 10 seconds and then eyeballing my overlay, if the FPS increased but did not remain for longer than a second I did not count it, if it dipped and the FPS had already stabilised, I counted it.
Code: Select all
8700k@stock 2133mhz C15-15-15-36(stock) // Stock CPU Stock Ram, 14.064ns latency
young 123-126
empty 20-21
dense 123-126
8700k@OC - 2133mhz C15-15-15-36(stock) // Overclocked CPU, Stock Ram, 14.064ns latency
young 125-129
dense 23-25
empty 123-126
Code: Select all
8700k@stock - 2666mhz C16-18-18-35(XMP) // XMP 12.00ns ram latency
young 125-128
dense 26-28
empty 126-130
8700k@OC - 2666mhz C16-18-18-35(XMP) // XMP 12.00ns ram latency
young 135-138
dense 28-30
empty 135-139
I stopped bothering to check stock clocks around this point to save time, stock comparison later on in my post against final memory OC.
Code: Select all
8700k@OC - 3200mhz C16-20-20-40(OC) // 10.00ns ram latency
young 143-145
dense 28-30
empty 141-144
8700k@OC - 3333mhz C16-20-20-40(OC) //9.6ns
young 145-146
dense 29-30
empty 144-148
this is where I left my RAM alone and started tightening timings, these aren't tight by many standards but compared to stock on the secondaries and things like tRFC and tREFI I did pick up a noticeable performance, albeit diminishing, total core to ram latency is 52.0ns benchmarked by aida64 (I have not yet adjusted tXP which can significantly improve latency in benchmarks such as aida64)
Code: Select all
8700k@OC - 3400mhz C17-22-22-42 // 10.00ns
young 145-147
dense 30-31
empty 143-145
8700k@stock - 3400mhz C17-22-22-42 // 10.00ns
young 139-142
dense 26-27
empty 138-140
up until now the die characteristics of each set did not effect the overall overclock other than loose timings, from here out I have removed one set of ram leaving 2x8gb in dual channel mode (c-die)
the loss of FPS observed here compared to 10ns runs earlier could be explained by t topology layout of motherboard or the loss of 16gb of ram from the system itself
Code: Select all
8700k@OC - 2x8gb 3466mhz C17-21-21-40 // 9.8ns
young 143-145
dense 28-29
empty 139-141
My steam client didn't like me changing the RAM out and detected that I was on a new computer, this reset my beta branch so for one run I did 3.30
Code: Select all
8700k@OC - 2x8gb 3600mhz C18-21-21-40 //10ns <!3.30 !>
young 138-140
dense 24-25
empty 137-139
8700k@OC - 2x8gb 3600mhz C18-21-21-40 //10ns <!4.0 beta 4!>
young 143-145
dense 28-29
empty 139-142
all memory settings were tested against to ensure they were stable and running well, my final memory stability was checked with Karhu RAM Test to 6400% (which is proposed to be 99.41% certain), all memory tests were done with CPU cache enabled to both ensure my cache/ring ratios were stable. I currently have a fan blowing over the RAM to ensure stability(this was not necessary however I am using a SFF case so the heat soak from my GPU can raise the RAM temperatures to uncomfortable ranges), RAM can be notoriously unstable at temperatures above 45c and I would not recommend just anyone go overclock their RAM without both understanding it from an in-depth guide and having adequate cooling on their case.
It's quite an interesting relationship between the overclocked CPU and the newly overclocked RAM, at a glance I assumed the stronger Single Core Performance (SCP) introduced by overclocking my processor is a significant gain towards the FPS however that is not the likely factor, by overclocking the CPU I was able to increase the cache/ring ratio to match my CPU Overclock at 4.7Ghz (CPU Multiplier minus 3).
The SCP increase appears to not have come in as a factor due to how loosely timed my RAM is, I believe I am near my limit of my processor itself yet still bottlenecked by my IMC, the observable differences in FPS without the CPU overclock in the dense save I would like to attribute to the lower ring ratio used on the CPU which ultimately bottlenecks my system however I am in some scenarios nearing the lows of 90% utilization on some cores which indicates even if I was to replace my RAM I would very soon hit that single core bottleneck.
My final thoughts would be, do not cheap out on your RAM, target under 9ns of latency on the kit you buy. A simple formula to work out latency at a glance is 2000 times CAS latency divided by RAM frequency so 2000 * C17 / 3400mhz is 10.0ns
I will return soon with other comparisons done but for now that is all the information I have to share, if you feel something was done erroneously please reply and I will address it.
Thanks, Hobo.