Comparison of Windows 11 scheduler behavior for AMD vs Intel processors in computing tasks
Comparison of Windows 11 scheduler behavior for AMD vs Intel processors in computing tasks
I manage multiple PrimeGrid jobs. Several resemble Prime95 and allow parallel processing. SMT/HT offers no advantage here, and in fact it may raise power usage without improving speed. The best setup is for each task to run on a dedicated core. In short, let's compare two comparable systems: 11700k and 7800X3D. Both are fully updated with Windows 11. All chipset and memory drivers are installed. On the 11700k I usually assign one thread per core, giving 8 threads for 8 cores. Without changes, the workload spreads across all 16 CPU threads, which isn't always balanced. This leads to a loss of efficiency—about a 10% drop in throughput. I call this "core smear" because it cuts performance by roughly ten percent compared to an optimised setup. The fix is either to turn off HT or apply core affinity so only BOINC gets one thread per core. On the 7800X3D, running eight threads still avoids the same imbalance, allowing full utilization of some cores and boosting speed. Running more than eight threads can overload the cores, lowering efficiency too, but it also wastes power. It's interesting how these systems handle threading differently; I'm curious about the design choice behind this behaviour. There seems to be an attempt to cluster threads from similar code for possible synergy and modest power savings. Since I don't own hybrid AMD or Intel chips, this will be a new challenge—especially if future versions require mixed performance profiles.
Are CPPC and CPPC preferred cores activated in the BIOS? Just having those two on can lead to similar problems even with older systems like W10.
Bios defaults are the standard settings, whatever Asus chooses. I plan to check what they’re configured for and how it affects things. It seems likely they use some priority system. I’m expecting a bit of confusion about my AMD CPU setup. Before the 7800X3D, I was using Zen 2 (maybe 3600), which had issues managing CCX—though I don’t remember that problem then. That probably changed around Windows 10. I own a Zen 3 laptop, but I rarely run heavy CPU tasks on it, so I haven’t experienced it there.
I’m having trouble locating that setting in BIOS, either by scrolling through the menus or using the search tool. It might not exist anymore, or it could be listed under a different title. The only name I saw was CPPC Dynamic Preferred Cores with four options: Frequency, Cache, Driver, Auto. I left it on Auto, which is what I set initially. Mobo is an Asus B650E-F running BIOS 3279, the most recent version checked last week. My guess is it might be trying to adjust performance settings, so Power Performance Plans could play a role. I’m currently considering switching to High Performance mode to test its effect. Now, what’s the distinction between Performance and Ultimate Performance? Do they represent default Windows options, or are they custom profiles AMD didn’t create? Edit: I’ve started researching further, but it seems this is just another complex issue! I’m still trying to figure out what CPPC actually does and how it influences core performance within a single CPU.
CPPC works with all Zen3 and newer CPUs. It doesn’t depend on whether you have an asymmetrical ccx setup or a 96MB L3 cache. It mainly tells the OS to focus more on certain cores because they either run faster or handle tasks better, though I’m not entirely sure about that part.
On Intel 11700K the scheduler usually distributes threads across physical cores first, which helps PrimeGrid achieve balanced loads after you set affinity. With Zen 4 7800X3D Windows tends to group threads onto SMT siblings before moving to other cores—a design choice aimed at gaming performance. This can cause issues for throughput tasks like Prime95 or PrimeGrid, as many cores sit idle and overall performance drops. Solutions such as disabling SMT in BIOS or forcing affinity per core work well, aligning with HPC practices of pinning MPI jobs to physical cores. No secret synergy exists; it’s a mismatch between the gaming-optimized CPU behavior and compute-heavy workloads. If AMD adds an option to prioritize core spreading, that would resolve the problem. Until then, affinity masks remain the best remedy.
It seems there are options to set affinity at startup for certain processes. You should keep SMT active, as it can provide advantages in those scenarios.
Right, that’s the trade-off, SMT on Zen4 X3D is great for lightly threaded / gaming, but the Windows scheduler packs threads in a way that kills pure throughput workloads like PrimeGrid. If you want SMT enabled for other cases but don’t want to hand-set affinity every run, you can automate it: Process Lasso Windows lets you define per-app CPU affinity masks, so BOINC/Prime95 always run 1 thread per core automatically. Or use start /affinity in a small batch script to launch PrimeGrid tasks pinned to physical cores. For more control, there’s the PowerShell Start-Process -ProcessorAffinity flag. That way you keep SMT available for the cases where it helps, but your compute workloads won’t get core packed and lose 30% perf. Until MS/AMD expose a scheduler toggle prefer core spread, this kind of per-app affinity automation is basically the cleanest fix.