3900x/Radeon VII watercooling: another newbie thread

[aigomorla]

sigh...
TEC's are energy hogs... not really worth it for the minimal gains you get and also the high risk of condensation, and having you coat or insulate your board.

Also just chilling the cpu wont really net you that much anymore without chilling the ram and ramping that up.
So its more of a dog chasing his tail since ram performance is also proportional to overclock performance.

 

[DrMrLordX]

sigh...
TEC's are energy hogs... not really worth it for the minimal gains you get and also the high risk of condensation, and having you coat or insulate your board.

Hence the need for temp sensors and voltage controllers. Temps start going below ambient? Back off the voltage, let things heat up. Running a TEC "balls out" is going to produce some sub-ambient temps which is one of the reasons why just throwing an eBay TEC in there and calling it a day is a generally terrible idea.

Also just chilling the cpu wont really net you that much anymore without chilling the ram and ramping that up.

That's . . . actually not true. Take a look at der8auer's LN2 charts for a relationship between clockspeed and voltage. I think his tested sample was pretty bad (I can beat his curve on water, up to about 4.4 GHz), but the basic idea is that Matisse is extremely sensitive to temp. It really needs to be kept as close to ambient as possible in the event that the end user does not want to deal with sub-ambient temps. Plus the L3 on these chips is so big that RAM performance matters less than it did on Summit Ridge and Pinnacle Ridge (depending on the application). DDR4-3733 is attainable, and it seems plenty fast enough. What the 3900x crowd wants/needs are all-core OCs in the 4.6-4.7 GHz range. Or at least 4.5 GHz in CBR20.

I was thinking that a traditional "TEC sandwich" setup would be misguided. Instead, someone with manufacturing cred should produce an IHS replacement with miniature clusters of thermocouples wired in parallel, situated on top of the chiplets. With the right control circuitry and sensors, they might be able to rig the thing up to ramp up voltage to specific thermocouple clusters in order to address hotspots on the CPU, and otherwise just keep the dice at or near ambient 100% of the time. Yeah, it would suck power, but short of DICE; LN2; phase; or other sub-ambient solutions that do require significant preparation (insulation/grease), I see no real alternative to fully waking up these chips. Putting them under a big rad isn't entirely cutting it. But I'm getting a bit off my own topic, so I'll leave the musing where it is.