Lot of good info in this guide, but also some bad advice, imo. Disabling c-states/speedshift/speedstep/hyperthreading with nvidia at max performance is just a waste, while skyrocketing temps and power draw. With all those settings disabled and a power plan with processor idle state set to disabled, my 65w cpu/110w gpu (12400f/4060) runs at 110w just sitting at the desktop with no apps open and nothing running in the background, just nvidia in systray. With this setup and a tweaked OS i get about 2-10µs average on max performance plan. You also lose your boost clocks, as all recommend settings lock your cpu at lower clocks. For instance, my CPU is rated for 4.4ghz boost, by disabling speedshift/speedstep, I get locked to 3.98ghz under any type of load. I was also hitting around 85c on CPU during gaming.
If I leave c-states enabled with speedstep/speedshift/hyperthreading enabled and nvidia at normal power, my system draws 33w idle at desktop with nothing open (yes, it's very low, it's heavily tweaked bios and system). I idle at about 38w-40w on a static text webpage. On the same setup, just switching to Higher Performance power plan, I get about 20-30µs average. My boost also doesn't get robbed, as I hit 4.4ghz with this setup. Yes, balanced power plan has DPC spikes, but they're under 200µs and way beyond suitable for the tasks I do of just browsing/listening to music/discord. I mean, if you need world record breaking latency to browse YouTube or edit a video, go for it.
The difference in game? Can't tell or feel a difference at 4k polling/240hz/240fps locked (I play older games). The only real difference I can feel with disabling everything like almost every guide always recommends, is that the windows desktop experience is a tiny bit snappier, but you can achieve 95% of the same feel on balanced profile, keep reading. The difference is so tiny, the tradeoff to achieve it is laughable compared to performance I get on my balanced power profile. Just setup process lasso or similar app to switch to your Ultimate Power Plan (tweak it, it's not really ultimate by default
) when your game exe is launched. That way you don't have to bother constantly manually switching power profiles and you get the super lower dpc interupt/execution only when your game is open and when on desktop you're sippin' power instead of full blasting your CPU to browse YouTube.
If you debloat windows (ntlite), tweak/strip bios, setup proper core assignment properly, and find perfect drivers through trial and error testing, that's all your really need. That's all I did and I'm sitting at 10-30µs when I need it and saving power when I don't. Power is cheap here, only 10 cents/kWh, but I love tuning for power efficiency as much as I do for latency, mainly why I prefer intel 65w cpus. All measurements done with a killawatt. With all these tweaks you see recommend all over the web telling you to disable c-states/speedstep/speedshift/hyperthreading, it's just impractical advice repeated over and over. You should only be following this advice if you're chasing the lowest latencymon world record for a prize of a million dollars, otherwise don't waste your time and work on tweaking other areas. I can't imagine how much more worthless this is doing on more modern cpus that have faster and smarter power saving wake/sleep core behavior than my old 12th gen 12400f.
My workflow for getting the lowest latency without 95% of the BS tweaks. Start with NTLITE and learn how to strip the OS yourself, don't bother with pre-made gaming iso's like this guide recommends, unless you don't have time or interest to do it yourself. If you're overwhelmed with NTLITE, look for a pre-made XML preset that strips for you, you just load it in NTLITE and it does all the work for you. The XML are also far easier to read/decompile than the gaming ISO floating around. You just open it in notepad and it's in plain english, that's it, no complicated scripts to understand. If you see something stripped that you need, just remove it from the XML and check for dependencies. Want something stripped that's not in the XML? Just add it. Or just use the GUI interface, as you can edit XML just double clicking enable/disable. This is also much safer than trusting some random pre-made gaming ISO for security reasons. Gaming ISO also add a headache to troubleshooting DPC spikes, because you have no idea what tweaks are applied across the system, unless you spend hours reading the source code, if it's even available.
With NTLITE, you start with fresh slate and a stripped OS with no tweaks applied, just a de-bloated os. Measure DPC as soon as your ISO install is done, get a baseline. Add drivers 1 by 1, measuring after each driver. If you notice large spikes after a driver install, you have to find an older version or dig into the latencymon info to try and pinpoint it exactly. Also, remove all devices from USB ports, only Keyboard/Mouse while you setup and move along your baseline. A simple thumb drive plugged can have the possibility to drive you crazy with random spikes you'll never even realize is the cause.
The most important thing for DPC isn't a 1-2µs reading, it's a stable reading. On my system, my spikes are anywhere from 10-30µs, that is very good. I think even at a 100µs baseline you wouldn't feel a difference and have the same amazing responsiveness as a 10µs system , as long as it's stable and spikes remain tiny. Keep in mind, these µs readings are microsecond measurements. 1µs = 0.001ms vs 100µs = 0.1ms. If you can feel a 0.1ms difference between interrupts and dpc execution, you must be super human. A stable 100µs system with stable and small spikes will feel better than a 10µs system with large spikes.
Focus on stabilizing your spikes, not chasing the lowest possible DPC number.
These super low idle numbers are useless, besides using it as a baseline to fix problems. Why a 1µs vs 50µs difference makes no logical difference - when you game, the game itself and nvidia are going to raise your execution/interrupt times. You already wouldn't be able to feel a difference between 1µs and 50µs, add in the noise of a gaming load which naturally raises these numbers, you make the indistinguishable difference between 1µs and 50µs even more indistinguishable. It's a waste of time and power draw. Again, focus on stabilizing, not number chasing. And if you want the snappy desktop experience/ultra smooth animations speedstep/speedshift disabled brings, just tweak your cpu power settings. And if you want to go the extreme, just force c1/c2 cstates, as they have the fastest wake times.
