YT YouTube2Text

YouTube2Text → Transcripts → I Spent a Month Optimizing my PC so You Don't Have To

I Spent a Month Optimizing my PC so You Don't Have To — Transcript

by bardOZ · 11,928 words · 1,681 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Ch YouTube. You might be asking
0:01yourself, who is this Italian guy? Why
0:04is he sitting on the ground? What's
0:05going on with these cables? And why is
0:07there a cat? Anyway, welcome to my video
0:10about how I spent many hours sitting on
0:12the ground trying to make my computer
0:14run as flawlessly as I possibly could
0:17because I really care about performance
0:19and about performing really well when I
0:21play my video games. I'm a competitive
0:22gamer, expert player, and world record
0:24holder name trainers, and I realized
0:26that it wasn't quite working the way I
0:28wanted it to work. I've recorded this
0:30video before because really the focus
0:32was about how I set up two computers to
0:35work together so that I could have 240
0:38fps VODs with absolutely no extra
0:41latency for my gaming computer. And a
0:42lot of that was this mess of cables that
0:45fix my audio problems. And I realized
0:47most people won't care about that. Most
0:49people that are seeing what I'm posting
0:51care about the FPS. So, I'm re-recording
0:53the video with just a focus on
0:55everything I've learned about optimizing
0:56actual FPS and performance and latency,
1:00and it will all be in this video. Now, I
1:03want to give you a little bit of a
1:05disclaimer. I'm never going to be
1:06selling you a course or personalized
1:09optimization stuff. I don't like that.
1:11So, a lot of people reaching out asking
1:13me, they want to pay me to tweak their
1:15computer and stuff. I don't like that. I
1:18really think that a lot of you guys are
1:20paying people that are honestly not very
1:22competent and they're selling you snake
1:23oil. I'm sorry to break that to you. And
1:26I think this is kind of like easy enough
1:28that you should be able to learn it
1:30yourself. Do it yourself or pay someone,
1:33but a lot less than I think most of
1:34these people are asking you for doing
1:36magical voodoo stuff. So, I'm going to
1:38try to keep this video simple. It's
1:40going to be tiers. The tier zero is
1:44making sure that your actual physical
1:47setup is correct cuz that's super
1:50important especially for thermals. Then
1:52the next tier will be overclocking and
1:54actual hardware tweaking which is where
1:5790% of your gains will be after the tier
1:59zero. Tier two is going to be software.
2:02Now, part of the software will be
2:04something that I've personally
2:05benchmarked and then at the very end I
2:08will give you the placebo part which is
2:10where I will mention some of these
2:12things that maybe you will find on
2:15Twitter from these people that are doing
2:16the LinkedIn style like clickbay posts.
2:20I will tell you which ones definitely
2:21don't work, which ones might work, and
2:24therefore I might be using myself and
2:27we'll try testing later on, and which
2:29ones just are not [music]
2:31something that you should do because
2:33they're actively bad. So, let's get
2:35started with the physical. So, tier one
2:39is the first step in your checklist. You
2:41should make sure that your physical
2:42components are okay. Obviously, you
2:44can't just upgrade your computer without
2:46spending money, but there's still very
2:48common mistakes that are made that can
2:50be fixed by physically going inside your
2:53case and changing things. For reference,
2:55my CPU is a 7800 X3D, my GPU is a 1490,
2:59and I've got two sticks of 32 GB Highix
3:02ADI RAM rated at 6,000 CL30 in their
3:05expo. So, here's a checklist that you
3:07should go through to make sure that
3:09before we move any further to
3:10overclocking or tweaks, your hardware is
3:12performing as expected. Number one, make
3:15sure that your RAM is in the correct
3:16slots. This is a very common mistake.
3:18You need to make sure that if you're
3:19using two sticks to properly use dual
3:21channel, you need to use the correct
3:24slots. This depends on your motherboard.
3:26So, look it up in your motherboard
3:27manual. For me, it's number two and
3:28number four. Another thing is make sure
3:30your GPU is being fed power properly.
3:32Especially if you have a GPU that has a
3:3512 pin connector and you're using one of
3:38the Nvidia adapters that give you 12 pin
3:41to 48 pin. If you only use three, it's
3:44going to work, but it's going to be
3:45power limited. So, keep that in mind.
3:47Number three, make sure that your CPU is
3:49cooled properly. Any modern CPU
3:52determines its clock speed as a function
3:55of load and temperature and voltage. But
3:58temperature is really important because
4:00it's not just about achieving the
4:01thermal throttle or you know the point
4:03at which your CPU is going to shut off.
4:05That's like extreme overheating. But in
4:08general, if you stay higher, both modern
4:11GPUs and CPUs will take that into
4:13account when they decide how to boost.
4:15So that means that even if you only
4:17average maybe 15 20 C higher when you're
4:20gaming, even if you're not getting
4:22officially thermally throttled, you will
4:24still not boost as high. So, make sure
4:26that you're cooling everything properly.
4:28This also means replacing your thermal
4:29paste or upgrading it. For example, I
4:31moved to using a cryo sheet, which is a
4:34sheet of graffine, which is a conductive
4:36material. So, you have to be careful if
4:37you do something like this. That goes
4:39onto the CPU and acts as the thermal
4:42connection between your heat sink and
4:44the CPU. Based on my testing, it
4:47performs better than any of the thermal
4:49compounds that I have access to right
4:51now. Another thing is if you're using an
4:52AIO, look up the best ways to mount it.
4:55Both in ways that won't kill the pump,
4:57which means that the pump can't be the
4:59highest point in the loop, which is
5:01pretty wellnown. But everyone will tell
5:03you to mount it on the top of the case
5:06in exhaust because that's safe for the
5:08pump. But I had very noticeable
5:10improvements in temperature for the CPU
5:13by mounting the AIO in the front and
5:16sandwiching it between two sets of fans.
5:19I'll talk about it later when I touch on
5:20my case. Number three, same thing for
5:23the GPU. Make sure that it's not being
5:26thermally throttled. And GPUs are even
5:28more sensitive. Nvidia has an algorithm
5:30and it's pretty clear to observe that as
5:33temperature goes up, your maximum boost
5:35goes down proportionally. So, the lower
5:38you can keep your GPU in terms of
5:40temperature while you're gaming, the
5:41more FPS you will get. So clearly you
5:44can do that much with GPUs compared to a
5:46CPU, but you can repaste it. I did it
5:49and I put PTM 7950, which is a phase
5:51shifting compound that's super popular
5:53now because it's cheap and it works
5:54great. And that had a massive effect on
5:56my thermals. So repasting a GPU is
5:59definitely more of a technical thing
6:00than repasting a CPU. So be careful if
6:03you decide to do it. Watch some videos
6:05before you attempt it. And of course,
6:07there's some risk that you're going to
6:08mess it up. So be careful. But I have to
6:10mention it because the gains are
6:12meaningful. Now number five, and this is
6:14often ignored, but the airflow in your
6:16case will help both your GPU and CPU and
6:20the other components like your RAM, your
6:21VRM stay in the right thermal range to
6:24perform. And you kind of need to
6:26experiment and figure out for your case
6:28what works best. I tested a lot of
6:30configurations. And for me, the key
6:32components were the AIO radiator mounted
6:35in the front, as I mentioned earlier,
6:36with cold air from outside going through
6:39three sets of case fans being pushed to
6:41the radiator and then into the case that
6:43gives plenty of cooling to the CPU and
6:46also a lot of positive pressure. So, the
6:48GPU has almost ambient temperature air
6:52to cool itself. Then the exhaust for me
6:54that work great were at the top and at
6:56the back. And ultimately, I tested with
6:58and without the glass panel on, and the
7:01GPU temperature is way lower when I
7:04don't use my glass panel. So, for me, no
7:06glass panel is better. But, of course,
7:08you're going to have more dust if you do
7:10that. Now, you should really take the
7:11time to test this out for yourself in
7:13your case, in your CPU, your GPU,
7:16because it's pointless for you to be
7:17overclocking or tweaking things if
7:19you're leaving most of the gains in FPS
7:22on the table by not optimizing your
7:23thermals. And I'm talking serious gains.
7:26Just changing this improve my overall
7:28performance by twodigit percentage
7:30points. But now assuming that you fixed
7:32your physical. Let's move on to
7:34overclocking and optimizing your
7:36hardware. Tier two. Okay guys, tier two
7:39hardware optimization. We start from the
7:41bias. We're going to be optimizing three
7:43pieces of hardware, the most important
7:45ones in this case. CPU, RAM, and GPU.
7:48We're not really going to be doing the
7:49GPU here, but we'll be focusing on CPU
7:51and RAM because the GPU is usually done
7:53in the operating system. Now, before we
7:55start, please update your BIOS. Even
7:58recently, I was helping a friend
7:59optimize his computer. I was just trying
8:01to see if I could achieve the same
8:02results that I achieve on my machine on
8:04his, which we did. But we realized his
8:06BIOS was really old. And we tried first
8:09specifically as an experiment not to
8:11update it to see if it would change
8:12something. And first of all, some of the
8:14good options that I need to tune wener
8:16even there because they were added later
8:18on in the new AGSA versions of the BIOS
8:21is also AMD. But even just the timings
8:24for the RAM, for example, weren't as
8:26stable on older versions of the BIOS.
8:28So, please update your BIOS. I know it
8:30almost sounds like your parents telling
8:31you to make sure that you wear a sweater
8:33when it's cold outside, but it's not.
8:35It's not like, "Mom, I'm not going to
8:37get sick." No, you will get more
8:39instability if you don't update your
8:40BIOS based on my testing. Unless you you
8:43have a faulty BIOS version that's the
8:46latest, but I doubt that's the case. So,
8:48please update your BIOS before you do
8:50any of this because then you might
8:52realize that oh, for example, maybe even
8:54instabilities that you were getting with
8:56your base system are not there anymore.
9:00Now, usually when you open your BIOS,
9:02depending on your motherboard, you will
9:03see something like this, which is a
9:04simplified version of your actual BIOS.
9:07I suggest to go to the advanced mode of
9:10your BIOS, so that you can get more
9:12settings. This is mine, for example, on
9:14the ASUS machine. In my case, I have AMD
9:17and my RAMs have expo profiles saved
9:21inside them. Now, what is Expo? It's
9:23basically a preset profile that tunes
9:26voltages, frequency, and timings
9:28automatically uh for values that are
9:30known to work well with that given
9:32stick. If you have Intel, it's called
9:34XMPP. You can use sticks that only have
9:38XMPP profiles on AMD and vice versa. But
9:41that means that you probably won't be
9:42able to use any of the automatic tuning,
9:45but you will have to do the timings
9:47yourself for which I highly recommend
9:49starting from a video from build zoid
9:51called build zoid easy timings for die.
9:55Heinix is a type of stick that you can
9:57have which is probably what you have if
9:59you're watching this video. If not, you
10:01will have to look up good timings
10:02yourself or or try yourself. I will say
10:05this, I know most people just go like
10:07enable XP and forget about it, but I got
10:10really big gains from tuning my RAM.
10:12We're talking like 10% more FPS compared
10:15to Expo, [music] at least depending on
10:17the scenario. Um, especially when you're
10:19CPU bound like most competitive games
10:21are. So, I know it's a pain in the ass.
10:23I know most people say it doesn't
10:24matter. And maybe for most people, it
10:27really doesn't because you don't care
10:28about the 10% extra FPS, but I am an
10:30extremely competitive gamer. I wanted
10:32the edge. So, I went through the effort.
10:35This is not going to be a guide on how
10:36to tune your RAM. I'm also not an
10:37expert. I highly recommend you watch
10:39someone like Bill Zoid or a thousand
10:41people that are more competent than me
10:42when it comes to actually tuning your
10:44RAM. I did it on my own. I'll show you
10:45the numbers that I managed to get, but I
10:47highly recommend that you go through the
10:49effort of doing it. It's also really
10:51hard to pay someone to do it for
10:52yourself because unless you have a setup
10:54like this where you can stream your
10:55BIOS, it's just very inconvenient to
10:57have someone doing it for you because
10:59it's a really long back and forth
11:01process. So I recommend you you learn
11:03how to do it. Um but start from your
11:05expo and make sure it works and then
11:07move to some someone else's timings like
11:09build zoid and then move from there.
11:11Memory frequency part of your of your
11:13RAM like it's the big number that
11:15determines the mega transfers per
11:16second. FK frequency is something AMD
11:18related that means infinity fabric
11:21clock. I'm not going to explain all the
11:23details here because I already am
11:24re-recording this for the second time
11:26because I was going too long on the
11:27first one. But TLDDR this number depends
11:30in a way with the one you want depends
11:33in a way from your memory frequency
11:35because there is such a value that it is
11:37synchronized with it depending on your
11:40memory controller on your CPU and uh
11:43another value called UCLK D1 mode which
11:45we'll talk about in a little bit but
11:47basically there's a value of this that
11:49minimizes latency and then as you
11:51increase you're going to increase your
11:53latency but also increase your bandwidth
11:55and to overcome the fact that they're
11:58not synchronized anymore. You need to
11:59get to a certain value that is now
12:01better for both latency and bandwidth.
12:04So just increasing this might not give
12:05you the best results unless you increase
12:07it by enough to overcome the loss in
12:10latency by not being synchronized
12:11anymore. Now I'm not talking input lag.
12:14I'm talking memory latency which is a
12:17value measuring nanconds. Remember
12:19there's a million nconds in a
12:21millisecond. Okay? So this is not like
12:23oh you lost three nonds. No human will
12:26be able to perceive three nanoseconds.
12:29Okay, we're probably able to determine
12:31stuff in the milliseconds. I'm now
12:33pretty confident that we are, but not
12:34nanconds. I'm sorry, guys. So, when I
12:37say latency, that's a value that implies
12:39a lot of other things. And lower RAM
12:41latency will improve your overall
12:43performance, but we're not talking
12:45milliseconds here. But anyway, in my
12:47specific case, I am lucky because I can
12:49go all the way up to 12 to 2200 at 6,000
12:52MHz, which is really good. I actually
12:54get a little higher than 2200 in real in
12:57the real world when I run my PC because
12:59my BCLK is slightly higher than 100.0.
13:02But please, this value BCLK, don't touch
13:04it unless you know what you're doing
13:06because you can absolutely corrupt all
13:07your data in your NVME, which I did, and
13:09I had to reinstall my Windows, which is
13:11a pain in the ass. And you can also not
13:14manage to boot because your GPU, which
13:17runs on the same PCIe bus that shares
13:19this base clock, might not work. You
13:21might have USB problems. Just be very
13:24careful if you if you tweak your BCLK.
13:26Now, in my BIOS, you can see we have DAM
13:28timing control, which is where you
13:29actually go tweak your RAM timings. It's
13:32a very important part of the of the
13:34tuning. The only thing I would recommend
13:36is if you get really slow boot times,
13:39you have to turn on memory context
13:40restore, which will automatically turn
13:42on power down enabled. In a lot of
13:44BIOSes, you need them both to be enabled
13:46or you will get black screens if you try
13:48enabled and disabled. Some people will
13:50tell you that you don't want this
13:51because it will make your memory less
13:53stable because it's not going to be
13:55training your RAM every time it boots,
13:57but it's going to reuse known uh values
14:00and maybe there's like a slightly
14:02different voltage from your power on
14:05your wall and it needs to retrain. I
14:07personally also this power down enabled
14:10used to give uh maybe still does
14:13slightly gives a little higher latency
14:16reading in ADA 64 which is a software
14:18used to test latency and therefore
14:20people were like don't turn power down
14:22enable on because you get much lower
14:24latency because your RAM has to wake up
14:26because it allows part of your RAM to go
14:28to sleep. In the real world this isn't
14:31meaningful, okay? Because your your game
14:33isn't constantly going to use different
14:35parts of your RAM. So it's not going to
14:37incur in this extra latency anyway. Yes,
14:40it might be measurable by something like
14:42ADA because it's going to be trying to
14:43randomly sample for latency, but it's
14:46not meaningful. And I would much rather
14:47take like 10-second boot times than 1
14:50minute every time. So I recommend using
14:52memory context store and power down
14:53enable, especially when you're going to
14:55be tuning stuff like your CPU and GPU.
14:58It makes it much faster to cycle through
15:01different tests. So because you don't
15:03have to wait a minute. So I personally
15:05use MCR MPD on UCLK V1 mode. You have to
15:10make sure that it's Ucl equals MCLK
15:12which is the memory controller on the
15:14CPU is synchronized with the frequency
15:16on your uh RAM on on the memory
15:18controller for your RAM. But if you're
15:20using very very high frequency on your
15:22RAM 7600 and higher apparently it's
15:25going to be realistically impossible for
15:26your ULK on your CPU to run that fast.
15:30So you need to use this UCLK equals MCLK
15:33divided by 2. your memory controller on
15:35your CPU is going to be running slower.
15:36Now, if you're going to be using 7600 or
15:398,000 MHz on your RAM, you probably
15:42already know what this is about. So, you
15:44probably already have it set to MELK
15:46divided by two. But just be careful
15:49because I know some people have like
15:50RAMs whose expo is 7600, then they set
15:54the expo, then they get boot problems or
15:57instability, then they lower their
15:59frequency uh by manually. The expo
16:03probably said Uclals MCLK divided by 2
16:06and then they end up with the UCL the
16:09memory controller on the CPU going way
16:11slower than it should uh add low memory
16:14frequency and that loses like makes you
16:16lose a lot a lot of performance. So if
16:18you're in the 6,000 to 6400
16:22MHz frequency based on my testing
16:24obviously you want your CLK equals MLK
16:27gear down mode. The old way of doing
16:29this was to find the values at which you
16:32would be disabled because you get uh
16:34lower latency uh even though you give up
16:37stability. Personally, based on all my
16:39testing, I would much rather have
16:41enabled for the extra stability even
16:43though maybe there's like a little extra
16:45latency, but uh it allows me to go
16:47deeper in all my other timings.
16:50Now, again, I said you have to do the
16:52timing tweaking yourself. So, I'll show
16:53you my timings are stable for my sticks.
16:56my 32x2 Heinix ADI dual rank sticks and
16:59I'm sure there's a lot of juice left to
17:02squeeze here. I'm not an expert RAM
17:04tuner and it's the most complicated
17:05part, but I'm pretty happy with my
17:07current results and they're very, very
17:08stable. So, that's what I'm keeping for
17:11now. I'm having fun tuning, so I might
17:13go deeper. I'll just show them to you.
17:15You can just pause the video if you want
17:17to try them yourself on your RAM if you
17:19have similar sticks. But
17:23this is it. I'm not going to talk about
17:24each individual timing because honestly
17:27even the experienced overclockers are
17:28finally admitting then that you can sure
17:31try to use logic behind these. To me the
17:34reality is that you need to like just
17:36keep trying to find which ones work best
17:39because you can really just logic your
17:41way out of these timings. If you don't
17:43know what any of these are just use expo
17:45or use the video that I'm going to link
17:47in the description from Bill Zoid. Now
17:49the next very important thing for tuning
17:50your CPU is precision boost overdrive.
17:53There's an AMD feature that allows your
17:55CPU to boost its clock, which is the
17:58most important part of the CPU depending
18:00on load. So, how much load is happening
18:03on the CPU, temperature, and voltage.
18:06You can if you don't turn it on, your
18:09CPU just won't give you all the the the
18:12juice that it can give you, unless
18:14you're doing a different type of
18:15overclock where you just raise the clock
18:17all the time. But for my 7800 XTD, you
18:20really can do it. It's completely
18:21locked. There's no multiplier that you
18:23can change. You can't really tweak the
18:25voltages manually. So, the only way to
18:27get maximum performance out of a 7800 XD
18:30is PBO. Some people will be thinking,
18:33well, there's another way. There's ELK,
18:35the asynchronous base clock uh tweaking
18:38on some motherboards. Yes, a few of you
18:40that bought those motherboards that have
18:42it because you are experienced
18:43overclockers will have a way to get even
18:46more performance. Most of us normal
18:48humans will have a normal motherboard
18:50and therefore our only way is PBLO. Now
18:53I've tested medium load boost. It
18:55actually uh gives me a little more
18:57performance. So I use it enabled
18:58precision boost overdrive enabled
19:00because if you actually read for me uh
19:03let me move my face well. I guess this
19:06works for me. It's basically saying that
19:08it allows the processor to run beyond
19:10the defined values as in the limits
19:13defined by default by AMD to the limits
19:16of the board. You can't go higher than
19:18the limits of the board. So what it's
19:20basically saying is that by using
19:23enabled it raises the limits to as high
19:26as they can go. Enhancement manual AMD
19:29eco mode even though enhancement sounds
19:30like it should be better than enabled
19:32actually aren't. Enhancement basically
19:34like changes the limits for yourself
19:36based on some temp temperature goal.
19:39Enable just raises them to as high as
19:41they can go. So use it to enabled.
19:42Overdrive scaler. Real quick, this value
19:45here is used internally by the PBO
19:47algorithm. It defines how long it can
19:49stay at the boost voltage and it will
19:52make your CPU degrade faster the higher
19:55the number goes. A lot of people say to
19:56stay at 2x. I haven't really measured
19:58the big of a difference and I'm okay
19:59with my CPU degrading faster. I don't
20:01really trust that the risk is that
20:03large, so I keep it at 10x. But
20:05experiment for yourself. This value
20:07here, on some CPUs, it will actually
20:09help it boost higher at the maximum
20:12level of boost from the PBO. On the
20:147,800 XPD, it can't go any higher than
20:1650/50. So, it does absolutely nothing.
20:19Same thing here. Doesn't I can't really
20:22tweak how high it's going to boost. Now
20:24for the curve optimizer which is the
20:26actual important part of this. This is
20:28what will allow your AMD CPU to boost
20:32higher and maintain its boosts higher.
20:35Now the curve it's optimizing curve
20:37optimizer is the voltage frequency
20:41curve. It's saying at a given voltage
20:44what frequency can the CPU be at? And
20:48you can shift with an offset so that say
20:51that by default at one volt your CPU is
20:54at four. You can shift it so that it is
20:58at four at like 0.9. Therefore, if it
21:02gets to one, it's going to be higher
21:04than four. So it's a way to overclock by
21:08undervolting because by undervolting
21:10it's going to be colder and then the the
21:12overgorithm will allow it to be at a
21:14higher voltage at the same temperature
21:16and therefore that or even the same
21:19voltage now but that voltage will have a
21:22higher clock connected to it. So you can
21:25do it per core or all cores. As I
21:27mentioned, maybe you can squeeze more by
21:29doing it per core. And you have to use a
21:31negative value if you want that a lower
21:34voltage will correlate to a higher
21:36frequency which is what we want. So it's
21:38technically an undervolt uh even though
21:41it's much more complicated than that.
21:43[music] And the value is just a number
21:46which is used internally to determine
21:48how many millolts per frequency it will
21:51shift. I'm pretty sure this number is is
21:52a little more complic in a little more
21:54complicated way than just like
21:56undervolting inside the algorithm for
21:58PBO, but whatever. Now, how low you can
22:01go depends on your CPU itself, your
22:05probably your motherboard, the quality
22:06of the of the power from your supply,
22:10the RAM. It's just complicated. You have
22:12to try and see what value like how low
22:14you can go. They will still allow you to
22:16boost pretty high. Now I can honestly I
22:20get pretty much the same performance for
22:22any number lower than 20 because of a
22:25different setting elsewhere that I will
22:26show you. So my CPU is still able to
22:29boost 50/50 which is the highest it can
22:32go 5.05 GHz on each core that is running
22:36my game which is what matters with no
22:38clock stretching which means that it's
22:40actually at that clock
22:43and effectively. Now why am I at minus
22:4627? because I'm kind of seeing how far I
22:48can go with it still being stable.
22:50Realistically, I think minus 23 is where
22:53I've noticed I have pull stability even
22:55like in synthetic benchmarks and testing
22:58and my CPU still has like imperceptible
23:01differences in games. So, I have 27 now,
23:03but I went as low as 32, but that
23:06absolutely does give me some instability
23:08or even 35. Same thing, some instability
23:10in some cases. This one you have to try
23:12yourself. I would start probably from 15
23:14or 20 negative all cores which will
23:17allow you to boost much higher and you
23:19will notice differences and then go from
23:20there. Go lower and see when you start
23:22getting instabilities.
23:24If you have a newer 9xxx uh chip, you
23:28can also use this really cool thing
23:29called curve shaper, which basically
23:32allows you to define different values of
23:35the curve optimizer per different
23:37situations, which is really nice because
23:39any CPU gets less stable at lower
23:41voltages and the boost depends on load.
23:44So an issue with P with a basic curve
23:46optimizer is that when you don't have
23:48load, the boost will be really low. So
23:50in the voltage frequency curve, you will
23:52be really far to the left. And what that
23:54causes is that you will go so low in
23:56voltage even though you would be okay
23:58when you're not playing being at a
23:59higher voltage that it will crash when
24:02idle but not when under load which is
24:04kind of annoying. So the curve shaper
24:06seems to help with that but it doesn't
24:07work on the 7800 XD on my motherboard at
24:10least. So sadly I can't test it but it's
24:12uh something that makes perfect sense to
24:14me. Now these settings here I use level
24:18two enabled disabled. Why? Core tuning
24:21configuration for gaming according to
24:22AMD is changing some settings of the
24:25prefetcher so that it's better for
24:27gaming. Even though some people will
24:29tell you to use legacy because they
24:30measure a lower latency in ADA 64, that
24:33to me is just an artifact of the
24:35preetching. But if this gives better
24:37FPS, that's what matters, not just like
24:39nanoseconds difference in a synthetic
24:41benchmark, which doesn't mean that it's
24:43actually real in the in the actual usage
24:44of the of the CPU. So I suggest using
24:48level two core performance boost you
24:49need enabled or it's not actually going
24:51to let the PBO algorithm work. Turbo
24:54game mode disabled because enabling
24:56basically makes your CPU uh an 8 core
25:008thread CPU no matter what it is which I
25:03think is just pointless unless you are
25:06so lazy that you don't want to do
25:07affinities in the operating system. But
25:11you can achieve this with just changing
25:13the affinities in the operating system
25:15without actually killing the rest of the
25:17cores which might be used for background
25:18stuff. So I don't recommend turning
25:20turbo game mode on even though they
25:22named it terribly. DG+ VRM. This one is
25:25kind of risky. I'll go really quick. If
25:28you want to change stuff here like I
25:30did, know that you might risk prying
25:32your CPU if there's a transient very
25:35high voltage and just degrade it faster.
25:38I'm okay with the risk. I'm not going to
25:40explain these here at the bottom. The
25:42TLDDR is that they're going to
25:44prioritize the stability of the voltage
25:46over thermal and allow the CPU to pull
25:49more current. This one, honestly, I'm
25:50pretty sure does absolutely nothing on a
25:527900 XCD because it's going to get
25:54limited way before it needs to pull more
25:57current. But this one does do something.
25:59So, I'm going to talk about this really
26:01quick because a lot of people have a lot
26:02of opinions about CPU line calibration.
26:04What this does is basically when your
26:07CPU is suddenly loaded, the voltage
26:10drops. It's called V droop. And CPU load
26:13line calibration is an algorithm that
26:15tries to flatten this curve so that when
26:18load is applied, it increases the
26:19voltage accordingly so that it can keep
26:21it more and more uh flat. This is risky
26:26because what happens is that by the time
26:29the voltage starts dropping and the CPU
26:31adds it, this there's a little bit of
26:33latency. So what can happen is also that
26:35a much higher voltage than what is
26:37reported is for a brief period of time
26:40fed to your CPU which could kill it.
26:43Some people think that it's related to
26:45some famous like 7800 XD like AM5 deaths
26:48on some ASRock motherboards. Personally
26:52I'm okay with the risk because it does
26:54increase performance by a lot and I
26:56think I figure out why. First of all uh
26:59the actual value depends on your
27:01motherboard. you have to look up which
27:02value is the flattest B droop. So the
27:05one where the motherboard is feeding
27:08more voltage and therefore more risk of
27:10frying your CPU. But what what happens
27:12is that remember how earlier with PBO we
27:15shifted the curve so that at a given
27:17voltage it can boost higher that is a
27:20value that's used internally to the CPU
27:22as we've mentioned and it will and I so
27:25some people are like you are giving
27:28voltage back from with the lowline
27:29calibration that you're taking away with
27:30PBO so that makes no sense but no
27:33because I tried and even at the same
27:35value of PBO as you increase load line
27:37calibration and therefore the voltage
27:39increases the actual boost increases
27:42too. So it seems like PBO shifts the
27:45curve and then it's almost like load
27:48line calibration is yes giving the
27:51voltage back but still at the shifted
27:53curve. So it's allowing the higher I go
27:57the more flat I go the higher it allows
27:59my CPU to boost even at a fixed PBO. So
28:03eight allows me to feed more voltage and
28:06keep the CPU more boost like get it to
28:09boost higher uh even at lower PBO. For
28:11me that's okay because my thermals are
28:13are great and my CPU never really
28:16overheats. But of course an effect of
28:17this apart from the risk of attrction
28:19voltage by killing your CPU is that it
28:22will effectively allow your CPU to boost
28:25higher at a higher voltage or or
28:27actually to it will allow your CPU to
28:30stay at a higher voltage and therefore
28:32boost higher so it will increase
28:34thermals. So you need to be careful
28:35because you're undervolting on one side
28:37and then giving that voltage back and it
28:39will boost high and it will generate
28:41more heat and of course it's pointless
28:42if you end up hitting a thermal
28:44throttle. So you need to make sure that
28:46you take all these things in
28:47consideration. But this one, no one
28:49really talks about it or they tell you
28:50not to use the maximum. But this one is
28:52massive in terms of actual performance
28:54gains. So I'm using level eight because
28:56again I'm okay with the risk and my CPU
28:58is really well cooled. This one is just
29:00if you want to be doing benchmarks and
29:02it changes some of the internal settings
29:03to do to get more numbers in benchmarks.
29:06I never changed this little thing here.
29:09Apparently, if you disable clock spread
29:10spectrum, which is a feature that allows
29:12your CPU not to stay at exactly its base
29:14clock, but go slightly up and down all
29:16the time to minimize electromagnetic
29:18interference. Apparently, this helps
29:19with stability, which makes logical
29:21sense to me. I haven't tested it a lot,
29:23but it seems like it's true. So, I have
29:25it disabled. This one, I don't really
29:27know. I'm testing because I don't no one
29:29really this isn't really documented, so
29:31I'm trying to figure out what it does.
29:33So, I've been testing it. That's why
29:34it's enabled. But, I don't recommend
29:35actually changing this one. For flex, I
29:37would leave it alone. It basically
29:38changes the algorithm for how the
29:40different CC the different CCDs which
29:42are the sets of cores in multiCCD CPUs
29:45or that might have some cores with the
29:47X3D which is the actual level three
29:49cache or not are prioritized differently
29:52according to values I will leave it
29:53alone cuz you can honestly probably do a
29:56lot of that work manually in process
29:58lasso later in the tier 2 level that
30:00we're going to see later the software
30:01level now the voltages I'm not going to
30:03really spend much time here I'll tell
30:05you SOC is really important because the
30:07higher you go, the higher you can have
30:10your ULK be stable, which means you can
30:13go higher memory frequency. But the
30:15higher you go, the less stable the FCLK
30:17gets, and also the more heat and power
30:20you draw, and also the more dangerous it
30:22gets for your CPU durability. This one
30:25you kind of have to play with yourself,
30:27figure them out. Some of these you will
30:28find in the build zoid, the ones for the
30:30DRAM. This I'm showing you what I have,
30:32but I'm not going to spend much time
30:34talking about them. Now, other there's
30:36just a couple more important things.
30:38First of all, you need to make sure that
30:40you have resize bar above 4G decoding
30:42enabled. This is really good because it
30:44allows games that Nvidia wide lists if
30:47you're using Nvidia to use resizable
30:50bar, which is basically it can address
30:52the memory in the GPU directly from the
30:56CPU in larger chunks than what would be
30:59otherwise possible. So, just just turn
31:01it on. It's probably going to be on by
31:02default. So, just wanted to mention that
31:04one. I highly recommend you turn off
31:06your integrated graphics unless you're
31:07using it for something because I did
31:09notice very big stability differences in
31:12mine, especially Steam. It's a known
31:14problem that I found on the Steam
31:15forums. Steam would actually crash my
31:17computer sometimes when I was
31:19overclocking because of the integrated
31:20graphics being on. Super weird. So, I
31:23disabled it and it fixed it. So, I
31:26recommend disable if you're not using it
31:27because why not?
31:30Then this one, experiment yourself. Some
31:33people will tell you that it's more
31:34stable with C states on. Some people
31:36will say that it's more stable with C
31:37states off. I honestly again don't
31:40notice a difference. And I've never
31:41noticed my CPU actually entering like C6
31:44or any of the sleep states no matter
31:47what, which is kind of weird, but that
31:49also probably means that this isn't
31:50doing really much on my CPU, even though
31:52a lot of people will tell you this does
31:54a lot. This doesn't change whether your
31:56CPU is in C 0 or one which is basically
31:59the like super like focus high
32:01performance or still like high
32:03performance but not uh focus it still
32:05goes 01 anyway so this doesn't control
32:07that so I'm not exactly sure now IO MMU
32:10if you disable this you will get more
32:12performance especially in something like
32:14CS but first of all it's a security
32:15feature so a lot of the things I will
32:17disable you need to know what you're
32:19doing if you have very important stuff
32:21on your PC it's your own risk I'm okay
32:24with it But just know that all of the
32:25virtualization features allow for bad
32:28viruses to get even deeper in your
32:31system and get access to more stuff
32:33because there's no virtualization. The
32:35drivers for a malicious attacker now get
32:38control over your memory CPU uh a lot
32:41more directly but it does give you extra
32:44performance but face it won't allow you
32:46to play. So depends on what you you play
32:50on what you you what you're doing. you
32:52will get more performance if you turn IO
32:54MMU off, but you won't be able to play
32:57Face It. So, right now it's disabled for
32:58me because I haven't played Face It last
33:00few days. It's really quick for me to
33:02enable disable it. I kind of go back and
33:04forth depending on whether I'm playing
33:05Face It or not. Another thing on the
33:07same topic, I am MMU. There's also SVM,
33:11which is CPU virtualization. This also
33:13gives me more performance when it's
33:14turned off. It also turns off memory
33:17integrity in uh Windows, which is a
33:20security feature. you won't be able to
33:21play face it if if you turn this off,
33:23but you get more performance. So, up to
33:25you if you want SPM off. Now, another
33:27quick thing here is this one was
33:29actually meaningful for me. The LDR, you
33:32need to make sure that your GPU is in
33:34the correct slot obviously, and it was
33:37for me, but my motherboard shares that
33:41PCIe lane with an NVME slot that was
33:44close to it and also another PCIe lane.
33:47So basically, if I had anything in those
33:49two, instead of being PCIe Gen 5 x16, it
33:54switched to PCIe uh Gen 4 8X, which is
33:59only like a 2 to 3% difference in my in
34:02a 4090. But, you know, I was trying to
34:05push as high as I as I could. So, I
34:07removed the NVME. Um, and you can even
34:09force it to be X16 like this, but I know
34:11auto mode will go X16 for me because
34:13I've got nothing on M.23. 2_3
34:17and PCI x16_2 which is the PCI e slot
34:21just under it and the NVME. So be
34:23careful because you might be using your
34:24super OP 1590 maybe at like PCI gen 48X
34:28because you you put an NVME in the wrong
34:30slot that you didn't realize. So use
34:32something like GPU Z to check the actual
34:35like bandwidth of your PCI lane for your
34:37GPU because you might be leaving a
34:39couple percentage points of performance
34:40on the table. Last thing to do here, if
34:43you go in the AMD overclocking area,
34:46which will have a lot of the settings
34:47that we've seen in the AI tweaker thing,
34:50only change them in one of the two
34:51because who knows which one is going to
34:53be prioritized if you change it in both
34:55is, but you will find this in build zoid
34:58stuff too. The nitro mode for DDR5.
35:02Again, I'm not going to spend much time
35:03talking about these, but know that as
35:05per usual, you get more performance, but
35:07less stability if you tweak these. Uh, I
35:09just wanted to show you what I have
35:10because a lot of people just want to
35:11know what I'm currently running. Apart
35:13from that, I haven't changed anything
35:15else in my BIOS. [music] So, you've seen
35:18everything. Quick recap. These are my
35:21voltages.
35:23This is my VRM setup. I have clock
35:26spread spectrum off. This is my core
35:30tunings.
35:32I'm running slightly higher BCLK, but
35:34don't do it. It's honestly doesn't
35:35really change anything to me either. I'm
35:37just trying to figure out what the limit
35:38is for me. I'm using 6,000 even though
35:40it's a 6118 because of the VCLK 2200 MHz
35:45Infinity Fabric. PBO is at the limits of
35:48the motherboard with 10x overdrive
35:50scalar and minus 27 all chorus curve
35:53optimizer and my DRAM is running uh with
35:57MCR and PD enabled UCLK equals MCLK
36:00obviously gear down mod enabled and
36:02you've seen the timings earlier but I
36:04will just quickly scroll through them
36:05again in case you want to stop and use
36:07this as a [music] as um uh timestamp
36:12reference then in advanced my only
36:14differences from talk are that I have
36:17right now SVM off just for the
36:19benchmarks of course I have rebar turned
36:22on then I have PCIe I made sure that
36:26it's on 16x my integrated graphics is
36:29disabled [music] my IO MMU is disabled
36:33and cate is disabled and my RAM
36:37nitro looks like this I also have a
36:41custom fan curve which I will quickly
36:43show it's pretty basic it's probably way
36:45easier to see like this. But basically,
36:48my CPU will uh ramp up to 100 when it
36:53hits 42 C.
36:56So when my CPU when my CPU temperature
36:59hits 42C, it will go from about 30% to
37:02100. My pump is always at 100% because
37:05the changes in its speed were bothering
37:08me in terms of sound. And my chassis
37:10fans are all at 100% all the time
37:12because they're really quiet and I want
37:14a nice air flow in general. So that's
37:17all for my BIOS. So CPU and RAM. Now
37:21let's quickly go check my GPU in the
37:23operating system. Okay. So now for the
37:25last part of the tier one, which is the
37:26hardware optimization. Let's talk about
37:28optimizing my GPU. Um, which implied of
37:31course overclocking it and a little else
37:36which means overclocking. And in my
37:38case, I also flashed a different bias,
37:40which up to you if you want to risk it.
37:43My GPU has a dual bias, so it was very
37:45safe for me. Um, but there's always a
37:47risk there. Um, I am pretty sure that
37:49there's risks involved with even just
37:51explaining how to do it based on the
37:53fact that no one seems to explain it.
37:55But you can look it up. There's a GitHub
37:57for a program called NV FlashK, which
38:00allows you to flash BIOSes that aren't
38:02compatible with your GPU. And then you
38:04can find BIOS on tech power up.
38:07um because they have a little nice
38:10database. Why did I do that? Because my
38:124090 by default would only allow me to
38:15overvolt it up to 1 070
38:19m volt. Originally 490s could boost
38:21could be pushed all the way to 1.1 volt,
38:24but then uh GPUs that were made later on
38:28were prevented from doing it. So if you
38:29flash a particularly old BIOS, you can
38:32get back to 1.1 molt, which allow you to
38:34get a little extra clock at higher
38:36temperature. Obviously, uh I remember
38:39that I also repasted my GPU with PTM
38:417950 or really I wouldn't have been able
38:43to squeeze any extra juice from
38:45overclocking. This is my uh everyday
38:47stable overclock where I can run OCT 3D
38:50Extreme and get no memory errors u ever.
38:53Basically, I ran it for hours. Now, when
38:55I got the 3D Mark scores, I wasn't using
38:57this BIOS, but I was using a different
38:59BIOS that allowed that allows me to lift
39:01the power limit at all. The power limit
39:03for the BIOS I'm using right now, which
39:04is the one that I use daily, is 600 W,
39:07which is the safe value for the
39:09connector, the infamous connector from
39:10Nvidia, the 12vt um uh 16 pin connector
39:15that is known to melt.
39:18I just for the benchmark I used a 100vt
39:21extreme overclock modified BIOS that
39:23allowed me to lift the power limit to a
39:251000 watt and I've seen it transiently
39:27hit like 750 which is not safe. So don't
39:30do it unless you know what you're doing
39:32and you okay with risking a little fire
39:34on your connector. Um and in case you
39:37want to know the overclock uh that I use
39:39on 3D Mark was this plus 205 plus 79 70
39:42uh 1700 which got me to 12.2 2 GHz on
39:45the memory and it allows this to boost
39:48up to 3.1 GHz um at a lower temperature.
39:51Of course, my fans were spinning at 100%
39:54and I started from about 28C uh on the
39:57GPU when I was doing the benchmarks. But
40:00my daily uh which really is basically
40:02the same performance overclock that I
40:04use is this. I don't lock my boost so
40:07that I allow it to get to lower power
40:09levels. But I do use Nvidia preferred
40:12maximum performance which means that it
40:14will never really go to the like 200 300
40:17uh hertz values, MHz values um that it
40:21would get. It doesn't really change
40:23much, but it allows you to be a little
40:25cooler at idle. Uh, of course, I use the
40:27maximum power and GPU temp limits, but
40:29they're never hit in real gaming. And I
40:31use 12.002 002 GHz
40:37memory clock which is a plus 1500 on my
40:40specific BIOS compared to its base
40:42memory values. So 140 on the clock so
40:45that it boosts to uh about about three
40:48GHz on the clock. Of course voltage I
40:51let it go all the way up to 1.1 volt.
40:53Maximum power limits 15 under clock. And
40:55then I have a custom fan curve similar
40:57to the one you saw in the BIOS for my
41:00CPU where at idle like now it's only
41:03running at 30% so that my PC is super
41:05quiet and then if it gets any higher
41:08than 40C they spin up to 100% and stay
41:12there. That way I don't let it just
41:15hover at like 70C, but when I'm gaming,
41:17for example, in Call of Duty, it's about
41:1945 C, which lets it be at 3 GHz all the
41:22time as I'm gaming. Um, with maximum
41:24stability. The software here is
41:26Precision X1. I've also used MSI After
41:28Burner a lot. Um, I just kind of like
41:30Precision more. I like the UI. Um, and
41:33it seems like it applies the voltage
41:35more consistently even though I can't
41:36tweak the curve like I can with MSI
41:38After Burner. So, of course, this will
41:41depend on what GPU you have, how
41:43comfortable you are, your thermals, but
41:44I just wanted to show what I did, what I
41:46use every day, and what I use for the 3D
41:48Mark course. Okay, here we are. Tier
41:51three, the tip of the iceberg, Windows
41:54optimization. There's a good chance that
41:56you're here hoping that I'll give you
41:57the next holy grail registry tweak
42:01that's going to improve your latency or
42:04whatever. And I will have to disappoint
42:08you. I fell for the same trap believing
42:11that there was some sort of tweak and
42:14that some of the things I was trying the
42:16popular ones were actually doing
42:17something I could perceive until I
42:19decided, hey, I want to be rigorous. I
42:22need to make sure that whatever is there
42:24I can measure. So I built my measuring
42:26tools which I've discussed on Twitter,
42:29Discord, and I'm going to make more
42:31videos about. But this video will be
42:33insanely long if I give you the details
42:35of my measuring devices and me measuring
42:39methodologies.
42:40Just know that absolutely none of the
42:45registry tweaks or the win32 separation
42:48or the interrupt affinities and policies
42:50or trying to optimize ISR and DPC and
42:53manually dealing with your interrupts
42:55did anything.
42:58All of them consistently change my click
43:01to photon latency by zero.
43:06I could reproduce to the 0.01 01
43:10millisecond precision with 100%
43:13consistency, zero effect from any of
43:17these tweaks. There's things that you
43:19can do in Windows to improve your
43:21latency, which I will explain, but
43:24they're very specific and they're not
43:27the tweaks that you think they are. So,
43:29let's define a ground rule based on my
43:32testing, and I'm open to someone proving
43:34me wrong. If a tweak has no effect on
43:37frame timing, so average FPS, 1% FPS,
43:40something that you can easily measure,
43:42then they do not affect latency. Again,
43:45I will gladly be proven wrong. But
43:47everything I've tested shows that if a
43:50latency difference is measured, it is
43:53only thanks to changes in frame time. So
43:56unless you see your FPS increasing,
43:58which you don't need knowledge or
44:01research to test, it's very easy to
44:03test. You're probably not getting any
44:06latency change from a tweak. So that
44:09said, clearly I've shown you these
44:11charts before and the tuned OS has a
44:14difference in FPS. How do you achieve
44:17that? There are five main things you can
44:19do. I'm going to keep them simple.
44:21Everyone has a lot of nuance. I can't
44:23cover it all in a single video that's
44:25already this long, but it's what I'm
44:27doing to show you the larger numbers.
44:28The most obvious one is you don't want
44:31stuff running on your operating system
44:33competing for resources with your game.
44:36This will cause a measurable difference
44:38in your average and especially low
44:40percentile FPS. So, you need to decrease
44:42the number of background things that are
44:44happening. So, you need to debloat your
44:46Windows. I do not recommend you install
44:50a custom Windows image. That's an insane
44:52risk. I don't see why anyone would do it
44:55unless you are kind of like in the Linux
44:57space where you're going to be checking
44:59the image yourself which no one
45:01is doing. You don't even know where the
45:03image comes from. Please, it's just it's
45:06just not wise to do. just if you want
45:10reinstall your Windows so that you start
45:12from the default level of blo in your
45:14windows and then there's some open
45:16publicly available tools like wind toys
45:19or Windows optimizer which are honestly
45:21unnecessary if you're willing to do some
45:23stuff yourself manually that are going
45:25to remove some of the more obvious
45:28bloated services that maybe I don't know
45:30your motherboard tried to inject like
45:32Asus motherboards tried to get you to
45:34install Armory Crate when you uh started
45:37with Windows
45:38If you remove those and you can easily
45:40just open your task manager and check
45:42what's using your CPU. No magic, no
45:44voodoo involved. If you debloat your
45:47Windows so that when your game is
45:49running, there's no other apps competing
45:52for any meaningful amount of CPU time.
45:55There's not going to be a change in your
45:57FPS. So I debloated largely manually by
46:01going through my process list and my
46:04services and finding stuff that was
46:06actually running and I could see running
46:08and removing it. And that had an effect
46:12on overall frame timing across long
46:16periods of time because the times where
46:18that process would have stolen resources
46:20by running in the background was now not
46:22happening anymore. I will add this bit
46:24here but I do not recommend it. There
46:27are two main Windows important services
46:30that actively compete for resources and
46:32are noticeable for games and these are
46:36Microsoft Defender and the Windows
46:38Search Indexer.
46:40These are two important services in your
46:44operating system.
46:46So, I do not recommend anyone turn these
46:48off, but I'm not going to claim that
46:52having them on or off will never make a
46:54difference. Obviously, it is more
46:56important for your operating system to
46:57make sure that you're not about to run a
47:00malware or actively running malware than
47:02it is for you to get at the highest
47:04possible frame time. So, if you disable
47:07the Windows offender completely or at
47:10least surgically disable the real-time
47:12protection, the malware analyzer, and
47:15you disable the Windows search indexer,
47:18but your searches are going to suffer
47:20even those in start. So, again, I do not
47:22I do not recommend it. you will have a
47:24lower chance of your FPS being affected
47:27by these background services. I'm not
47:29recommending you do it. I did it for
47:32testing and could measure changes by
47:34removing Defender completely through
47:36Windows 11 Pro group policy settings,
47:40which I do not recommend you do. But I
47:42want to be fully honest and say that
47:45those do affect your frame times. But
47:48the risk versus reward unless you're
47:51trying to like get a high score on a
47:53benchmark or something in terms of FPS
47:56benchmark, not aim training benchmark
47:58are to me not worth it. Number two tweak
48:01that actually works is I could
48:03consistently statistically meaningfully
48:06reproduce using any sort of VSYNC or
48:10G-Sync implementation. So that it means
48:13also adaptive sync, free sync, whatever
48:17technology name you want to use. That
48:18same concept of trying to synchronize
48:20your refresh rate with your frames, add
48:23latency. Even if you do all the correct
48:26steps, you use the minus 3 FPS limit.
48:29What I'm not going to mention these
48:30because those are if you're interested
48:32in having a great G-Sync experience,
48:34which I recommend unless you're trying
48:36to chase that lowest latency possible.
48:40But I could easily measure and reproduce
48:42an increase in latency. We're talking
48:45about 1 2 3 milliseconds on my testing
48:49framework if you use any of the
48:52synchronization algorithms and systems.
48:55So if we're talking about purely
48:57optimizing for latency and when I say
48:59latency I mean input lag I suggest that
49:02you don't use bsync or g-sync. Third
49:04thing, and this is a universal truth, a
49:08pretty easy to understand concept, I
49:10would say, do not lock your frame rate
49:13to your monitor hertz unless your game
49:18specifically doesn't play well with
49:21uncapped frame rate. Because I'll give
49:23you a quick example. Say that you are on
49:25a 60 Hz monitor. That means that you
49:28will see a frame on screen every 16.6
49:32milliseconds. If you get 60fps and you
49:36don't have any synchronization, that
49:38means that you have a chance of seeing a
49:40frame that is up to 16.6 milliseconds
49:43old after 16.6 milliseconds. But if you
49:48say have infinite FPS at 60 Hz, that
49:51means that every 16.6 milliseconds, the
49:55frame visible on screen will be fresher
49:59than if you were getting lower frame
50:01rates. So, even if you are already
50:05achieving your monitor's refresh rate,
50:08more FPS will always decrease input
50:11latency because you will see the same
50:13amount of frames on screen per second,
50:16but they will be closer to now in actual
50:21time. Some games do play poorly with
50:25uncapped FPS. So, test on your game. If,
50:28for example, the price to pay for
50:31locking your FPS is getting way worse
50:34frame times on average or getting really
50:37bad dips, then it's not worth it,
50:40obviously. But you can measure the
50:42expected change yourself. You take your
50:44FPS and you do 1 divided by FPS. That
50:48gives you the frame time between two
50:50frames. And that's the theoretical
50:53maximum amount of time that can be
50:56elapsed since that frame was generated
50:58when it's shown on screen. FYI for the
51:00very technical people, this is a large
51:02simplification because that's not how
51:04actual rendering works. We would need to
51:06talk about frame buffer cues, Nvidia
51:08reflex, the presentation model. It's
51:11more complicated than that, but in
51:12general, it's true. If you don't change
51:15any other variable, higher FPS means
51:18lower latency. And this is super easy to
51:20reproduce and super easy to calculate by
51:22how much changing frame rate increases
51:24your latency.
51:27So if you can don't cap your FPS, let
51:30them run higher than your monitor
51:32refresh rate. Number four, I know that
51:36you probably think full screen
51:37borderless
51:39is the same in modern day as full screen
51:43exclusive. And that is true if a certain
51:47set of conditions are fulfilled.
51:50The safer approach unless you have the
51:53technical knowhow to verify that those
51:55conditions are fulfilled is to use full
51:58screen exclusive. For example, I have a
52:02540 Hz monitor on a 4090 which I have
52:06and any other 4000 series Nvidia GPU. to
52:09implement 540 htz through the SC. It's
52:13going to end up using two display heads
52:16inside the GPU. That means that the
52:20display acts as a tiled display inside
52:23Windows, which means that mo doesn't
52:26work. Which means that I don't get the
52:29flip presentation model optimizations
52:31that have made for a certain amount of
52:35people in a specific setting. full
52:37screen borderless equal to full screen
52:39exclusive. So in my setup, which isn't
52:42that uncommon, and there are other ways
52:44that could force your MP not to be on, I
52:47have an very easy 100% reproducible on
52:51any game, decrease in frame time, and of
52:55course an increase in input lag if I use
52:57full screen borderless. Same thing
52:59happens if I use full screen optimize
53:01and I use overlays like the Steam
53:03overlay. So in my case I have to use
53:07full screen exclusive
53:09on my main gaming monitor and I have to
53:12avoid using overlays.
53:14Now,
53:16there are again cases in which full
53:18screen borderless is equivalent to full
53:20screen exclusive. But unless you are
53:221,000% sure that your MO clip
53:25presentation model optimizations from
53:27Windows are working correctly, it's
53:31safer for me to say just play full
53:33screen exclusive so that you don't risk
53:36not having the flip presentation model
53:38working. Number five, and honestly the
53:41most tweakingish of the changes that is
53:45surprising
53:47that is kind of surprising that it works
53:49is and you need to test this on your own
53:52CPU for my CPU and a lot of other CPUs
53:56that I've seen. If I prevent my games
53:59from running on core zero, physical core
54:02zero, which means that if you have
54:04hyperthreading on would be for my
54:06architecture, logical cores 0 and one,
54:10if I prevent my game from running on
54:12logical cores 0 and one, I get increases
54:15in both average and 1% FPS up to 8%
54:19average and 22% 1% in CS2, which is
54:24insane, and you wouldn't expect that to
54:26be the And I've tried figuring out why.
54:29And I think the reason is just that most
54:31legacy code probably including a lot of
54:34low-level kernel stuff in Windows tends
54:37to spawn its thread on core zero and
54:40especially the low-level Windows stuff
54:42will obviously have a higher priority
54:44level compared to a user space process
54:46like your game and therefore probably
54:49end up preempting your game thread more
54:52often on core zero.
54:55This is super easy to test. Open your
54:57game, open your task manager, go to the
55:00details tab, find your game process,
55:03right click, select the affinity, and
55:05try removing core 0 and one. Or if your
55:09architecture defines the two logical
55:11threads for your core zero as something
55:13like zero and maybe 14. This one you
55:16have to look up yourself.
55:19Disable those. basically disable both
55:22logical cores from the physical core
55:24zero and see if something changes. It is
55:26important that you disable both of them
55:28or you will actually make things worse.
55:30If you do that correctly, for me in
55:33every game that I've tested, this
55:35increases both average and 1% FPS. So,
55:39this is a little bit magical, but again,
55:41easily justifiable, easily testable. You
55:44don't need to do research. You don't
55:46need to measure latency with a complex
55:49electronical uh end to end latency tool.
55:52Just change it and see if you have FPS
55:55changes. If your FPS changes, then it
55:57affected your latency. Now, if you
56:00change it from the task manager, you
56:02have to do it every time your process
56:03starts. So, up to you what you use, but
56:06I've been using process lasso because
56:08it's basically just a tool that allows
56:10you to define rules that then apply
56:12these affinities manually. and I just
56:15applied it to my steam.exe because these
56:17affinities are inherited by processes
56:20spawned by a specific parent process in
56:22most cases. You need to check your game.
56:25So, I just set up a rule that my
56:27steam.exe
56:29only runs on my cores that are not the
56:32physical core zero and I don't have to
56:35think about it anymore. Now, someone
56:37would say they're using something that
56:38process lasso introduces overhead. Sure,
56:41you have something running in the
56:42background to check if a process
56:44started. I could not measure any impact
56:47on frame time or latency, of course. So,
56:49I'm okay with process loss. So, I think
56:51it's a pretty cool software. Up to you,
56:53but this one was very measurable to me,
56:56so I had to mention it. And it's the
56:57only real tweak that actually did
56:59anything. And that's it. Literally
57:02nothing else I tried had any measurable
57:04effect. I'm going to mention now briefly
57:06because I said I would two changes that
57:09are probably in the placebo territory,
57:13but that might maybe theoretically do
57:16something. I'll be quick because I don't
57:18think you should waste actual time doing
57:20this yourself. One is that I bought a
57:22PCIe card that uses an as media ASM 3142
57:27XHCI
57:29controller. Again, can't go into
57:31details. this video will be too long,
57:33but just know that it's basically going
57:36to make it so that your mouse packets
57:38are handled by a different piece of
57:40hardware that will only be handling
57:42those and will pull only your mouse,
57:45which is the XHCI controller on the card
57:48and also using a different driver
57:50because I specifically installed as
57:52media's implementation of the USB driver
57:55and not using the Windows default one.
57:58This technically could make it so that
58:01there's more free
58:05real estate for your mouse interrupts
58:08not to collide with anything else or be
58:10moderated through something called IMOD,
58:12which is interrupt moderation. But if
58:14you stop to actually think about it and
58:16do some math and some statistical
58:17analysis or some testing, which I've
58:19done, it's actually way more likely that
58:22it does absolutely nothing. And I could
58:24measure no statistical difference
58:26between using my CPU connected USB,
58:29which is another thing that I tested not
58:31to be meaningful, motherboard connected
58:33USB, or this PCIe as media card with its
58:37own XHCI
58:39controller and its own driver. So maybe
58:42depending on how many peripherals you
58:43have connected to a specific XHCI driver
58:46on your hub on your motherboard, using
58:50your mouse on its specific PCIe card
58:53could maybe do something. I couldn't
58:55find a situation in which it could be
58:57measurable, but it's only €23.
59:00If you are seeing issues with, for
59:02example, running 8k Hz even after you do
59:04everything else that I've mentioned,
59:06maybe try spending €20 on a PCIe USB
59:08expansion card. Up to you. The number
59:12two placebo probably tweak is trying to
59:16change your interrupt affinity mask and
59:20priority. The priority itself, I'm
59:22pretty sure absolutely does nothing when
59:25you have MSI enabled, but again, can't
59:27go in details because it will be too
59:29long. But changing your affinity mask
59:32does do something easily reproducible.
59:34If I assign my for example PCIe XHCI
59:38controller to have an affinity mask for
59:41its interrupts to for example only
59:45handle interrupts on cores 14 and 15.
59:48Then I can see that the DPCs which is
59:52part of the whole interrupt pipeline ESR
59:55causes the DPC the DPC which is the
59:57actual handling of the interrupt happens
1:00:00on those cores. So I can see a world
1:00:03where making sure that the popular ones
1:00:06are usually tuning your GPU interrupt
1:00:09affinity mask and your mouse
1:00:13actually your mouse's controller
1:00:15interrupt affinity mask. I can see a
1:00:17world where depending on what's going
1:00:19on, this makes it so that your mouse
1:00:23doesn't have to compete for interrupts
1:00:25with something else. Same with your GPU.
1:00:28But again, I could find no real world
1:00:31scenario that I could run and measure
1:00:34any difference. So again technically it
1:00:38makes sense and you can measure
1:00:40differences if you use ISR
1:00:44and DPC analysis tools like the Windows
1:00:47performant analyzer toolkit. You will
1:00:49see these changes do something at the
1:00:52interrupt level. But if you stop to
1:00:54actually do the math, you will realize
1:00:57that again, unless there's a very
1:00:59obvious visible problem, the likelihood
1:01:02of that affecting input lag is basically
1:01:05zero. If you're really interested, go
1:01:08lurk my Twitter and you will find a
1:01:10number of threads where I went through
1:01:12the math and shown why statistically it
1:01:15is extremely unlikely that optimizing
1:01:17your interrupts at the ISR and DPC level
1:01:22actually affects input lag. But again,
1:01:25maybe in some specific case, these two
1:01:27tweaks might do something. If you have
1:01:31nothing better to do with your time and
1:01:32you also want to waste time like I did,
1:01:34try them out for yourself, make sure
1:01:37that you're not being placebo into
1:01:38thinking they're doing something and let
1:01:40me know if you can measure any change.
1:01:41And now that we're done with the placebo
1:01:43part of the Windows optimization as
1:01:45well, we're done. I can tell you there's
1:01:49a very high chance that anything you can
1:01:51tell me to try, I've already tried. I've
1:01:53tried Win32 priority separation. I've
1:01:56tried the hidden power plant settings
1:01:57that you can tweak. I've tried all the
1:02:00registry keys are supposed to play with
1:02:02the power saving features on your GPU
1:02:04and CPU. If you're actually playing a
1:02:07game running a DirectX app, I could not
1:02:11measure any difference down to the 0.1
1:02:15millisecond. If you're confident that
1:02:19you have the holy grail of tweaks, the
1:02:22registry secret key that will cause
1:02:25anything to be measurable in some
1:02:28rigorous statistically significant way,
1:02:30please let me know so that [music] I can
1:02:33openly and happily claim that I found
1:02:36the one tweak that works. But unless you
1:02:38can prove
1:02:40that something that you're tweaking will
1:02:43consistently statistically probably
1:02:47cause two frames that to be different
1:02:51versus the scenario in which you didn't
1:02:53have the tweak then it is logically
1:02:56impossible to claim that that has a
1:02:59perceivable effect by human. And I'm not
1:03:01saying because it happens faster than a
1:03:03human. I'm saying that unless it's
1:03:06happening so that the frame that is
1:03:08being sent to your screen is actually
1:03:10different because the effect of the
1:03:13thing for example the interrupt from the
1:03:14mouse caused it to be visibly different.
1:03:18What is being shown to the user is
1:03:20exactly the same and therefore there is
1:03:220% chance that a user could even feel
1:03:25the difference even assuming they had
1:03:27infinite discerning capabilities in
1:03:29terms of frequency or latency. So unless
1:03:33someone can tell me what it is that can
1:03:36cause that, I will keep saying that
1:03:38unless you see a natural change in frame
1:03:40time, you are not going to see a change
1:03:43in latency. And just as a last thing,
1:03:45let me clarify why I'm saying that if
1:03:47there's a change in frame time, you will
1:03:49see a change in latency. This is very
1:03:51easy to understand. Frame rates are
1:03:54usually under a,000 FPS. When you reach
1:03:571,000 FPS, you now have things happening
1:04:00at 1 millisecond frequency or higher.
1:04:04When you're below that, which is where
1:04:06the vast majority of your gameplay will
1:04:07be, everything is in the greater than 1
1:04:10millisecond order of magnitude. So, for
1:04:14example, if your average FPS goes from
1:04:18400 to 250 average FPS, that means that
1:04:21the time that elapses between two
1:04:23consecutive frames decreases by 1.5
1:04:27milliseconds, and that's measurable, but
1:04:29that's also very easy to calculate. You
1:04:31don't need magical tweaks to figure that
1:04:34out. So, higher FPS is always going to
1:04:38be lower latency at a at an order of
1:04:40magnitude much greater than anything
1:04:43these tweaks ever claim to do. So, if
1:04:46you want to focus on something, focus on
1:04:48ways in which you can increase your FPS
1:04:49and share them with the world because
1:04:51they will probably not be the secret 100
1:04:54registry tweaks that these people are
1:04:56trying to sell you. So, that's it. This
1:04:58is my optimization video. I wasted an
1:05:01ungodly amount of time on it. I recorded
1:05:03it multiple times. I wanted to be direct
1:05:06but also accurate. And I'm sure there
1:05:09must be some inaccuracies considering
1:05:12considering how much I yapped about
1:05:14this. So, please let me know so that I
1:05:16can have a pin comment where I clarify
1:05:18anything that I wasn't accurate or I was
1:05:20actively wrong about. And if you
1:05:23actually watched everything up to this
1:05:25point, thank you. I do not expect many
1:05:27people will do it, but I am super
1:05:29grateful. And I'm also super grateful
1:05:30for everyone on Twitter and Discord that
1:05:32was actually open to discussing these
1:05:35topics without preconceived notions,
1:05:38without their bias. I think that was
1:05:40super helpful. And I'm also thankful to
1:05:42all the people that probably felt like
1:05:44their business was being challenged. all
1:05:46the tweakers that actively tried being
1:05:49negative in my comments and even
1:05:51threatening me because they fueled my
1:05:54will to actually go through build my
1:05:57electronics and do all the testing which
1:05:59took many many days and of course the
1:06:02biggest thank you to my patreons whose
1:06:05names you will see on screen now I still
1:06:07am shocked that anyone is willing to pay
1:06:09considering I haven't really put much
1:06:11behind the payw wall but considering how
1:06:13much I've learned in my research Here,
1:06:15I'm planning on putting a lot of very
1:06:17technical content behind the payw wall.
1:06:19For example, the specifics of how I
1:06:22built my electronics down to my STDM32
1:06:25code, my setup for its ADC, my setup for
1:06:29its timer, things that most people
1:06:31probably won't care about, but maybe the
1:06:32few crazy people will care about. And
1:06:34since they took me so much to figure
1:06:36out, I feel like it's fair for me to put
1:06:38behind the payw wall. Of course, the
1:06:40results of my testing will always be
1:06:42public, and I'll make more videos about
1:06:44it. Now that I'm finally done with this
1:06:46video, I'll also go back to actually
1:06:47reviewing preiprols, which I have a lot
1:06:49of backlog for, and trying to get world
1:06:52records and scores in aim trainers,
1:06:54which I still want to do. So stay tuned.
1:06:57I'll see you soon. Ciao.

About this transcript

This page contains the full transcript of I Spent a Month Optimizing my PC so You Don't Have To by bardOZ, generated from the public captions YouTube serves with the video. The transcript has 11,928 words across 1,681 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.