I just had a frustrating morning with my 245KF. I have a huge Excel sheet from work that takes about 260 seconds to calculate on a Xeon 6248R (company server) with 48 threads. So I thought it would be a good idea to test how my 245KF does. WELL, initial impressions were great. I got the lowest time of 209 seconds with only 14 available threads for the computation workload. But then in an attempt to better my score with RAM overclocking (it booted up with 5600), I raised the clocks to 8000 which I know it can do easily with my RAM kit. That worked and was stable in y-cruncher so I benchmark the sheet and...the only way to describe what happened next is a whole string of expletives.
In my previous two or three benchmark attempts, all cores were being used and CPU utilization was hitting a full 100%. After the RAM overclock, the system went full wonky and started throttling the CPU utilization in between the computation so that it was fluctuating between 40 to 100% with major dips. I know about clock throttling but this is the first time I've seen a CPU do "utilization throttling" where the whole thread goes to sleep for a second or two, dropping the utilization across the cores. The time to complete came out higher than 300 seconds which was a HUGE regression.
So RAM overclocking is either broken in the BIOS (I'm on stock Gigabyte BIOS) or this CPU cannot work full tilt with RAM overclocking. There's also weird behavior in Win Server 2019 (so Win10 behavior may be similar) where the workload starts on all cores but after a few seconds the Lion Coves are put to sleep. So here's what happened:
Attempt 1: Lion Coves enjoying their nap during the entire run
Attempt 2: Reboot and this time, for about 10 second intervals multiple times during the run, the Lion Coves participated.
Attempt 3: Reboot and my hopes got raised when I saw the Lion Coves holding on for more than 30 seconds. That turned to dismay when they went to sleep after about 60 seconds and the rest of the run continued with the E-cores.
The workload is a simple sumif() of 76000 rows from data of 1 million rows.
My stupid guess is that without the Win2019 scheduler being hybrid aware, the Intel Thread Director is "learning" as it goes so maybe after about a month, it will start trusting the Lion Coves through the entire run. It's still a very weird behavior making me wonder, "What are they so scared of about Lion Coves running on an unsupported Windows Scheduler that they specifically designed logic to forcefully put them out of commission????".
I'm all ears for anyone with similar experiences. I'm in no mood to update the BIOS because I don't want to trade one fix for a myriad of new potential issues.