AWD-IT 465X Ryzen 9 3900

[HEXUS]

AMD's 12-core 65W CPU gets an outing.

Read more.

[edmundhonda]

£2000 and they don't bother setting the fans up?

Pull the other one.

[philehidiot]

I've just slapped together my AMD 3900X system with a Corsair cooler like this. I just finished troubleshooting the few little niggles and I'm about to move on to BIOS set up.

The fan speed was so high (cooler moved off an Intel system where it was extremely quiet) I thought I'd screwed up my CPU / heat sink mounting.

The behaviour is that it sits at low temperatures and quietish but the tiniest little thing (even opening a browser) will cause temperatures to spike quickly and the fans to ramp up, where the temperature quickly is brought under control. The stress test results in a peak temperature of 69C which is excellent for the chip and I can't see my mounting / TIM stuff being a problem.

What I suspect is happening is the Corsair is trying to be really quiet and is configured with the Intel chips in mind, which have very restrained boosting behaviour. Who, before Ryzen, would have bought a cheap, slow AMD chip and then liquid cooled it? These profiles have probably been refined on Intel chips over the years.

My solution is likely going to be to increase the base fan speed slightly, allowing more baseline cooling for the variable clock speed which can respond very quickly but not as fast as the liquid cooler. There's a response time from increasing fan speed to the cooler liquid hitting the CPU and I suspect the very variable and responsive AMD boosting system is something for which AIOs need properly optimising.

[philehidiot]

Fan speed problem is an interaction on my moobo (may be relevant here) is between the "smart fan" tech where it tries to ramp up and down the fans to the smallest change in temperature. The Ryzen master CPU temp data is clearly using greater averaging than the mobo direct data as, whilst the refresh rate is the same, the peaks and troughs are lower. I expect there are a couple of things going on:
1) The smart fan tech is trying to make immediate changes to the fan status where there is a disparity between the speed of instruction and the transmission of actual liquid. I think it gets itself in a tiz as it ramps down, then the hot fluid arrives, it ramps up, so it cools down. The CPU sensor and radiator state are assumed to be in the same thermal state. They are not.
2) The rapid and highly unpredictable (from the POV of the fan controllers) temperature of the CPU is very different from previous gen performance systems. Intel had a Tau duration and then couldn't boost again for a couple of minutes. This made boosting predictable whereas XFR and so on makes the thermal output of Ryzen, for all intents and purposes, random (linked to workload and CPU thermal state). This means that even with the allowance of 5C delta before changing fan speeds, you're going to get rapidly variable temperatures.
3) Smart fan is varying both pump and fans together where they should probably be controlled independently for any kind of "intelligent" functioning.
4) The mobos appear to be seriously over volting Ryzen. Mine was at ~1.44V. Insane. The chips will often run at 1.3V apparently and 1.35V is a safe bet. Mine is at 1.44V at idle. In fact, all my default voltages appear very off the mark. XMP says my DRAM should be at 1.35V for current settings, but it's 1.2V. There's some weirdness with the IF interconnect voltages that I haven't worked out and load line calibration seems very wappy. This is all stuff I've yet to get my head around but is conspiring to throw more power and heat at the sillycone. At idle my CPU is consuming ~35W which is very high in my book. I have a properly set up Vega64 using ~5W at idle in the Windows 10 environment. In laptops, this architecture can do way more than my system is doing at idle in a sub 25W envelope.

The solution for me is to mimic the averaging effect of the Ryzen software in that the fans need to be told to ignore temperature changes below 65C and to ramp up like the proverbial mofo when 70C is breached.

To tune this smart fan stuff needs to be turned off and the CPU and CPU_opt (or whichever system fan header the AIO pump is connected to) set to RPM fixed mode. My fan controller gives the option of a fixed RPM below a certain temp and full poooooowwwaaarrrr above a certain temp. I chose 70C as my PC is usually ticking over or at full chooch, with no in between. To choose a tickover RPM I turned the fan chooch down notch by notch until the temperature at idle started to rise. I then went 10% higher and that's my idle setting. I did the same with the fans and the pump speed. I also tried turning off the pump to see what happens to the temperature and how fast. Small delay and it rises fast and harrrdddd.

End result for me? 40% pump and 40% fan speed at tickover for casual use with temperatures of 42-44C. I expect when I finally wrestle the CPU voltages off auto, those temperatures will sit in the 30s.

The way to differentiate this between a pump failure or poor TIM / heat sink application is to either turn off the pump and see what happens (a failed pump will result in no change in temps when its RPM is put to zero), turn or down the fans and turn up the pump and listen. Don't do these for long periods, duh - a few seconds whilst keeping an eye on temps is all that's needed. And finally a stress test. If there's a problem with your TIM application or your spreader-sink connection in some way, a stress test will boil and throttle your CPU, possibly resulting in a shutdown.... eyeball temps with a finger on the "stop" key. Mine topped out at 69C whilst stressing all 12 cores but I think I got lucky in the sillycone lottery if that's right. With my tweaks it topped out at 75C. Not sure why the different but it may have been that this took me all day to figure out and it was warmer in here by the end of it.

So, to be fair to these system builders, they may not have been faced with just a fan set up problem, but also a mobo voltages problem where the stock mobo power delivery is totally, unforgivably mental and we're still very much learning what is required for normal, non overclocked operation.