KB5132061 (Win 11 26H2): ~9% lower Cinebench R23 multi-core and ~7% lower CPU-bound FPS on Ryzen 7 9850X3D, A/B tested. Can anyone reproduce?

TL;DR: On my 9850X3D, KB5132061 costs ~9% in Cinebench R23 multi-core and ~7% in CPU-bound Cyberpunk 2077 (720p). Clock, power and temperature are identical with and without it, so the CPU simply does less work per MHz. Uninstalling the KB restores the score, reinstalling brings the drop back. Single-core is not measurably affected. I’m looking for Zen 5 (or Zen 4) owners willing to run a quick A/B, and for anyone who knows what in this KB could cause it.

System

  • Ryzen 7 9850X3D, MSI MAG B850 Tomahawk Max WiFi, BIOS 2.A64 (AGESA ComboAM5PI 1.3.0.1d), microcode 0x0B40403B
  • 2x16 GB Kingston KF560C32-16 at XMP 6000, FCLK 2000
  • RX 7900 XTX, AMD Chipset Software 8.09.28.1124
  • Windows 11 Pro 26H2: build 26300.9974 with KB5132061, 26300.9711 without
  • Same BIOS settings in every run: XMP only, PBO off, Curve Optimizer 0

Cinebench R23 Multi
Same system state for the first two rows: Memory Integrity off, Wi-Fi/BT/USB audio disabled, GPU driver only (no Adrenalin), HWiNFO logging every 2 s.

state runs scores median
KB uninstalled 4 22,583 / 22,861 / 22,972 / 22,868 22,865
KB reinstalled 3 20,633 / 20,791 / 20,982 20,791 (-9.1%)
with KB, normal daily state (day before) 5 20,915 / 20,582 / 20,854 / 20,938 / 20,923 20,915

In all runs: effective clock ~5,190-5,200 MHz, PPT ~137-140 W, Tctl ~82-84 C. Run-to-run spread on this machine is 0.2-0.6%, so a 9% gap is far outside noise.

Single-core: 2,198 with the KB vs 2,204 before the update, which is below my noise threshold for single-core (~1.8%).

Idle with HWiNFO running: without the KB 4.6% CPU / 30.7 W package, with the KB 9-13% / 43-57 W. My guess is more expensive kernel transitions, but that is a guess.

Cyberpunk 2077 built-in benchmark, 1280x720 (CPU-bound)
Same session state as above. The first run of each series is discarded as warm-up (it reads ~4% high on my setup), runs paired by position in the series.

run with KB without KB delta
2 202.43 223.95 +10.6%
3 201.70 215.60 +6.9%
4 199.00 209.30 +5.2%
median 201.70 215.60 +6.9%

Caveat: the without-KB series drifts down (224 to 209) and I don’t know why yet. Temperatures were flat at ~69-70 C. So treat the exact size as approximate. The direction is consistent, though: every run without the KB beats every run with it.

Ruled out, one variable at a time
Memory timings and FCLK, Curve Optimizer, background apps, antivirus, Memory Integrity (HVCI), AMD chipset and GPU software including the Ryzen Master SDK, Wi-Fi/BT/audio drivers, thermals.

Hypothesis (unverified)
The timing lines up with AMD’s TSA bulletin for Zen 5 (AMD-SB-7073). Does KB5132061 enable that mitigation on Zen 5, and is there a supported way to check or opt out of only that one? If anyone knows how to verify it, I’d like to hear it.

How to reproduce (~20 min)

  1. Pause Windows Update first, otherwise it reinstalls the KB.
  2. Boot, wait 5 minutes, close background apps. Run Cinebench R23 Multi 3 times with “Minimum test duration” set to Off.
  3. Optional: log HWiNFO (clock, PPT, Tctl). It costs a few % by itself, so keep it the same in both states.
  4. Settings > Windows Update > Update history > Uninstall updates > KB5132061. Reboot, wait 5 minutes, repeat step 2.
  5. Post your CPU, board, BIOS/AGESA, both build numbers and all scores, including runs that show no difference.

Reported to Microsoft (Feedback Hub, plus Microsoft Q&A thread 6071549) and AMD support. I’m back on the KB for daily use while I wait for answers.

HWiNFO logs and README: https://drive.google.com/file/d/1XP9TUIZqkmiS7g6XwEfWbIw1nTCBqGO5/view?usp=drive_link
Feedback Hub: https://aka.ms/AA1458oh

Formatted with AI help; all numbers come from my own runs and logs.

Check if VBS is getting turned on by default by the new KB.

To prevent it from getting turned on, disable SVM in BIOS.