Really no reason to try and optimize your system just for a better Cinebench score. The only time I run it, is when I build a new system or upgrade a system. I run it along with other benches (3DMark, AIDA64 Extreme, etc.) just to make sure I'm getting appropriate scores. Anyway, have fun.
Really no reason to try and optimize your system just for a better Cinebench score. The only time I run it, is when I build a new system or upgrade a system. I run it along with other benches (3DMark, AIDA64 Extreme, etc.) just to make sure I'm getting appropriate scores. Anyway, have fun.
I switched RAM in my NUC from Kingston Impact 3200 CL20-22-22 to Patriot CL18-19-19 (self-modded SPD). It's about 5GB/s and 5ns latency improvement. Both are dual-rank Hynix. The scores went up from 603 to 606 in multi-core and from 99 to 101 in single-core. Totally worth it
I switched RAM in my NUC from Kingston Impact 3200 CL20-22-22 to Patriot CL18-19-19 (self-modded SPD). It's about 5GB/s and 5ns latency improvement. Both are dual-rank Hynix. The scores went up from 603 to 606 in multi-core and from 99 to 101 in single-core. Totally worth it
11700k set to 250W power limit. 2024 doesn't go near 200W so this is practically unlimited. I do see 2nd+ runs within a set scoring slightly lower than 1st, so there is probably still a little thermal effect going on. We're talking about <2% in general, although I need to run more tests to check.
Ram is XMP 3200 dual rank.
1: XMP 3200 dual channel: 820
2: XMP 3200 single channel: 745
3: SPD 2133 single channel: 627
I could hear the CPU fan spin slower as I reduced the ram settings, so the CPU can't work as hard if it can't get the data fast enough.
Going from step 3 to 2, that's +50% peak bandwidth for +19% score. Going from step 2 to step 1, that's +100% bandwidth for +10% score.
I'll need more data points to try and work out how the scaling goes. To remove the CPU thermal element I might repeat with turbo off for example. I just wish the test doesn't take so long.
I don't have enough density of single rank ram to see if rank makes much of a difference vs 2R. Checking timing influence might be the last thing I do.
Got round to charting the results I can use. Others were variable clock so can't be put on this chart. Vertical axis is an indication of points normalised for CPU cores and clock. Horizontal axis is an indication of the ram potential vs CPU potential. Basically further to the right means more ram potential relative to CPU potential. Hmm... looking at the chart some more, I could have simplified it more. The CPU cancels out since it is on both axis so it is basically score vs ram potential. However if I later decide to compare different CPUs on the same chart, CPU normalisation will be required.
The data points are, bottom left to top right:
CPU 3.6 GHz, ram 2133 single channel
CPU 3.6 GHz, ram 3200 single channel
CPU 3.6 GHz, ram 3200 dual channel
CPU 1.8 GHz, ram 3200 single channel
CPU 1.8 GHz, ram 3200 dual channel
The ram I used was dual rank. All 3200 runs were at the same XMP timings. The 2133 run was JEDEC timing so may be worse than just speed alone.
By reducing the CPU clock, we're increasing the ratio of ram potential so I can extend the chart to the right without insane speed ram. It's pretty flat on the right which is what we might see with older versions of Cinebench. However with higher end faster CPUs, especially with only dual channel DDR4, there may be some limiting going on. The higher bandwidth of DDR5 might put that off somewhat, but I can't say without more data.
The two dots in the top middle ideally should match up. There's actually a 3% difference which is about on the limit of run to run variations, especially as I didn't do that many runs given how long it takes. I was just happy if I got two runs at each setting within 2% of each other. We could get into other theories like caches playing a part, but again there isn't the data to support that.
A very rough result for my 7920X, assuming an average clock of 3.6 GHz puts it around 280, 19. Since its clock varies during a run this wont be exact. This implies 11700k has about 20% higher IPC in this use case, which might not be far off.
Edit: why didn't I think of it earlier. The single thread results probably wont be ram limited to any significant degree and may give a better "unlimited" indication. Could take a while to run though?
This site uses cookies to help personalise content, tailor your experience and to keep you logged in if you register.
By continuing to use this site, you are consenting to our use of cookies.
