Intel Graphics Software Performance Boost ignores power limit

[idmsa]

I was playing with some overclocking settings on my A380 and noticed some strange behavior.

For testing, I imposed the minimum power limit of 21W from the default power limit of 43W on my card. The GPU seems to respect this power limit as expected if the Performance Boost slider is set to 0%. But if Performance Boost is increased even to 1%, the GPU will suddenly jump to 2150 MHZ and pull about 40W. If the Performance Boost slider is raised from there, the clock speed will stay the same but the GPU voltage will drop. Why do lower power limits become irrelevant once the Performance Boost slider is touched? Does Performance Boost force the clock frequency to be at least 2150 MHZ? This is not intuitive to say the least.

The Performance Boost option also behaves somewhat weirdly at the default 43W power limit. I'll give an example at the 50% Performance Boost setting. Instead of picking a medium clock frequency and medium voltage, the GPU appears to rapidly fluctuate between 2150 MHZ at 0.83 V and 2600 MHZ at 1.01 V. In general, I noticed that as I increased the Performance Boost, the high frequency would get higher at a similar voltage, while the low frequency was always 2150 MHZ but with a reduced voltage. To explain this fluctuation, I was thinking that maybe there aren't any clock states between 2150 MHZ and 2600 MHZ, so the GPU has to switch between them quickly to stay near the power limit. Regardless, I found that this fluctuation seemed to result in more unstable frametimes. Additionally, this 50% Performance Boost yielded a mere 2% performance uplift, which is much less than I'd expect. So maybe this is a bug as well?

[jbruceyu]

let things straight up:

why the minimum power limit stops working once you touch performance boost?

well, on Arc cards, the Performance Boost slider doesn’t work like the overclock sliders you see on NVIDIA or AMD. Instead of just pushing clocks higher, it shifts the voltage-frequency curve and nudges the GPU toward higher boost states. The A380 has a hard-wired floor around ~2150 MHz. The second you raise Performance Boost (even by 1%), the card basically says, aight, I’m not going below this floor anymore. That’s why your 21W limit gets ignored, the GPU jumps straight into that ~2150 MHz state, which naturally pulls closer to 40W. so yeah... Performance Boost effectively forces a minimum of ~2150 MHz, which is why your lower power limits stop making sense once you touch it

why the card bounces between ~2150 MHz and ~2600 MHz?

Unlike NVIDIA/AMD, Arc doesn’t really give you neat “in-between” clock states. It works with bigger, fixed bins. At 50% Performance Boost and the default 43W limit, the GPU is basically power-balancing by hopping back and forth between:

• ~2150 MHz @ 0.83 V

• ~2600 MHz @ 1.01 V

That constant flip-flopping is how it stays under the power cap. The problem is, that bouncing shows up as unstable frametimes, what feels like microstutter to you

Why the performance bump is so tiny (~2%)?

Since the card spends most of its time hanging out in the lower state (2150 MHz), the quick bursts to 2600 MHz don’t really move the needle... You don’t get a smooth “middle ground” clock, just jerky toggling, which cancels out the performance gains you’d expect. This is why Arc overclocking feels kinda unrewarding right now. The boost logic is still rough around the edges, and there’s no way to manually control those in-between bins. 

about bug or design?

This looks like design, not a driver bug:

• The 2150 MHz floor is baked into the hardware.

• The oscillation is just the power governor doing its “bin-hopping” thing.

That said, the stutter and weak uplift make it clear Intel could improve how these transitions work or maybe open up finer control. You’re not the only one noticing this

what i'm try to say the bottom line of this is Performance Boost locks ~2150 MHz as the floor, so low power limits won’t stick. The back-and-forth between bins is expected, but it causes frametime instability and gains are small because the GPU can’t stay at the higher bin consistently under the cap.

...If you care more about smoothness than raw FPS, you’ll probably be happier leaving Performance Boost at 0% and just adjusting with the power limit

..hope this helps you

[idmsa]

Thank you, this is incredibly insightful. I was beginning to suspect this was a design decision, albeit probably a wrong one based on the poor overclocking results and frametime instability. The lack of documentation or other information online about this had me doubting whether this was universal behavior or just an issue on my system.

I'm assuming that the bins are built into the hardware in some way, given that Battlemage does support manual tuning of the V/F curve, though I don't know how well it works. If it was just a firmware limitation, Intel would surely have backported that feature to Alchemist, right?

Clearly, we have observed that any amount of Performance Boost will set a clock speed floor of 2150 MHz. I can't imagine there is a hardware reason for this being necessary. It's beyond me why they didn't just implement Performance Boost to shift the V/F curve up and call it a day. I'm certain there's an explanation for it, but it's not really acceptable that the power limit is thrown out the window.

Do you have any experience with Arc OC Tool? I'm wondering if it lets you configure a custom clock state that the GPU can sustain. However I'm hesitant to try it myself.

[AlphaTop89]

@idmsa this whole setup smells more like a design choice than a straight-up bug, but yeah… it feels like the wrong call when the end result is janky frametimes and weak OC headroom. Intel really ain’t do us no favors with the lack of docs either, so I get why you were second-guessing if it was just your rig or everyone’s.

On the hardware vs firmware side, you nailed it. Those bins are pretty much baked into Alchemist’s silicon. Battlemage getting manual V/F tuning is proof that Intel built that flexibility into the next gen, but Alchemist is locked the way it is. If it was just firmware, they probably would’ve patched it back by now. And yeah, that 2150 MHz floor once you touch Performance Boost? That’s just how the P-states were coded. Doesn’t mean it had to be that way, just means that’s how they shipped it. Like you said, they could’ve just shifted the V/F curve up instead of tossing the power limit out the window. Why they didn’t? Only Intel knows. 

But as for ARC OC tool, I’ve seen folks mess with it, but I’d be straight with you: it’s more experimental than polished. You can tweak custom states with it, but stability’s hit-or-miss, and it ain’t officially blessed. If you’re hesitant, you got good instincts , Arc still feels real picky when you start messing with bins it don’t like. If you do try it, keep expectations low and watch your temps/power draw close. this is how the card behaves by design. Battlemage should give us more freedom, but Alchemist’s boost logic is just stubborn like that. If smooth frametimes matter more than raw MHz, you’re better off chilling with Performance Boost at 0% and tuning the power slider instead

[idmsa]

Thank you guys for the info 🙂