New info:
was on the practice server, everything was fine, frame times CPU/GPU under 5msec.
Server ran out, had to quit and rejoin as the server does not "turn over" anymore.
Next I know stutter of the worst kind.
All I did was rejoining the practice server and now I have frame times of >5msec CPU and <15msec GPU?? These are averages, I think there were drops way below that.
View attachment 723295
I had to restart my PC to get into the >7msec GPU times again.
Restarting CM did not work:
How is that possible.
Zu Hülfe
@RasmusP
Sadly I have absolutely no idea what happened there
If it happens again, please drag the CSP Render Stats app bigger to get the details displayed. There should've been a few things with massively increased times.
My own pc had been running like poorly for a long time. Could not figure it out. All vitals, cpu, gpu, ram, were fine on the hardware monitor.
Turns out it was my harddrive. Even though there was a few hundred gigs free, on a 3 terrabyte hardrive that 20% free space isn't enough. Deleted a load of games I don't really play and got it back to 2 tb or about 66% full, ran the defrag and all is well now.
[...] a 3Tb toshiba mechanical. Has good stats on random read speeds too. But for fast transfers it seem to start bottlenecking at about 66% capacity.
Maybe the NVme SSDs are better but have had the same issue with a standard SATA ssd hitting 80-90%
This is wild! Apparently you had some massive fragmentation?
The issue is that the "needle" needs to jump around too much instead of actually reading/writing, if the disk becomes too fragmented. A full defrag should usually help.
SSDs have a similar issue, although not related to fragmentation at all. SSDs don't care for files being all over the place, but they have "Multi Level Cells":
tekmart.co.za
Modern SSDs either have
Triple
Level
Cells or
Quad
Level
Cells.
The more levels, the slower the access. To be fast, the controller writes everything in SLC-style until the write-process stops. Then the controller moves all the written data into TLC or QLC style.
If you don't have enough empty space left, the controller needs to write the data in MLC, TLC or even QLC style, which is a lot slower.
So you usually see "SLC Cache" or "DRAM Cache" in SSD specifications. And that's also why bigger SSDs are often faster than smaller ones. They simply have more "cache", if they aren't full.
Important to know:
HDDs will slow down both for writes AND reads, since the "needle" has to jump around.
SSDs will only slow down for writes, since they lose the SLC cache. Reads will always be the same speed.
And for SSD shopping:
Buy TLC, if possible with DRAM/DDR cache and check the "IOPS", not the standard transfer rate. I've never seen the transfer rate when copying files. It's not about PCI-e 3.0 or 4.0 or 5.0, it's about the SSD controller managing small files.
Another thing to notice: TBW (max writes before cells might die) is a lot higher, the fewer levels the cells have.
It's not really an issue anymore, but if I only pay 5€ more for 1.5x the lifetime, I'll happily pay the 5€. Doesn't matter if it's 20 vs 30 years
One shouldn't really need it, but the compression tool in ac manager can save a lot of disk space (from 250 to 50 gb for me).
The other thing I want to throw into this discussion are the "extra fx" in the custom shaders patch settings. From my experience, these settings have a huge impact on fps. Just disable extra fx and leave everything else as it is. This is a simple way to get some extra power for very demanding tracks/grids or some large traffic maps.
Thanks for the hint to the compression! Looks interesting!
About extra fx: I mostly agree, but I prefer to work with the brilliant presets feature of CM, since extra fx has TAA (Temporal Anti-Aliasing), which I couldn't race without.
Before TAA, I used 4x MSAA + forced SGSSAA via nvidia profiler and guardrails etc. were still shimmering/flickering due to aliasing.
The TAA of CSP is the best TAA I've seen yet. No noticeable ghosting and extremely smooth.
I only use 2x MSAA in the normal AC settings to free up GPU performance.
It needs some sharpening though, which I apply via the "graphics adjustments" extension:
But yeah, the rest of extra FX like Local Reflections, Ambient Occlusion, Light bounce etc. really tank the performance.
Local Reflections are 99% only on the graphics card though, so if you have enough GPU power but your processor is struggling, SSLR are still possible and look absolutely gorgeous.
But they are heavy on the graphics card!