Modding a 2x 140 mm All-in-One CPU Watercooler onto the AMD Radeon VII

[W1zzard]

Reddit user CarbonFireOC wasn't happy with the temperatures he's been seeing on his Radeon VII, and decided he wanted watercooling. Unfortunately, that early since product release no GPU waterblocks are available for the Radeon VII. What he figured out was that the waterblock of EVGA's CLC 280 all-in-one cooler will fit. This cooler, which is made by Asetek and licensed by EVGA is designed for Intel & AMD motherboards and many similar variants exist from vendors like Fractal Design, NZXT and others.

While performing such a mod on a $700 graphics card is not for the faint-hearted, it shows that with a little bit of creative ingenuity you can achieve amazing results without wasting top dollar.

His 24/7 stable settings on Radeon VII are 2122 MHz core, 1265 mV, and 1252 MHz memory, resulting in a 3DMark Firestrike graphics score of around 33,000. Even at such a high voltage, the watercooling keeps temperatures very low at 40°C GPU, 60°C Hotspot.





View at TechPowerUp Main Site

 

[londiste]

1265mV sounds rather high.
What's the power consumption like? VRM temp? Hotspot temp?

Logging temps in Battlefield 5, my hotspot is around 40C and temps in the 30C range.

I was able to run a 2200Mhz 1310mV no problem. Just for giggles, I managed to play a round of BF5 at 2270Mhz.

Now we have hit a package power limit of sorts, they don't go much above 40C GPU and 60C hotspot during a bench run. It will hit 70C hotspot under a stress test though and needs another 50mV to stabilize it.

This is a really good question, I don't have a meter on. We have hit a possible hardware power limit with the card. I can only guess its around 500W. The registry powerplay table is set to allow 1000W, so I don't think the limit is there.

 

[ZoneDymo]

I actually messaged Gamersnexus to try to do just this (well as soon as the new drivers are out so we could try some OCing as well)

 

[londiste]

The guy suspects (but has not measured) that card uses 500W if not more. This is twice the power consumption for extra 300MHz. This is pretty far on the steep upwards efficiency curve.

AMD cards are much more overclock friendly than Nvidia cards, there are simply far less limits. Any Turing or Pascal card has a rather limited Power Limit range and basically a hardcoded Voltage limit around 1.09V. This will not get you far and subverting this is not trivial.

 

[Deleted member 172152]

My solution is much simpler: Higher case fan speeds since they are really quiet even at max speed, plus putting my leftover fan in the bottom-rear of my case blowing air towards where my Radeon VII will be. I'm expecting it to warm up slower and maybe keep slightly lower fan speeds, which would be nice since historically the gpu always ends up being the loudest component in my system! XD

 

[Flyordie]

So basically... AMD should have done an AIO model of this card just like they did with Vega64.

Point proven.

 

[Deleted member 172152]

So basically... AMD should have done an AIO model of this card just like they did with Vega64.

Point proven.

Maybe, but the clockspeeds aren't worth the power consumption. 1900MHz plus low temps would have been sweet though.

 

[Vayra86]

So... great. 7-10% extra performance at the power draw of 2x2080ti.

Bonus points for the modding effort of course, but still, painful.

 

[Flyordie]

Maybe, but the clockspeeds aren't worth the power consumption. 1900MHz plus low temps would have been sweet though.

Eh, power consumption for the high end cards are meh. My RX Vega64 runs 1800Mhz just fine and while running F@H only uses about 190-200W while doing that. Gaming @ 1800Mhz uses about 260-270W.

As for your comment directly- I agree there too. Plus it'd give you the ability to vent that heat right outside the case instead of keeping it tucked inside the case.

They can do an AIO version. Vega64 and Fury X proved it. I doubt any other AIB will do it. So AMD needs to step to the plate and actually do it.

 

[phill]

Think this is great, I just wish they would lower the volts down a bit...

Wonder if when they release their brand new GPU, they will surprise us all.... We can but dream

 

[EarthDog]

Love to see this! Cant wait for additional oob solutions to arrive.

 

[siluro818]

1.25GHz HBM2 - me wants....

 

[XXL_AI]

why do you mod? because it sucks the power like an thirsty elephant in africa. just wait amd to make low power version, you won't die*



*I can't promise because all of the amd hardware sucks more power from the powersupply

 

[the54thvoid]

Hats off to the guy. Modding is its own thing - It's not about being the best, it's about being creative. And this is creative.

The power draw is an irrelevance as it's not an everyday thing. All modders push the limits somewhere - it's a bit naive to harp on about power when you're dissecting an expensive gfx card.

 

[Enterprise24]

33K FS @ 2122 is very impressive. My 1080 Ti watercooled + XOC @ 2190 / 1600 is just a bit faster at 34.1K
If he try to bench at 2270 1.31V he will beat heavily OC 1080 Ti for sure. Not counting that performance from future drivers will widen the gap further.
Looks like there is tons of OC potential on Radeon 7 if you can properly cool it. This remind me of Kepler days where you can squeeze 40-50% more performance from selected GK110 with watercooling.

 

[Jism]

The guy suspects (but has not measured) that card uses 500W if not more. This is twice the power consumption for extra 300MHz. This is pretty far on the steep upwards efficiency curve.

AMD cards are much more overclock friendly than Nvidia cards, there are simply far less limits. Any Turing or Pascal card has a rather limited Power Limit range and basically a hardcoded Voltage limit around 1.09V. This will not get you far and subverting this is not trivial.

It's not a voltage, but current limit, on Nvidia cards. We dont know the real reason behind it but my best bet would be that the proces is not suitable for high current and it could burn the chip out (to put it simply).

why do you mod? because it sucks the power like an thirsty elephant in africa. just wait amd to make low power version, you won't die*



*I can't promise because all of the amd hardware sucks more power from the powersupply

It really comes down to the workload. Amd has a different approach then Nvidia and the downside is that it consumes a bit more power. A TDP is simply a outer limit the card could use. Does'nt mean when your gaming it's doing the same. However, you have so many tools these days to make a card more efficient. Click on Chill. Click a Vsync. And before you know your 180W TDP AMD Card is doing all of a sudden barely 100W and still spits out 75 frames per second with low frame times.

 

[londiste]

It's not a voltage, but current limit, on Nvidia cards. We dont know the real reason behind it but my best bet would be that the proces is not suitable for high current and it could burn the chip out (to put it simply).

It is both. There are Temperature, Power and Voltage limit bits which do get triggered separately.

A TDP is simply a outer limit the card could use. Does'nt mean when your gaming it's doing the same. However, you have so many tools these days to make a card more efficient. Click on Chill. Click a Vsync. And before you know your 180W TDP AMD Card is doing all of a sudden barely 100W and still spits out 75 frames per second with low frame times.

That depends on point of view. Card going for the max possible frequency while staying under TDP should be the preferred approach for power and clock management. Nvidia is currently doing it better than AMD.
The few times I have tried Chill, it tends to either introduce judder/stutter when altering moving speed or does not give a large benefit. It is cool tech but not one I would prefer to use during gaming.

 

[Jism]

I really have no issues with that. Set the FPS limit like a few FPS above your monitor's refreshrate, i.e 78 or 80 on a 75hz panel. This way it renders really smooth in my opinion.

 

[londiste]

Maximum FPS limit is not the problem and is not that interesting. This is effectively just an FPS limit.
Minimum FPS limit is the cool tech part and tends to be problematic for me. This lowers FPS when there is not much going on on screen.

 

[XXL_AI]

It's not a voltage, but current limit, on Nvidia cards. We dont know the real reason behind it but my best bet would be that the proces is not suitable for high current and it could burn the chip out (to put it simply).



It really comes down to the workload. Amd has a different approach then Nvidia and the downside is that it consumes a bit more power. A TDP is simply a outer limit the card could use. Does'nt mean when your gaming it's doing the same. However, you have so many tools these days to make a card more efficient. Click on Chill. Click a Vsync. And before you know your 180W TDP AMD Card is doing all of a sudden barely 100W and still spits out 75 frames per second with low frame times.

a bit more? if you use nvidia-smi to decrease the power limit to the half, nvidia is still faster than 2x power consuming amd cards.

 

[HwGeek]

33K FS @ 2122 is very impressive. My 1080 Ti watercooled + XOC @ 2190 / 1600 is just a bit faster at 34.1K
If he try to bench at 2270 1.31V he will beat heavily OC 1080 Ti for sure. Not counting that performance from future drivers will widen the gap further.
Looks like there is tons of OC potential on Radeon 7 if you can properly cool it. This remind me of Kepler days where you can squeeze 40-50% more performance from selected GK110 with watercooling.

Since the PCB is on the high end- AIB partners just need to release Waterblock model or take their existing massive Air cooler and mount on VII, I think they can sell very well for $749 a card with OC potential of 2200Mhz core/1200MHz Memory with good normal noise levels.
We need to remember that RTX2080 FE runs ~1900Mhz AVG while gaming, so Radeon VII @1900MHz/1050Mhz will perform very nice without OC.

 

[londiste]

We do not know what the power consumption at 2200MHz is. I seriously doubt a massive air cooler would be sufficient.

 

[John Naylor]

AMD cards are much more overclock friendly than Nvidia cards, there are simply far less limits. Any Turing or Pascal card has a rather limited Power Limit range and basically a hardcoded Voltage limit around 1.09V. This will not get you far and subverting this is not trivial.

In what respect ? What I mean is, are we taking core and memory clocks or are we talking results in fps ? From what I gave seen in TPus testing, of seven "gamer model" 2080 Tis, the highest core or highest memory clocks do not correlate with the highest fps . And if that correlation isn't there, then what conclusions can be drawn from greater core / memory OC ability ?

The reason I asked is that there's one significant story that I haven't seen in print as yet regarding what we are seeing these days regarding the fps gains obtained in overclocking GFX cards. Starting with the 2xx series, AMD stared aggressively overclocking their cards in the box. When the 290x came out, it was designated the new king ... it lost the throne to the 780 Ti a week later. However, when both cards were manually overclocked, the 780 retained 2nd place and the 290x slipped to third. With the 980 Ti and Fury X looked competitive outta the box, but the 980 Ti OC'd 27-30% fps improvements above reference and the Fury X was in single digits. The 10xx series wasn't quite as OC friendly but still a large gap between the two camps.

With the AIB 2xxx series cards, we are seeing much lower OCs both in how much more we get outta the box and how much more we get above reference. The factory OC'd cards are gaining around 8% above their "in the box" condition, same as the Radeon VII. We are also seeing the FE "A series" GPUs being used as the new "reference". Compared to **that** reference, the AIB's typical gamer cards have ranged from 112.1% OC (EVGA X Ultra) to 116.5% (MSI Gaming X Trio). I couldn't help wondering whether with 2xxx, they mighta tweaked up the OCs a bit more than in the past given that the 1xxx to 2xxx improvements generation to generation were less impressive,

Another thing that I found intersting was that in the OC test, the 8 cards tested, 7 fell into vendor's gamer line and 1 in the premium, "bawlz to the wall" card. The 7 ranged rather tightly, from 218.0 fps OCd to 226.6 .... however the premium card (Lightning Z) jumped to 236.7. I'm curious as to what's happening here, recent iterations of the Lightening, Classified and Matrix had fallen well short of expectations and Im curious as to what improvement(s) gets credit for that jump. It is wporth noting that all 7 "gamer' cards tested has Micron Memory (T61K256M32JE-14:A) which ranged from 2005 to 2065 OC'd (EVGA was only 1960) and the Lightnng has Samsung (K4Z80325BC-HC14) which hit 2090. The Lightning had an OC'd core tho of only 2100, Only 2 cards were lowewr. The MSI Trio was 2085 and outside the Lightning had the highest OC'd fps; The EVGA X Utra was 1960 and finished last other than the FE .

 

[EarthDog]

In what respect ? What I mean is, are we taking core and memory clocks or are we talking results in fps ? From what I gave seen in TPus testing, of seven "gamer model" 2080 Tis, the highest core or highest memory clocks do not correlate with the highest fps . And if that correlation isn't there, then what conclusions can be drawn from greater core / memory OC ability ?

In many respects... but what he alluded to was that AMD cards don't really have a 'hard' power limit like NVIDIA cards do. So, overclocking isn't limited as it is with NVIDIA cards where getting around the power and voltage limits are hard mods, not software changes.

Part of the reason you are seeing various results with different (slightly less) clocks is because something isn't stable (memory ECC is taking over for example) or it is dropping bins and the reviewer is reporting a peak clock or a clock that isn't sustained due to power limits or temperatures which will obviously vary with card as each has a different power limit set for its model.

It has been well know for generations that overclocking is more of a silicon lottery than 'card level' as to what overclocks more.

 

[londiste]

For an nvidia card, power limit is usually not the worst problem, voltage limit is. Every Pascal and Turing card I've had is power limited at stock. Somewhere at 10-15% additional power limit they start becoming voltage limited. Even though you can increase power limit for example 30% you will not the full benefit as voltage limit cuts you down first.

AMD has power limit that can be subverted but not voltage limit as such. It least I do not remember Vega having one and temperature/power is what remains the important limit.

With the AIB 2xxx series cards, we are seeing much lower OCs both in how much more we get outta the box and how much more we get above reference. The factory OC'd cards are gaining around 8% above their "in the box" condition, same as the Radeon VII. We are also seeing the FE "A series" GPUs being used as the new "reference". Compared to **that** reference, the AIB's typical gamer cards have ranged from 112.1% OC (EVGA X Ultra) to 116.5% (MSI Gaming X Trio). I couldn't help wondering whether with 2xxx, they mighta tweaked up the OCs a bit more than in the past given that the 1xxx to 2xxx improvements generation to generation were less impressive,

When it comes to Nvidia, Boost 3.0 and Boost 4.0 are a thing. Reference clocks are far below any clock a Pascal or Turing card does with reasonable cooler on it. To put it simply, cards will increase clocks until they hit one of the limits - temperature, power, voltage. Well, there is a max number of bins it can boost but that is generally enough to reach to ~1900MHz and something else will limit it first.
AMD has a boost system that is essentially similar.