AMD Radeon VII Chilling & Undervolting @ [H]

What about pairing a undervolt alongside Radeon Chill? Plus I'm curious about setting the Radeon Chill's min/max and how that could impact boost in regard to frame dropping the more unwanted side of the min/max/average frame spectrum. The hardware dropping some of the high frame rates so it can run cooler and boost higher on the minimum frame rates would be a worthy trade off. I find the minimum wattage and frequency on the undervolt really intriguing on the Radeon VII.
If there is a way to keep the minimum frequency/wattage identical that would probably improve the efficiency average a fair amount depending on workload.

That's why with a bios mod I tweak each p-state with a undervolt. Where you can run into issues though is if you set voltage too low and you run into corruption issues from the GPU being under fed voltage relative to frequency bad things happen at that point on screen. Too far of a undervolt turns into something reminiscent of a NES cart blow nostalgia experience quickly, but a less quick and easy fix at the same time.
At least AMD provides a means to undervolt w/o a bios mod though with ease so it's far safer to test w/o running into a complicated mess to fix.
 
Last edited:
For undervolting to be providing any raw performance benefit means to me that AMD shipped the Radeon VII with a too low max TDP.

Great article as always, and surprising results.

I always assumed undervolting would reduce performance, but I see that is not the case.

I've been having a lot of fun with overclocking the Radeon VII. Wattman has a ton of convenient options and it makes it easy to try things with minimal hassle.

As similar to the review, I also did not have much luck with Chill. I found it reduced performance too much and I'm usually on the edge (4K, etc.) and don't have the performance to spare.

One question: did you compare simply setting an fps cap in software (for example in RivaTuner) and would this result in similar savings?

Yeah, as explained in the article (and for which I have some first hand experience) when the board or bios hits max TDP then it will throttle regardless of temperature because the card will refuse to allocate any more power to the GPU and RAM.

Interesting article. Would using a waterblock in a custom loop increase performance or just reduce temps for undervolting?

I found that a waterblock can lead to further GPU clock headroom not just from lower temps, but because there is a little extra power available before you hit TDP because you aren't powering the fans anymore. It's not much more MHz, but every little bit helps.
 
^ You can probably undervolt and marginally overclock some aspect of virtually every GPU released since a damn Voodoo GPU with a bios mod. This is hardly new or specific to AMD, but nice try.
 
^ You can probably undervolt and marginally overclock some aspect of virtually every GPU released since a damn Voodoo GPU with a bios mod. This is hardly new or specific to AMD, but nice try.

Who said this is an AMD exclusive thing? Anyway, while you MAY be able to undervolt GPUs from the Voodoo era, they won't clock any higher automatically, like more recent gpus do :) (and not all gpus were voltage modable)
 
Great follow up guys! Now we just need some damn Waterblocks for this card and things should get interesting
 
Interesting article. Thanks!

Chill seems like mostly a fail to me. Huge hit to performance, while most of the time only having a marginal impact on temps, and sometimes even getting hotter.

The way to go seems to be the undervolt. It behaved exactly as expected, which was good to see. Slightly lower max clocks, but much better average clocks. Sure it got hotter most of the time, but that's because you are also getting higher frame-rates. If you really need the temp lower, maybe couple undervolting with a frame limiter or vsync.

It would be interesting to see how this picture changes if you put a waterblock on it.
 
Very interesting review. These Radeon VII cards are looking like a really solid card. As AMD irons out some driver/bios kinks and tweaks are applied to the cards (undervolting) these cards are looking more and more competitive to the 2080/1080ti.
 
Nice analysis Brent , but can you explain something because otherwise i'm unable to understand :
since undervolting can result all these that you mentioned above (lesser power , same performance) ,then why exactly AMD doesn't sell this card with lesser voltage at default :eek:???
THERE MUST be some downside ,otherwise this means that AMD 's engineers sabotage their own GPU ,AND their customers as well , by making them pay more money at electrical bills for no reason whatsoever:wacky: !!!
(*i don't believe that AMD 's engineers are idiots , so there must be a reason for selling this card with these default settings. For example , when undervolting the components of the GPU , could this affect their longevity negative or positive? There must be some explanation which justifies their decision about the amount of voltage that they are using as default :confused:!!! )

Much akin to leaning out the A/F ratio in your car. Or adding more boost w/o more fuel. Same thing. You can make more power but if ANYTHING happens, bad fuel, over boost, etc you're going to blow the motor.

Remember them saying about lowering voltage whilst ensuring games still run? That's why AMD doesn't undervolt. They want to make sure the card is as widely compatible with the most games, with the fewest issues.

This article is a game changer for me as well. I would have never thought to undervolt. I'm going to start playing around with my Vega 64.
 
Undervolt and overclock test!!! wonder if you can set it to say 1025 and get a few more MHz on the core etc.
Yes, it's possible. I have more stability tests to do, but I'm currently at 1020 mV and 1880 MHz on core and it seems good (tested about 5 games, 30 minutes each).

At these settings I'm getting a nice boost in performance, around 5 to 10% improvement, the same as with 2000 MHz on core but at much lower voltage (at 2000 I had to pump the voltage way too high, around 1.2, and it still had some stability issues).

I can probably tweak more, but I'm happy with the performance right now so I'll probably just game with it for a bit and make sure it's solid over time.
 
Last edited:
Hahahahaha @ Brent_Justice :) Thank you dude!

great updated summary quote :)
and i quote " We have done further testing to further identify why GPU temps increase while undervolting. Though we did not touch the fan speed, the culprit seems to be the fan speed. "

Thanks dude, will continue reading now:)


Edit :

So, with MSI Afterburner, using the default fan curve settings; normalised through zero, reduces my idle temp ( 40`c @ default - fans off ) to just above ambient = ~ 29`c for 200-350rpm. This gives rise to gaming senarios ~ 60`c+ ( +- 5-8`c depending on ambients )

Worth investigating if Wattman ( :-( ) not playing nice!

Thanks for update, interesting result, the auto reduction in fan-speed rpm!!!!

Edit : spelling + clarity.
 
Last edited:
Who said this is an AMD exclusive thing? Anyway, while you MAY be able to undervolt GPUs from the Voodoo era, they won't clock any higher automatically, like more recent gpus do :) (and not all gpus were voltage modable)
There was a bit of a implied AMD jab and not as if Nvidia doesn't also do similar, but sure they have more resources to do a better job at squeezing out more performance and efficiency no kidding that's readily obvious. There is a point where it's simply not worth the time and effort for either team to test for every single chip they produce at least w/o A.I. doing all the leg work. It's defiantly not worth it beyond a certain point for humans to do all that stuff awhile still offering a good product at a good price point to the end user. I never said all GPU's, but I do think the vast majority you can do some form of a undervolt and/or tiny overclock with a bios mod to reduce power or increase performance or a mix of both. The Q9550 vs Q9550S is a nice example case on the CPU side the bulk of Q9550's were able to undervolt to exactly the same vcore of the energy efficient model that was primarily the same with some small notable differences in spots. You make a fair point on the boost subject though it's another matter separate from TDP though I guess you can say the can intertwine a bit. You can bake in voltages with p-states for min/max frequencies though if you've time to test for and the results are really more ideal done right, but boost is a ease of use type of solution where again there can be a bit of wiggle room to be had one way or the other in terms of power efficiency or frequency performance and mixed with temperature and power targets I guess. These things get more complex to handle the more intricate they become. They aren't going to carefully and perfectly every single CLK frequency range to voltage to absolute performance/efficiency/stability realistically speaking by hand though maybe the Skynet overlord's will for world dominance.

Interesting article. Thanks!

Chill seems like mostly a fail to me. Huge hit to performance, while most of the time only having a marginal impact on temps, and sometimes even getting hotter.

The way to go seems to be the undervolt. It behaved exactly as expected, which was good to see. Slightly lower max clocks, but much better average clocks. Sure it got hotter most of the time, but that's because you are also getting higher frame-rates. If you really need the temp lower, maybe couple undervolting with a frame limiter or vsync.

It would be interesting to see how this picture changes if you put a waterblock on it.
They only seem to have tested default chill settings far as I can tell so I don't know how accurate that assessment really is. Having a option isn't really a fail. The results on the Wolfenstein benchmark for chill were intriguing it appears like it would exhibit very minor jarring micro stutter in that scenario because it had really small jitter. If they set it to 70/144 min/max that seems to align well with that and show's it's potential upside to keep frame rate consistency smooth and linear. Think of chill like a min/max frame rate cap now provided boost can still work within that it could even be beneficial to bumping up min/average frame rate with less complication compared to manually baking in voltage to frequency in regard to temperature and power limits.

*UPDATE*

We have made an update to the conclusion regarding GPU Temp / Fan Speed, please check it out right above The Bottom Line

https://www.hardocp.com/article/2019/02/24/amd_radeon_vii_chilling_undervolting/6
This falls in line with my GTX960 undervolting experiences. As the voltage changes I happen to notice the fan curve RPM can skew one way or the other. You can basically raise or lower the the temperature target to compensate to readjust the RPM. The whole fan curve to temperature is a complexity a visual GUI graph display of how it all ties together like motherboard bios have for fan control is extremely helpful. That's a area a bios mod tool like BIOS Tweaker could improve some to be a bit more user friendly.

"The drawback is increased fan noise doing this. That means undervolting does actually reduce fan noise by way of reducing fan RPM, but also increases GPU temp. The overall best way is going to be manually customizing voltage plus fan speed for your preferred settings." Once you readjust the fan curve you can get a more gradual temperature rise and thus a lower average RPM over time from a undervolt. I found that I preferred to reduce the fan temperature targets on curve points to compensate for the RPM. You can alternatively bump up the RPM instead, but then it ends up loader than before. You could not adjust it at all, but then idle up thru full load temperatures will all be higher from reduced RPM's. The most concerning one is really the highest temperature fan curve point since it's the one most tied to throttling and overheating related significance.
 
Last edited:
Just so people know. The stock voltage of this GPU is not a static number which is unlike Vega 56/64. The stock voltage is depends on the quality of the chip. Mine is 1052mv. We have seen some as low as 1032mv and some as high as 1.1v. Maybe the story could clarify this. Since my Radeon VII has a lower stock voltage it would likely perform better than the test GPU in this article.

Edit: also if you update the BIOS to Version 106 that AMD just release it resolves some of the fan curve issues. It spends much more time in the 900-2000rpm range now which is tolerable. AMD has released both v105 and v106 in about 1 weeks time at the same web link.
 
Just so people know. The stock voltage of this GPU is not a static number which is unlike Vega 56/64. The stock voltage is depends on the quality of the chip. Mine is 1052mv. We have seen some as low as 1032mv and some as high as 1.1v. Maybe the story could clarify this. Since my Radeon VII has a lower stock voltage it would likely perform better than the test GPU in this article.

Edit: also if you update the BIOS to Version 106 that AMD just release it resolves some of the fan curve issues. It spends much more time in the 900-2000rpm range now which is tolerable.

Please test :), data will be invaluable to dicussion :)
 
Very interesting results, kinda makes you wonder why AMD didn't set the default voltages lower.
 
Please test :), data will be invaluable to dicussion :)

I wish I had the time. Its quite a noticeable difference from the stock bios and v105 which fixed the UEFI/secureboot issue. I can say /AMD at reddit has some people confirming the same thing.

Edit: I raised a ticket with AMD for the changes.
 
Last edited:
Just so people know. The stock voltage of this GPU is not a static number which is unlike Vega 56/64. The stock voltage is depends on the quality of the chip. Mine is 1052mv. We have seen some as low as 1032mv and some as high as 1.1v. Maybe the story could clarify this. Since my Radeon VII has a lower stock voltage it would likely perform better than the test GPU in this article.
I encountered CPUs and gfx cards that are good at lower volts but dont do well at higher volts.
They had a max Hz limit which was very difficult to go beyond.
I'm interested if recent fab methods have changed anything and whether your overclock results are better.
 
Has anyone tried compute heavy loads while undervolting? I'm curious to see if they require more voltage than gaming.

I'm also curious along these lines dude; .............when playing around with mining, i found on both R9 380 + RX 580 that; maximum efficiency gave me the best return; just was never willing to leave PC on, with significant Gpu noise all day/night :-(

But with undervolt + Chill + Fan tweaks etc.........:), compute heavy workloads could get a serious boost...... maybe by many times in the efficiency sweet spot : interesting :)

My R5 2600 @ 3.400GHz to 3.800Ghz @ 1.100v F.L.F stable...... ( with vcurve like this maybe, but for Gpu? )

Edit : spelling.
 
I wish I had the time. Its quite a noticeable difference from the stock bios and v105 which fixed the UEFI/secureboot issue. I can say /AMD at reddit has some people confirming the same thing.

Edit: I raised a ticket with AMD for the changes.

Cool, thank you :)
 
*UPDATE*

We have made an update to the conclusion regarding GPU Temp / Fan Speed, please check it out right above The Bottom Line

https://www.hardocp.com/article/2019/02/24/amd_radeon_vii_chilling_undervolting/6
Hi, Brent. On my system I seem to be getting slightly better performance (by about 2 - 3fps) with fan on default.

If I set a custom fan curve, I can audibly hear the fans spinning up, but then performance is ever so slightly less.

Any idea why that is?
 
Last edited:
Seems like the fan speed algorithm heavily relies on Vcore. Great work Brent. Playing around with Vcore for any of the GCN derivatives is definitely a labor intensive step, but worth it for enthusiasts.
 
Still really keen to know if the undervolt combined with chill can work in tandem and if so how that might alter results a touch plus I'm curious how straying from the default chill min/max settings could be of a benefit. Hows it tie in with boost in relation to the minimum/average frame rate in correspondence to chill's min/max frame settings with setting the chill min higher and max lower? Can you get a bit better average minimums doing that with chill?
 
So looks like there is more fan fixes for Radeon VII. Figured I would add this since it was a part of the article testing and fan noise has been my biggest gripe.

Radeon Software Adrenalin 2019 Edition 19.2.3 Highlights

Fixed Issues
  • Battlefield™ V players may experience character outlines stuck on screen after being revived.
  • Fan speeds may remain elevated for longer periods than expected when using Tuning Control Auto Overclock or manual fan curve in Radeon WattMan on AMD Radeon VII.
  • ReLive wireless VR may experience an application crash or hang during extended periods of play.
  • Zero RPM will correctly disable in Radeon WattMan on available system configurations when manual fan curve is enabled.
  • A loss of video may be intermittently experienced when launching a fullscreen player application with Radeon FreeSync enabled.
 
So looks like there is more fan fixes for Radeon VII. Figured I would add this since it was a part of the article testing and fan noise has been my biggest gripe.

Radeon Software Adrenalin 2019 Edition 19.2.3 Highlights

Fixed Issues
  • Battlefield™ V players may experience character outlines stuck on screen after being revived.
  • Fan speeds may remain elevated for longer periods than expected when using Tuning Control Auto Overclock or manual fan curve in Radeon WattMan on AMD Radeon VII.
  • ReLive wireless VR may experience an application crash or hang during extended periods of play.
  • Zero RPM will correctly disable in Radeon WattMan on available system configurations when manual fan curve is enabled.
  • A loss of video may be intermittently experienced when launching a fullscreen player application with Radeon FreeSync enabled.
Cool, been receiving loss of video periodically, will have to upgrade to 19.2.3.
 
Cool, been receiving loss of video periodically, will have to upgrade to 19.2.3.

Are you undervolting / overclocking Gpu ( have had such display anomalies with both senarios!), or do you have enough NB/SocV to ensure vidoe-out / display?, or maybe the Cpu needs a nudge, or DDR on V, to stabilise system-wide?
 
Has anyone tried compute heavy loads while undervolting? I'm curious to see if they require more voltage than gaming.
In general even less voltage, your not using the full GPU as in ROPs which can generate a lot of heat. Then again that was with old Vega, new Vega?

A+ article, also fun and different then the endless mundane ones that all copy each other. For original OCing Vega, temperature made very huge differences in how one can push things, wondering if that follows suit with Vega VII. Look very much forward to Kyle's experience with water cooling here.
 
Are you undervolting / overclocking Gpu ( have had such display anomalies with both senarios!), or do you have enough NB/SocV to ensure vidoe-out / display?, or maybe the Cpu needs a nudge, or DDR on V, to stabilise system-wide?
Vega's FE's not V, sorry. Just was replying to new driver info.
 
EK stated blocks would be ready in a months time, not sure what you guys had planned, but would be great if you can swing it! If not, all good, someone will put something up out there. Appreciate the time invested! behaves similarly to OG Vega, I wonder if soft powerplay table mods will boost performance further once properly cooled
 
So looks like there is more fan fixes for Radeon VII. Figured I would add this since it was a part of the article testing and fan noise has been my biggest gripe.

Radeon Software Adrenalin 2019 Edition 19.2.3 Highlights

Fixed Issues
  • Battlefield™ V players may experience character outlines stuck on screen after being revived.
  • Fan speeds may remain elevated for longer periods than expected when using Tuning Control Auto Overclock or manual fan curve in Radeon WattMan on AMD Radeon VII.
  • ReLive wireless VR may experience an application crash or hang during extended periods of play.
  • Zero RPM will correctly disable in Radeon WattMan on available system configurations when manual fan curve is enabled.
  • A loss of video may be intermittently experienced when launching a fullscreen player application with Radeon FreeSync enabled.

And yet still major known issues which can directly affect our testing and your testing, not happy about that myself since this is what we are trying to test:

Known Issues
  • Changes made in Radeon WattMan settings via Radeon Overlay may sometimes not save or take effect once Radeon Overlay is closed.
  • Radeon WattMan settings changes may intermittently not apply on AMD Radeon VII.
  • Performance metrics overlay and Radeon WattMan gauges may experience inaccurate fluctuating readings on AMD Radeon VII.

They have got to fix the Wattman issues, it's always been very buggy
 
The fan issue is not fixed for me. Once a custom curve is set, after the fan ramps up it won't go back down (even after exiting the game, I have to restart my computer).

I've settled on just using the default fan profile for now until the bugs are sorted out, and it seems to work well, but definitely would like a custom curve if possible.

Also, do we know when Radeon VII clock/temp etc. readings will be available in Afterburner? I know there is a AMD OSD but things would be much easier if AB worked.
 
HBM voltages can also be tweaked or adjusted. Since HBM2 should be similar to original Vega HBM2 where OCing can be greatly increased if temperature is less than 60c, preferably less than 55c. Undervolting HBM2 (cooler) also helps in OCing them which for Vega made a big difference both in bandwidth and latency. All HBM2 are not the same, my VegaLC UV HBM2 is very limited, while the Vega FE's are not.
 
Protips:

-Cooler base is not very flat, Vega 7 Die may not be very flat as well.

-Add washers to rear to increase cooler pressure on die and get better temperatures.

-Powerplay tables work for Vega 7
 
Really didn't think about that. Reminds me of the old Athlon days, bare naked dies cracking in the corners.

Its not the die that cracks in this case, its the substrate or the internal HBM layers. Realistically not much of an issue unless you're going LN2, otherwise any upgraded air or water cooler would put you at risk with Vega/2 as any cooler could cause "incorrect" clamping force.
 
Its not the die that cracks in this case, its the substrate or the internal HBM layers. Realistically not much of an issue unless you're going LN2, otherwise any upgraded air or water cooler would put you at risk with Vega/2 as any cooler could cause "incorrect" clamping force.
Yep, you will never see it.
 
Hah, I was just about to come in here and post about how you can't lose 40 watts of power and have it "make sense" that the GPU is hotter just because it is running at higher clock speeds. The thermal solution hasn't changed. But, then I got to the updated conclusion and it all came together.
 
Back
Top