ISLC and its impact on memory

[Parisien_entraide]

Hello

After various tests with LatencyMon, both with and without ISLC, I noticed that it causes me to lose memory bandwidth in the MOC test (I’m not providing the link because I don’t want to be considered a spammer).

On the other hand, I do see some benefits elsewhere. For example, the usual Nvidia driver no longer appears as the main source in LatencyMon (2-minute-30-second test).

---------------------------------------------------With --------------Without ISLC---------------Différence

Read ------------------------------------------ 56,1 Go/s-----------72,5 Go/s---------------+29 % sans ISLC

Write ------------------------------------------67,0 Go/s---------- 81,2 Go/s ---------------+21 % sans ISLC

Copy -------------------------------------------34,9 Go/s-----------37,9 Go/s-----------------+9 % sans ISLC


This board : With ISLC - Without ISLC - Withou ILSC but with the program "Timer Resolution" (0.500ms)

tableau_comparatif_latencyMon.png

There was no Standby List cleanup. Here is my ISLC configuration:

Performance Timer: TSC (Synthetic 10 MHz)
Current Timer Resolution: 0.504 ms
Maximum Timer Resolution: 0.5 ms
Minimum Timer Resolution: 15.625 ms
Desired Timer Resolution: 0.5046 ms (set manually following the test from TimerResBenchmark)

The following options are enabled in ISLC:

Enable custom timer resolution
Use GlobalTimerResolutionRequest


HPET

uselegacyapicmode No
useplatformclock No
useplatformtick Yes
disabledynamictick Yes




CONTEXT

My configuration

- MSI Z890P P WiFi (everything is up to date; BIOS and drivers from MSI sources)
- CPU: i7-265KF
- RAM: 32 GB DDR5 Corsair Vengeance CMH32GX5M2B6000C38, 6000 MHz, CL38 with XMP (38-44-44-96-140)
Intel 200S Boost enabled, as well as the new RAM Performance Mode
“This memory extension mode optimizes DDR5 sub-timings, reduces latency, and increases performance by up to 14%” (MSI source), depending on the profile. I selected the High Efficiency profile, which is the maximum setting.
For example, the tRFC setting changes from 480 (idle) to 140 under load (the maximum supported by the RAM), depending on the Bitsum ParkControl profile.
I’m using the Intel-recommended default settings for PL1 / PL2.
- Everything is managed by Intel’s AI NPU.
- Microsoft Windows 11 Professional (x64) 25H2 Build 26200.9168, fully up to date
- WD_BLACK SN8100 2 TB PCIe 5.0, installed in a PCIe 5.0 slot
- Nvidia RTX 5070Ti 16GB
- HAGS enabled, MPO, etc.

- I use reliable tweaking tools such as WinHance (I’m aware of the impact of the modifications I make).

In all of my tests, I experienced no crashes or instability, including while gaming.

---------------

Edit : Interessant...

https://randomascii.wordpress.com/2020/ ... le-change/
https://learn.microsoft.com/en-us/windo ... eginperiod

[Wagnard]

Hello

After various tests with LatencyMon, both with and without ISLC, I noticed that it causes me to lose memory bandwidth in the MOC test (I’m not providing the link because I don’t want to be considered a spammer).

On the other hand, I do see some benefits elsewhere. For example, the usual Nvidia driver no longer appears as the main source in LatencyMon (2-minute-30-second test).

---------------------------------------------------With --------------Without ISLC---------------Différence

Read ------------------------------------------ 56,1 Go/s-----------72,5 Go/s---------------+29 % sans ISLC

Write ------------------------------------------67,0 Go/s---------- 81,2 Go/s ---------------+21 % sans ISLC

Copy -------------------------------------------34,9 Go/s-----------37,9 Go/s-----------------+9 % sans ISLC


This board : With ISLC - Without ISLC - Withou ILSC but with the program "Timer Resolution" (0.500ms)
tableau_comparatif_latencyMon.png




There was no Standby List cleanup. Here is my ISLC configuration:

Performance Timer: TSC (Synthetic 10 MHz)
Current Timer Resolution: 0.504 ms
Maximum Timer Resolution: 0.5 ms
Minimum Timer Resolution: 15.625 ms
Desired Timer Resolution: 0.5046 ms (set manually following the test from TimerResBenchmark)

The following options are enabled in ISLC:

Enable custom timer resolution
Use GlobalTimerResolutionRequest


HPET

uselegacyapicmode No
useplatformclock No
useplatformtick Yes
disabledynamictick Yes




CONTEXT

My configuration

- MSI Z890P P WiFi (everything is up to date; BIOS and drivers from MSI sources)
- CPU: i7-265KF
- RAM: 32 GB DDR5 Corsair Vengeance CMH32GX5M2B6000C38, 6000 MHz, CL38 with XMP (38-44-44-96-140)
Intel 200S Boost enabled, as well as the new RAM Performance Mode
“This memory extension mode optimizes DDR5 sub-timings, reduces latency, and increases performance by up to 14%” (MSI source), depending on the profile. I selected the High Efficiency profile, which is the maximum setting.
For example, the tRFC setting changes from 480 (idle) to 140 under load (the maximum supported by the RAM), depending on the Bitsum ParkControl profile.
I’m using the Intel-recommended default settings for PL1 / PL2.
- Everything is managed by Intel’s AI NPU.
- Microsoft Windows 11 Professional (x64) 25H2 Build 26200.9168, fully up to date
- WD_BLACK SN8100 2 TB PCIe 5.0, installed in a PCIe 5.0 slot
- Nvidia RTX 5070Ti 16GB
- HAGS enabled, MPO, etc.

- I use reliable tweaking tools such as WinHance (I’m aware of the impact of the modifications I make).

In all of my tests, I experienced no crashes or instability, including while gaming.

---------------

Edit : Interessant...

https://randomascii.wordpress.com/2020/ ... le-change/
https://learn.microsoft.com/en-us/windo ... eginperiod

I suspect the timer resolution, try without the timer resolution set. Technically a higher resolution(<1ms) is more demanding. If its not helping the bandwidth, try changing the polling rate to a way longer one to see if it help. Maybe polling too often the memory is causing this.

[Parisien_entraide]

Hello,

I did a lot of tests with ISLC, BIOS settings, Windows etc... So it was a bit long.


The result:

The problem did not come from ISLC, but from the test program used which should only act on a single CPU, etc... With AIDA latest version everything is correct


Without ISLC (AIDA64 Test)

Read: 95,132 MB/s,(about 99% of the theoretical read)
Write: 88,094 MB/s, (about 92% of the theoretical writing)
Copy: 89,299 MB/s, (about 93% of the theoretical copy)

With ISLC:
Read 95 854 MB/s ( 99% of the theoretical read)
Writing; 88,101 MB/s (92% of the theoretical writing)
Copy: 89,790 MB/s (94% of the theoretical writing)



I had tested with 1ms and more and other parameters but it didn't change anything. So the best for me is:

For information my ISLC configuration

2026-08-23_163759.jpg

[Wagnard]

Thanks, for your time on this, that reassuring.

[Parisien_entraide]

That’s normal—I had highlighted something involving ISLC, so I had to find out why

Otherwise, to provide some additional context:

As a side note, I manually apply the “High performance” power plan because, with my version of Core Parking (I have the paid version), the Bitsum mode switches between a normal/idle mode and the specific Bitsum power profile under heavy loads. During AIDA tests, the load is not high enough to trigger the lower-latency mode and achieve better benchmark scores.

HPET

• uselegacyapicmode : No
  useplatformclock : No
  useplatformtick : No
  disabledynamictick : Yes

Setting useplatformtick to “Yes” was sometimes recommended for very old systems to solve audio or video synchronization issues—and it did work for me. However, on Windows 11 25H2 ans hardware released from 2024/2025 onward, it is counterproductive.

It should therefore be left set to “No” to achieve the lowest possible latency, rather than following current optimization “tutorials” that claim the opposite.

[asprunes]

That’s a good clarification. In my experience, power-plan and timer settings can interact in ways that make isolated benchmark results misleading. If Core Parking is dynamically switching profiles based on load, the benchmark may not actually be testing the same scheduling state every time. I also agree that older HPET recommendations shouldn’t automatically be applied to modern Windows 11 systems. The safest approach is to establish a baseline with default settings and then change one variable at a time, checking both latency and real-world performance.

[Parisien_entraide]

I tested many different possible combinations without actually overclocking the PC. (Obviously, running the RAM in XMP mode is technically an overclock, but that's the only thing I changed.)

As for the rest, I disabled HDCP to reduce latency a little, stopped unnecessary services, cleaned things up (thanks, Autoruns ), and when I launched WinHance, everything was already set up for the optimizations. (Same thing with WTOOLS, by the way.)

Obviously, ReBAR is enabled, as are MPO, HAGS, etc.

In the end, I never have any issues in games. Even with my previous PC, I never had any problems either. This time, though, I had to dig deeper and look closely at the power plans, how DDR5 was being managed, etc. The HPET settings were important as well.

What I've learned now is that, when properly configured, ISLC helps improve RAM read, write, and copy performance, and especially overall responsiveness. There were also some improvements in latency, and I no longer have the Nvidia driver causing problems.