Monday, June 27th 2022
Windows Defender can Significantly Impact Intel CPU Performance, We have the Fix
Kevin Glynn, aka "Uncle Webb," our associate software author behind popular utilities such as ThrottleStop and RealTemp, developed a new utility named Counter Control, which lets you monitor and log the performance counters of Intel Core processors since 2008 (Core "Nehalem"). During development for ThrottleStop, Kevin discovered a fascinating bug with Windows Defender, the built-in security software of Windows, which causes significantly higher performance impact on the processor than it should normally have. Of course a security software is bound to have some (small) performance impact during real-time protection, but this is much bigger.The first sign that something is happening is that HWiNFO will be reporting a reduced "Effective Clock" speed when the CPU is fully loaded. A much bigger problem is that when Defender is affected by the bug, performance of your machine will be significantly reduced. For example, a Core i9-10850K running at 5.00 GHz all-core loses 1000 Cinebench points (or 6%). Such a performance loss has been reported by owners of Intel Core 8th, 9th, 10th and 11th Gen, both desktop and mobile CPUs, on both Windows 10 and Windows 11. AMD processors are not affected.
The underlying issue that costs so much performance is that Windows Defender will randomly start using all seven hardware performance counters provided by Intel Core processors, which includes three fixed function counters. Each of these counters can be programmed in one of four modes, to configure at which privilege level it counts—Disabled, OS (ring-0), User (ring>0), and All-Ring levels. Since these counters are a shared resource, it is possible that multiple programs want to access these counters at the same time.
Popular system utilities like HWiNFO, OCCT, Core Temp, and ThrottleStop, all set these counters to "mode 3" or "All-Ring Levels." Since they all set the same mode, there's no issues with multiple programs using the same counter. Windows Defender on the other hand will set these counters to "mode 2", at what looks like random intervals, for random durations of time. This can happen when a computer first boots up or it can happen at any time after that. While Windows Defender is running in the background, it can start and stop or continuously try to change these counters to mode 2 at any time. Just to clarify, the performance loss will happen even without any monitoring software running—Defender will still use excessive CPU time.
The issue is not with the Intel hardware, as setting the same timers as Windows Defender manually has no negative performance impact. Also, if these counters are manually overwritten, Defender detects that, immediately stops whatever it is doing and performance returns to normal—without any negative effect on the ability to detect viruses in real-time.Our Counter Control software monitors and logs the "IA32_FIXED_CTR_CTRL" register of Intel Core processors, located at MSR 0x38D. This register provides access to the three fixed-function performance monitoring counters mentioned before. Counter Control will inform users if any software is using the Intel fixed-function counters, and for how long they've been in use. Typical values reported by Counter Control look like this:
If "Windows Defender Boost" is not checked, the counter will be initially cleared. This stops the Window Defender algorithm but ThrottleStop will no longer try to keep one counter running while using ThrottleStop and it will not keep that one counter running after you exit ThrottleStop. This allows a person to use ThrottleStop without having to worry that ThrottleStop might be doing something to Windows Defender that it should not be doing. After ThrottleStop starts up, if that timer is not being used, after 10 minutes or so, Windows Defender will check that timer, see that it is not being used, and will be able to start its mysterious performance-eating algorithm again.
Let us know your experience in the comments of this article. It'll be interesting to see how widespread this issue is, we have confirmed (thread at TPU, thread at OCN) it to be happening on many systems in recent months. If we make enough noise, I'm sure Microsoft will look into why they need that many timers in Defender, why there's such a big performance hit, and fix it accordingly.
As always, let us know your thoughts and questions in the comments. Also let us know if you didn't understand certain technical details, so we can improve this writeup.
Counter Control is available as free download in our downloads section.
The underlying issue that costs so much performance is that Windows Defender will randomly start using all seven hardware performance counters provided by Intel Core processors, which includes three fixed function counters. Each of these counters can be programmed in one of four modes, to configure at which privilege level it counts—Disabled, OS (ring-0), User (ring>0), and All-Ring levels. Since these counters are a shared resource, it is possible that multiple programs want to access these counters at the same time.
Popular system utilities like HWiNFO, OCCT, Core Temp, and ThrottleStop, all set these counters to "mode 3" or "All-Ring Levels." Since they all set the same mode, there's no issues with multiple programs using the same counter. Windows Defender on the other hand will set these counters to "mode 2", at what looks like random intervals, for random durations of time. This can happen when a computer first boots up or it can happen at any time after that. While Windows Defender is running in the background, it can start and stop or continuously try to change these counters to mode 2 at any time. Just to clarify, the performance loss will happen even without any monitoring software running—Defender will still use excessive CPU time.
The issue is not with the Intel hardware, as setting the same timers as Windows Defender manually has no negative performance impact. Also, if these counters are manually overwritten, Defender detects that, immediately stops whatever it is doing and performance returns to normal—without any negative effect on the ability to detect viruses in real-time.Our Counter Control software monitors and logs the "IA32_FIXED_CTR_CTRL" register of Intel Core processors, located at MSR 0x38D. This register provides access to the three fixed-function performance monitoring counters mentioned before. Counter Control will inform users if any software is using the Intel fixed-function counters, and for how long they've been in use. Typical values reported by Counter Control look like this:
- Not Used - 0x000: The three fixed function counters are stopped. None of the counters are presently being used.
- Defender - 0x222: All three fixed function counters are programmed to mode 2. This is the value that Windows Defender sets these counters to when it is using them.
- Normal - 0x330: Two counters are programmed to mode 3. One counter is programmed to mode 0 and is not being used. This is normal. Most monitoring programs that use these counters will program the counter control register to this value.
- Warning - 0x332: This is shown when two counters are being used normally by monitoring software while the third counter has been set to mode 2, likely by Windows Defender. This is a warning that two different programs might be fighting over control of the shared counters. You might see the counter control register constantly changing between 0x222 and 0x332. This is what you will see when running HWiNFO if Windows Defender is trying to use the IA32_FIXED function counters at the same time.
If "Windows Defender Boost" is not checked, the counter will be initially cleared. This stops the Window Defender algorithm but ThrottleStop will no longer try to keep one counter running while using ThrottleStop and it will not keep that one counter running after you exit ThrottleStop. This allows a person to use ThrottleStop without having to worry that ThrottleStop might be doing something to Windows Defender that it should not be doing. After ThrottleStop starts up, if that timer is not being used, after 10 minutes or so, Windows Defender will check that timer, see that it is not being used, and will be able to start its mysterious performance-eating algorithm again.
Let us know your experience in the comments of this article. It'll be interesting to see how widespread this issue is, we have confirmed (thread at TPU, thread at OCN) it to be happening on many systems in recent months. If we make enough noise, I'm sure Microsoft will look into why they need that many timers in Defender, why there's such a big performance hit, and fix it accordingly.
As always, let us know your thoughts and questions in the comments. Also let us know if you didn't understand certain technical details, so we can improve this writeup.
Counter Control is available as free download in our downloads section.
257 Comments on Windows Defender can Significantly Impact Intel CPU Performance, We have the Fix
Start and exit. It is very easy.
However, I do not know how to use Counter Control.
Do I start Counter Control, click "Reset Counters" and exit?
If so, will the Defender problem be permanently fixed as long as the PC is running?
I've learned messing with windows the hard way.
If it can't be disabled or uninstalled without an app, leave it alone or risk messing up future updates.
Zeb
I've seen Windows Defender use 100% of a core on low-powered devices like old Bay-Trail Celerons etc. 100% of one core is either a quarter or half of the entire performance those devices have to offer.
The inclusion of Defender on Windows Server 2019 and 2022 is also a welcome addition.
Any sufficiently popular software will more likely be targeted as you're guaranteed to have a much bigger attack surface. More chance of success, more reward for the hacker.
What is Control Flow Guard in Windows; How to turn it On or Off (thewindowsclub.com)
How to Disable Virtualization-Based Security (VBS) in Windows 11 | Beebom
- you can also just disable virtualization in the bios to do the same effect and then test performance.
The point I was trying to make is that security apparatus in windows looks like it might be a 5% hit, and is in some very synthetic benches, but it's closer to 15% in real games when taking into account IO/memory latency with just those two settings. The system virtualizes the kernel and then nannies memory access with extreme overhead.
We have a juicy chunk of performance left on the table when it comes to OS and software optimizations. Will be interesting to see also if CPU accelerators will come into play for some of these tasks.
Source: Me. I'm well known as a ME security researcher.
15500 vs 16700 Cinebench R23 score. I start TS with windows, and use .ini to stop 5 sec later.
Read the op and see for yourself here's the first clue
Not to many people stare are task manager especially when doing other things.
here my log
2022-06-27 13:38:28 00:13:48 0x330 Normal
2022-06-27 13:42:37 00:03:07 0x000 Not Used
2022-06-27 17:15:47 03:33:10 0x330 Normal
2022-06-28 12:31:11 00:00:43 0x000 Not Used
2022-06-28 19:17:36 00:01:47 0x330 Normal
Plus running CB23 before and after I get same score, even after reset counters.
when running CB23 I get 100% load so as I said before it doesn't seem to affect 12th gen or at least my 12th gen config.
As far as sleep goes I always disable sleep with powercfg /hibernate off right after windows install so can't really test that, never use sleep.
I just let the monitor go into low power mode