TR Forums

Just had a thought cross my mind. These days, manufacturers (CPUs and GPUs alike) are pushing their cards to the ragged edge from the factory with dynamic frequency boost algorithms. With TR's introduction of inside-the-second analysis, the industry quickly became more in-tune to improving frame-time consistency and largely solved the "problem". However, with dynamic clock rates becoming increasingly common, perhaps we're effectively taking a step backward in "smoothness" in order to consistently squeeze out every ounce of performance a system can muster? Also, your "experience" may vary greatly depending on how much cooling you've invested in?

I certainly noticed an improvement in smoothness going to water cooling on my GPU compared to stock. Perhaps placebo effect but it felt pretty dramatic. It might just be higher sustained clocks in general but more cooling is certainly more better these days.

Same thing with the CPU on air versus under water.

It’s certainly worth looking into. I would imagine much of it depends on how the dynamic adjustments are implemented. For example, if a GPU’s thermal limit under x conditions is that it can run at 2025mhz, I imagine it would be preferable from a smoothness perspective to run at a steady 2000mhz, rather than repeatedly banging the RPM needle off the redline and then having to throttle back.

That was from TH's AMD 5700/5700XT review. Similar outcomes could probably be found in the CPU arena as well I presume.

It would seem (with reference coolers) that Nvidia is generally maintaining a more consistent clock rate than AMD. Due to odd dips, their Y-axis scale makes it pretty difficult to garner a % here, but it would seem the 5700XT clocks fluctuated as much as 15% (ignoring those odd dips). Similar results at PCPer

As TheEgg said, it all hinges on the "boost" duration each manufacturer is choosing to use. From the chart above, it would appear AMD is being more aggressive with their dynamic boost than Nvidia.

Gamer's Nexus also got similar results with the Sapphire 5700XT vs. the stock cooler. The Sapphire cooler was able to keep the card cool enough that it didn't throttle at ~40 dBA, whereas the reference cooler had a hard time doing that at 50 dBA.

https://www.gamersnexus.net/hwreviews/3498-sapphire-rx-5700-xt-pulse-review

DPete27 wrote:

Just had a thought cross my mind. These days, manufacturers (CPUs and GPUs alike) are pushing their cards to the ragged edge from the factory with dynamic frequency boost algorithms. With TR's introduction of inside-the-second analysis, the industry quickly became more in-tune to improving frame-time consistency and largely solved the "problem". However, with dynamic clock rates becoming increasingly common, perhaps we're effectively taking a step backward in "smoothness" in order to consistently squeeze out every ounce of performance a system can muster? Also, your "experience" may vary greatly depending on how much cooling you've invested in?

NVIDIA has been doing this for a few generations now, so it should be easy enough to go back and look at existing frametime data to see if it's better or worse than the "static clock" era from a consistency standpoint.

Certainly the majority (all?) of reviews leave everything on auto. Therefore, dynamic clock boosting is left to its own devices and consequently that behavior is displayed by proper frame time captures. Now that more reviewers are looking at frame times, dynamic clock boosting may be making some cards look "worse" than they actually would be if you were to manually "lock" the chip frequency.

Frame-to-frame behavior is easy enough to test pretty much anywhere you can get a fixed scene. Dynamic clocks do add a couple fps of wobble on a framerate counter, but it isn't a difference I can feel.

As far as lower-frequency swings, it'll depend on the game a lot. If naturally slow frames / scenes burn more power than fast ones, dynamic clocks will amplify variance, and if it's the inverse it's a win over fixed. A game could easily have situations cutting both ways.

I've run both ways quite a bit with a 480 (945 mV / 1160 MHz versus a full custom curve with a similarly low power limit) and haven't noticed any particular difference in feel, but net consistency is very rarely GPU-bound for me and I actively trade away consistency to improve latency at any opportunity I get. I've mostly settled on fixed now, but that's for perf/W reasons, not variance.