TR Forums

It's an AB350 Gaming-ITX/ac. It turns out that not only does it not see fit to tell me the CPU's temperature, it doesn't do fan control on it either. It's entirely running on the basis of a thermistor under the socket. In 2017. On a $110 board. What the hell were they thinking?

If I were more concerned about slow noise changes than reliability, I would be almost OK with this. Instead, I'm interested in low idle noise and spinning the fans way up fast when needed. As is, they take 30+ seconds to respond to 16 threads of prime95, and that's while set to be fully ramped by 9C (nine) over idle.¹ Not only that, but the sensor only finds a 17C difference between idle and full-bore prime95. My computer's environment varies seasonally by 10C. If I want anything resembling normal fan spool rate, I'm going to have to adjust my fan curves four or more times per year (and the ramp will have to be compressed into <5C, meaning huge and audibly annoying steps). All of this put together means I'm just going to have to idle my fans much higher than I would like to.

I haven't yet seen a legit temperature from this CPU. Nothing that runs on Linux knows how to ask the CPU itself (for now).

¹ I didn't even set it to 16 threads. I had SMT disabled for now due to that one bug (I haven't run into it, just a preemptive measure at least until everything else is set up right), and it for some reason glitched SMT back on at that boot.

Your phrasing is making this sound like it's an OS-specific issue, rather than the title's claim of a mobo (or possibly manufacturer?)-specific issue. Could you clarify on that? How does it behave under Windows?

The difference is that with Windows, there are purportedly tools that know how to talk to the CPU directly to get the real temperature. It solves the visibility issue, but not the fan control issue unless you chain something together with Speedfan or something complex like that.

The board's monitoring chip has an input for a temp direct from the CPU, it's just pegged at -128C, and the fans follow the thermistor under the socket so presumably they meant it to be that way.

Really this is all about ASRock. Maybe they've messed up other boards like this, maybe not. I wouldn't know.

Shobai wrote:

Your phrasing is making this sound like it's an OS-specific issue, rather than the title's claim of a mobo (or possibly manufacturer?)-specific issue. Could you clarify on that? How does it behave under Windows?

Sounds like he's talking about two separate issues -- 1) the motherboard's ability to control fan speed; and 2) lack of Linux support for direct temperature monitoring of Ryzen.

AMD has been very slow in getting real CPU temperature support (meaning data directly from the CPU itself, not external board sensors) reporting enabled for Linux. It's a known issue and will affect you the same way on any RyZen motherboard.

I'm a little more concerned that the off-chip temperature sensors aren't working, however.

If I'm reading his post correctly he didn't claim the board sensors aren't readable. Just the on-die one.

just brew it! wrote:

If I'm reading his post correctly he didn't claim the board sensors aren't readable. Just the on-die one.

The way he worded it sounded like there's a thermistor under the socket that he can't read that's controlling the fan automatically without any ability for him to override the fan settings.

chuckula wrote:

just brew it! wrote:

If I'm reading his post correctly he didn't claim the board sensors aren't readable. Just the on-die one.

The way he worded it sounded like there's a thermistor under the socket that he can't read that's controlling the fan automatically without any ability for him to override the fan settings.

Yeah, but he's also complaining that the thermistor does not respond quickly enough anyway, and he wants to use the on-die sensor instead.

Alright, sorry, I'll try to start over.

There are supposed to be two main ways to get information: directly from the CPU, and via a monitoring chip on the motherboard. Directly from the CPU is borked for now on Linux, as Chuckula said. The monitoring chip is an NCT-something-something (I forgot) using the NCT6775 driver, and it's pretty well-supported. Lm-sensors knows how to handle it, and the only issue that should come up is that it can't read high voltages so a lot of values are rescaled in hardware to fit its inputs and have to be manually scaled back. For instance, you have to double in1 to get the real Vcore.

This NCT chip has inputs for several thermistors as well as what look like a couple of inputs that let it get the temperature as reported by the CPU. None of the latter are hooked up.

The motherboard theoretically has as good fan controls as most modern motherboards (I always configure that in its firmware interface), but it can't control them on information it doesn't have. It's operating on the data provided by this same monitoring chip. Not only is this terrible data in general, but it turns what were previously very nice fan control features into misfeatures that may make the CPU throttle in situations where it has no business needing to throttle.

Here's the thermal monitoring both the mobo and I have to work with:

The two that look reasonably scaled are the only two that ever change. 32C is a very low idle, and it gets maybe as high as 52C under prime95. If put under prime95 from 32C, it may take 30 seconds getting to 43C. I know how I want my fans to behave, and I can't make it happen on data that shoddy.

Yeah, this sounds like typical behavior for every AMD CPU I've ever owned. Load up as many temp monitoring apps as you can think of. Every one will give you a different reading of the CPU temp, plus 1 extra from the BIOS. The real fun is trying to decide which one you think is right! Or, are any of them right?

The socket thermistor is readable everywhere and not at all troublesome from a software/monitoring side. I set a fan speed to 1%/C to verify the mobo is seeing the same temp I am, and it is.

Previous AMD CPUs have had weird non-linear stuff going on with that PECI line (at least I think it was PECI). AFAIK, Ryzen's only weirdness there is the 20C offset on -X SKUs. It just doesn't matter if nothing's reading via PECI.