Intel 12th / 13th gen CPU latency

[alcoholbob]

DPC latency is way higher on Alder Lake/Raptor Lake compared to Intel's older monolithic designs, this is why Moore's Law had that DAW guy on talking about the latency issues with hybrid CPU designs. This is more of a problem with modern hybrid CPU designs, both AMD's chiplet and Intel's big/little design have tradeoffs that make them worse than traditional monolithic designs in latency sensitive applications. And this is why Nvidia is still holding on to dear life for monolithic dies, but the reality is eventually the computing demands will eventually outstrip minor quality of life issues with end user latency and everyone will have to give up on monolithic dies eventually and live with the imperfect world of having a complicated thread director try to guess what's the best way to breakup and assign a workload.

 

[JoeRambo]

To improve system latency one needs to take active steps to do so:

1) Memory latency optimization helps everywhere
2) Making sure your CPU and Uncore clocks are alway at 100% while at same time keeping deeper power states. It takes eternity from CPU PoV to change clocks from 800Mhz to 6Ghz, cause they need to ensure voltage safety bands are followed and so on.
3) I/O power savers are especially nasty, something like PCIE power states are going to hurt bad
4) Looking around for DPC latency problems with tools - esp on USB side, every device needs to be accounted for in USB tree software. Your motherboard might have LED controller that dangles on USB and introduces 50ms stutters if driver is/is not installed. Power saving management and common sense tactics ( connecting devices directly to Intel's stuff, not some USB hub designed by idiots )
5) Some Windows tuning helps a lot

Those tips only go so far. I have been disabling HT since 10C Skylake. Now with 12th / 13th GEN i am also disabling marketing redacted cores.

The end result is incredibly snappy and responsive system, but it takes effort to achieve it.

The redacted word is used as a pejorative here in the U.S. and members have complained about it's use in this post. Hence, I removed it. I do understand you may have been using its textbook definition so not going to infraction you for it. -Moderator DAPUNISHER

 

[igor_kavinski]

The end result is incredibly snappy and responsive system, but it takes effort to achieve it.

If you don't mind, could you share a screenshot of the LatencyMon tool on your system so we know how a properly tuned system compares to the 10850K results shown in part 2 of the youtube video?

 

[FangBLade]

And what does frame chasers say?

 

[coercitiv]

Can't just be a Windows bug?

It sure can, but it can also be a number of other things. Here's my favorite, considering I own the board:

After spending hours and hours investigating the root cause of high latency in latencymon using ProcessMonitor, Windows Performance Recorder/Analytics. I've found the root cause and I profoundly hate MSI.

Disconnected every possible usb hub/device, removed driver and was using the computer with a PS/2 keyboard. Found out, using USBVIEW, that there was always one device left on my system. This device was registered as Bright Light.

MSI Z690-A PRO doesn't support built-in rgb but motherboard can act as a controller. Since I don't have any RGB I've never installed the MSI Mystic Light. No device were shown as uninstalled or anything. So I installed the horrible Dragon Center and Mystic Light and boom!! latency gone. Highest measured interrupt to process latency went from 14000 to 391. My system had over 6h of system interrupt in the last month.

Final fix: Disable USB2 port2 in the bios and the device is gone completely.

Imagine blaming the Intel "IO driver" for high latency, only to discover your mobo manufacturer has left a USB port hanging into the void.

 

[igor_kavinski]

So mobo crapware installation needed to fix latency????

How did things get so BAD???? (I really want to spam my post with more question marks)

Having to use USBVIEW to investigate is insane. Does every mobo owner need to scour forum posts to figure out how others owning the same mobo fixed their issues? It should JUST WORK out of the box. PERIOD.

And thank you @coercitiv for sharing.

 

[JoeRambo]

On topic of DPC latency, Nvidia finally released the drivers fixing DPC latency bug with Ampere+. Notes say Ampere, but my 4090 was also hitting it and now mostly seems to be fixed.

Fixed General Bugs

• Increase in DPC latency observed in Latencymon for Ampere-based GPUs [3952556]

Looking fine now, used to be >300 for DPC in Nvidia driver.

EDIT: looking at their video, they might have been hit by that bug, those 302 microseconds and nvld... is givaway

 

[Timur Born]

Unfortunately these videos are sloppy. They describe a problem, demonstrate it with hard to reproduce situations (system specific tasks like Premiere) and offer no real analysis (CPU load, SSE load, memory, etc). Like so often DPC "latencies" are just thrown in without proper understanding (or at least explanation) of their meaning.

The files examples seem to be done from a Desktop folder, but we don't know if the Desktop isn't redirected to a network drive or synchronized via Dropbox or Onedrive (both of which try to push users to do so). That "file icons are rebuilt slowly in explorer" situation usually happens on network drives and then you also get the Ethernet controller in the mix with the Intel I255-V being very (!) problematic (even revision 3).

The 2nd Youtuber at least offered that "Word" test that is more easily reproducible. Alas, nothing out of the ordinary happens on my 13900K.

If this is PCIe lanes related then I would have to ask if their M.2 drive is maybe installed in a slot that is connected to the chipset instead of the CPU? If it is CPU related then I would like to know which C-states are enabled in their respective BIOS and whether Windows power profiles use "autonomous" mode (Speedshift) and/or core parking (Windows defaults differ per CPU).

Or testing with E-cores disabled.

From what I understood the second Youtuber did that. At least he mentioned that he thought this to be E cores related, but it seemingly wasn't.

 

[BFG10K]

Followup testing on Ryzen.

 

[sandmann771]

I am a software developer in the hardware-related area. I am amazed at the demands placed on attentive observers of the problem. I have tested more than 20 PC systems. Of these, 5 were 13th generation i7 systems. Only these 5 systems produced real-time peaks. As a user, I can say that the new Intel systems are unusable for audio DAWs. I think Intel and Microsoft have known this for a long time. It is enough to install a PC with a simple, reduced Windows without virus protection and Internet access and use the optimizations for audio applications. If the 10th generation i7 runs without peaks within 10 minutes and the new 12th and 13th generations don't, then that's the case and there's a problem. The technology specialists can say whatever they want. Then there is a problem and these systems have no use for real-time applications and are therefore a step backwards for me as a user.