It seems strange to think, but your graphics card—a piece of hardware seemingly unrelated to the process of making music—could be one of the single biggest culprits when it comes to poor audio performance. Here's why—and what you can do about it.
One of the most heart-breaking things to see on music forums is the spectre of someone who has just spent thousands on a shiny new computer only to find themselves suffering from terrible, productivity-killing audio glitches. I often find myself reaching out to them to try and help diagnose the problem, and when I do my first question is always the same: "What sort of graphics card do you have?"
If the answer is "NVIDIA", then I've already got a pretty solid guess about what's going on.
The reason? NVIDIA GPUs are probably the single biggest drags on audio performance, a result of the huge DPC latency their drivers generate. It doesn't really matter how fancy your computer is or how expensive your audio interface is—if you have an NVIDIA graphics card, chances are it's going to cause you problems when it comes to realtime audio playback.
A Snarl-Up on the Data Highway
Every composer will instinctively feel a chill when they hear the word "latency". We generally think of latency as the delay between hitting a key on a MIDI keyboard and hearing a sound from our computer, which is largely a function of our audio interface and its drivers.
But there are other forms of latency that have little to do with audio drivers and are instead related to the way your CPU handles competing requests for its time. This is something called "deferred procedure call latency"—or "DPC latency" for short. Here's how it works:
Imagine your CPU is a busy highway being shared by millions of vehicles. The highway has multiple lanes of traffic (cores). In this analogy, your realtime audio processes are the equivalent of ambulances—they need to keep those lanes as unobstructed as possible because they have a LOT of numbers to crunch in just a few milliseconds in order to give you the feeling of instant responsiveness. Your audio samples are the patients. If the ambulance is delayed, corruption occurs and those patients don't make it to their destination (your audio interface) alive.
Now your realtime audio processes may demand the use of those lanes, but other drivers do too—and since each core can only process one task at time, any delays will result in a bottleneck. With so many drivers sharing the road, the operating system acts as a traffic control system—an army of traffic cops directing cars into lanes and occasionally stopping some to check their paperwork.
Some of those drivers are hauling oversized loads and have complicated paperwork, meaning they block the entire lane for an extended period of time while the traffic controllers check them over. This is where our ambulance can run into trouble. If it gets stuck behind one of these drivers (resulting in a lengthy "deferred procedure call"), it can't be processed quickly enough. By the time the ambulance is stuck, it's too late to switch lanes—they're all occupied. The patient, your audio, convulses on the gurney—and promptly flatlines. Click. Pop. Stutter.
NVIDIA's Drivers are a Massive Road Hog
In our highway analogy, you can think of NVIDIA's graphics drivers as an enormous eighteen-wheeler cutting across multiple lanes, spewing black smoke out its exhaust, adorned with offensive novelty mudflaps and a bumper sticker that reads "HOW'S MY DRIVING? CALL 1-800-LOL-BITE-ME". When this guy barrels through, he couldn't care less about your sad, pathetic excuse for an audio driver. This is his road, and he'll use it however he damn well pleases.
Now in fairness—and just to back away from this slightly tortured metaphor for a moment—there are some good reasons for NVIDIA drivers to be so demanding. Graphics cards are extremely complex beasts, and high-powered ones have to work hard to manage all the gaming optimisations, AI tools, ray tracing support and other features you might have bought them for. And to keep these cards from, you know, catching fire, NVIDIA constantly adjusts their power states—creating extra work for your CPU, even when your GPU isn't doing much.
Of course, AMD and Intel cards also manage power states and system resources—but their drivers are leaner and far less aggressive about wresting your CPU's time away from other drivers. That may partially explain why they are often outclassed in gaming benchmarks—but it also makes them much more reliable when it comes to realtime audio stability. (Not coincidentally, this is the reason we only put AMD GPUs in our machines.)
If you have an NVIDIA card, you can see the problem for yourself by downloading a free tool called LatencyMon. Run the programme for a few minutes, even while idling, and if you look at that line that says "Highest reported DPC routine execution time (μs)", I'll bet the driver listed there is nvlddmkm.sys—the NVIDIA Windows Kernel Mode driver. (If you don't see it, click over to the Drivers tab, sort by Highest Execution Time (ms) and you'll almost certainly find it in the top two or three.)
In some (though not all) cases, you may even see a message appear like this:
"Your system appears to be having trouble handling realtime audio and other tasks. You are likely to experience buffer underruns appearing as drop outs, clicks or pops. One or more DPC routines that belong to a driver running in your system appear to be executing for too long."
This is very bad news for your audio indeed. Let's take a look at why.
Sample Rates, Bit Depths and Buffers—Oh My!
You'll likely be familiar with the concepts of sample rates, bit depths and buffer sizes—but just in case, here's a quick refresher. (Feel free to skip ahead if this is old hat to you.)
Sample rate relates to the frequency at which audio is sampled (i.e. recorded) or played back digitally. For example, 48kHz means 48,000 samples—tiny slices of audio—are captured and/or reproduced per second.
Bit depth, meanwhile, determines the resolution of each sample—for instance, 16-bit audio has 65,536 possible amplitude values per sample, while 24-bit can have over 16 million values, resulting in greater dynamic range and more accurate sound reproduction. Encoding that extra resolution means more data is required per sample: 2 bytes per 16-bit sample, 3 bytes per 24-bit sample and 4 bytes per 32-bit float sample.
Finally, and most critically, the buffer size refers to the number of samples that are grouped together and sent to and from your audio interface at a time. These packets represent a trade-off between responsiveness (low buffer sizes) and stability (high buffer sizes), and the amount of data they contain scales with the bit depth. A low buffer size of 128 samples will actually mean sending 256 bytes if you're working with 16-bit audio, 384 bytes for 24-bit and 512 bytes for 32-bit—while at a higher buffer size of 512 samples you can quadruple all of those numbers.
These values directly determine your audio processing latency, i.e., the amount of time it takes for your CPU to process instructions like MIDI inputs or audio playback. The formula to calculate the processing time is:
Latency (in milliseconds) = (Buffer Size / Sample Rate) × 1000
For example, with a buffer size of 128 samples and a sample rate of 48kHz, your audio processing latency would be: (128 / 48,000) × 1000 = 2.67 milliseconds. This 2.67ms is your CPU's deadline—your system must process each buffer within this time to avoid glitches. If anything delays it, like a misbehaving driver, you'll hear crackles and dropouts. More on that in a moment.
While bit depth doesn't directly affect the calculation, it does increase the amount of data that needs to be processed within that 2.67ms window. Higher bit depths require more processing power and memory bandwidth, making your system more vulnerable to pops and stutters whenever there's an interruption.
Highway to Hell
Now let's examine the direct impact NVIDIA's driver can have on audio performance by returning to our highway analogy.
Remember our audio "ambulances" that need unobstructed use of the lane? As mentioned above, with a 128 sample buffer at 48kHz, a new ambulance must reach its destination every 2.67 milliseconds without fail. Each ambulance carries exactly 128 audio samples that must be delivered safely and consistently every 2.67ms.
Now, most modern CPUs won't actually take the whole 2.67ms to process those 128 audio samples. The amount of time it will actually take depends on how fast the CPU is and how high the bit depth is (remember that 32-bit float requires twice as much data as 16-bit). It will also be highly dependent on just how much number crunching is required to produce each sample—more (and more complex) plugins require more processing time.
But whatever the case, when our NVIDIA driver's eighteen-wheeler hogs a lane on this highway for 2.3ms (as in the LatencyMon screenshot), it is consuming nearly 87% of the available time window. If one of your audio ambulances happens to be using the same lane, it will be forced to wait while the graphics driver uses the CPU.
This leaves just 0.343ms for your DAW's ambulance to reach its destination—a brutally short window, even for the fastest CPU. In this tiny fraction of a millisecond, it must process all those virtual instruments, plugins, and mixing operations coming from your DAW. The odds of missing the deadline skyrocket. And if your audio fails to reach its destination on time, your audio interface has nothing to play, which is why you will start to notice stutters, clicks and dropouts. The patients—your precious audio samples—haven't made it.
Larger buffer sizes will of course help here. Using our calculation above we can see that a 256 sample buffer size takes the processing deadline up to 5.33ms, 512 samples to 10.67ms and 1024 to 21.33ms. Obviously a 2.3ms spike from your GPU isn't going to make nearly as big a difference in those scenarios. But the trade-off is that the higher the buffer size is, the less responsive your audio will be. I personally find a buffer of 256 samples works well for me most of the time, in that it's the maximum size that I can play my MIDI keyboard without any noticeable lag—but anything above that starts to feel uncomfortably delayed.
Whatever your settings, though, you ideally want to aim to get your DPC latency down below 1ms for a professional audio system. Anything above that puts your audio playback at risk.
So What Can Be Done to Fix The Problem?
It's important to say that not every composer or producer with an NVIDIA card will experience these issues. Older, lower-powered NVIDIA GPUs, for instance, tend not to have as many features and therefore cause fewer problems. And the degree to which DPC latency affects your audio performance will depend on just how complicated your DAW sessions are, how powerful your overall system is, and the sizes of your sample rate, bit depth and buffer.
(The quality of your audio interface doesn't make much difference at all here, by the way: any latency it introduces will be around the speed at which it converts analog sound to digital data and vice versa, which happens before and after the CPU processing we're talking about.)
But if you are experiencing audio dropouts and that infernal nvlddmkm.sys driver is lighting up LatencyMon, there are a few things you can do.
Install the NVIDIA Studio Driver
By default most people will have installed NVIDIA's Game Ready Driver, which is designed to take full advantage of your GPU's capabilities—but often at the expense of stability. For professional use cases, NVIDIA also offers a Studio Driver which can be downloaded from their website. This driver is tested more thoroughly, updated less frequently and is generally more stable than the Game Ready equivalent. It also reduces DPC latency somewhat, although it isn't a panacea: expect c.20-40% improvements rather than a complete eradication of the issue.
Before installing, I recommend running the Display Driver Uninstaller tool from Guru3D to fully remove the old Game Ready Driver. NVIDIA's own uninstaller doesn't do a great job of cleaning up after itself, leaving pieces of the old driver on your system that can cause issues when a new one is installed.
(NOTE: Not all NVIDIA GPUs are supported by the Studio Driver, especially not older or lower-specced models.)
Switch to the Ultimate Performance Power Scheme
Your Windows power scheme is a major factor in determining how your GPU's power states are managed. By enabling the Ultimate Performance Power Scheme, your GPU will run at a higher power state and "gear down" far less frequently, reducing DPC latency spikes. This does, of course, mean a marginally higher electricity bill—but for realtime audio, this is a step I recommend whatever sort of GPU you have. (All OPUS 101 Pro Audio PCs are set up this way when we ship them.)
The Ultimate Performance Power Scheme isn't available on most systems by default, but you can enable it by running Command Prompt as an Administrator and typing (or copying) the following:
powercfg -duplicatescheme e9a42b02-d5df-448d-aa00-03f14749eb61
(Yes, I know that looks a bit dodgy, but this is genuinely how it's done!)
You can then go to Control Panel > Power Options and select Ultimate Performance.
Adjust Settings in NVIDIA Control Panel
There are a couple of NVIDIA-specific settings you can tweak that may help to reduce DPC latency:
- Right click on your Windows desktop and select the option that says NVIDIA Control Panel.
- Look for Manage 3D Settings in the left-hand pane. From here, you should see a tab called Global Settings.
- Under Global Settings, set Power Management Mode to Prefer Maximum Performance.
- Set Maximum Pre-Rendered Frames to 1
- Finally, if you have a G-SYNC compatible monitor, look under Display on the left-hand panel and uncheck the option that says "Enable G-SYNC, G-SYNC Compatible".
- Click Apply to save your changes.
Disable Windows Game Mode and Game Bar
Windows includes several gaming-focused features that can cause interruptions and compete for resources with your audio workflow. These features prioritise frame rates and game capture over consistent, stable performance. Here's how to disable them:
- Press Win+I to open Windows Settings and click on the Gaming tab.
- Select Game Mode from the sidebar on the left and toggle Game Mode to Off.
- Click on Game Bar from the left sidebar. Switch Record game clips, screenshots, and broadcast using Game bar to Off.
- If available, you should also toggle Open Game bar using this button on a controller to Off.
- In the same Gaming section, click on Captures. Set Record in the background while I'm playing a game to Off.
- Click Apply to save your changes.
Switch to an AMD or Intel Graphics Card
Finally, it has to be said: your best bet is just not to use an NVIDIA GPU in the first place. AMD offer competitively priced Radeon GPUs with performance that is very close to (and sometimes even better than) NVIDIA's line-up. And just to show I'm not simply an AMD fanboy, I should also mention that Intel's Arc GPUs get excellent reviews and seem to have very low DPC latency as well.
If you do go down this route, make sure you watch some tutorials about uninstalling and installing GPUs as it's not always a straightforward exercise. You will also want to follow the instructions under Install the Nvidia Studio Driver about using Display Driver Uninstaller to remove the NVIDIA drivers.
Of course, if you are working on a laptop, you may not have a much of a choice here! But the good news is you're also less likely to run into these problems as laptop GPUs are generally less resource-intensive. I'd still recommend following the optimisation tips above though.
Conclusion
This whole issue essentially boils down to a battle between your GPU's desire to flex its muscles and your audio's need for consistent, predictable, realtime processing. Your computer is constantly racing against a clock determined by your audio buffer settings—and every time it fails to deliver the necessary processing on time, you're going to know about it.
The tweaks mentioned above can make a small but noticeable difference if you're stuck with an NVIDIA card—and it's worth mentioning that I know many composers and producers who make brilliant music with these cards every day. In many cases you may not notice any problems at all—but your NVIDIA card will be creating a bottleneck, even if you haven't run into it yet.
If you are purchasing a new computer (preferably one from OPUS 101 Pro Audio), then AMD and Intel GPUs will give you an easier ride when it comes to audio work, and I really recommend getting one. Ultimately, what matters most is getting your system to a place where the technology fades into the background. After all, the best computer is the one you don't have to think about because you're too busy creating an unbroken flow of amazing music.