So I would to get some information concerning clock stretching and the zen4. I am running a 7900x with a per core pbo curve. I am getting some great single core performance with boosts up to 5.8+ on my best cores, with tangible gains on bench marks. My question slash issue is this, No matter how far I walk back my negative curve per core settings I get intermittent clock stretching. Is this just unavoidable? and I should move on? or is this something that can be completely resolved?
I have settings that I can share: I just thought I will start with this as an opener.
Clock stretching should not occur with curve optimizer, at least it historically didn’t with 5000
There was a bug if you didn’t change the CO to zero, save then disable CO it would keep the offset despite disabling it
So I returned my bios to defaults and ran some prime 95 small FFTs. I noticed that even running at default with boost enabled that some cores are experiencing some clock stretching intermittently. Is this normal behavior for zen 5? or a cpu in general. seems to begin about 6min into a workload.
from what i gather elsewhere if your using offsets on the pbo you need to adjust your load line calibration. (LLC) as the stretching could be a result of vdroop.
also clock stretching is apparently a safety feature?..
Its really odd, tonight I am going to run workloads with coreboost off, so that it will just be running expo profile, and no other overclocks or boosting behavior Ill see if the behavior continues. I have some screen recordings I will upload once I have them converted to mp4 format. here are a couple screen screen recardings with one with defaults with coreboost enabled , and the other with coreboost disabled both are prime 95 small FFT AVX 512 instruction
So I have a working theory about what I have been observing. What I believe what is occurring is a very quick change of workload for each core. Prime95 workers will occasionally cycle into a self test. I am not sure of the exact timings of the test cycles but I have observed the first test cycle happening about 6min into the workload. which coincides with my first observations of what I thought was clock stretching. And after I observed the prime95 activity further I have also observed the same behavior when the workers cycle into a self test, The effective clock will drop on that core briefly. And once the self test is complete and the worker loads the core with calculations again the effective clock will raise and show that the core is working at the given frequency. I believe this is happening to quickly for the core clocks to register.