3090 LA - high load average

Hi all,

Can’t get stable work of 3090-s on HiveOS.

GPU: 2x3090 premium brands, no thermal issues
Motherboard: H61M-P20 (G3) (MS-7788) MSI (V1.6 04/17/2012)
CPU: 2 × Intel® Celeron® CPU G1620 @ 2.70GHz
Disk: USB 2.0

Ambient temp at ~20-25C
Cards temp at 52C max, usually 40-49C

HiveOS, ethermine, t-rex (newest)

On Windows, I got stable work for 5 days at memory junction temp not exceeding 94C with the following overclocks: -200, 1200, fans 100% (all on MSI afterburner), 273-299W power limit (printed by t-rex), cards at 48C, ambient at 28-29C.

When I turned to HiveOS reflecting same overclocks (C-200, M2400, F100%, PL300W) the rig keeps crashing - LA (load average) increases and rig stops hashing, but still consume watts at the wall. Very frustrating.

Sometimes recovers after rebooting, sometimes not.

Log from t-rex at HiveOS shows following message:
“t-rex exited (exitcode=0), waiting to cooldown a bit”

But cards at <50C before and during crash !!!

Other meaningful prints could be:
“TREX: Can’t initialize device [ID=1, GPU #1], cuda exception in [initialize_device, 96], CUDA_ERROR_UNKNOWN”
“WARN: GPU #0(000100): ASUS GeForce RTX 3090, intensity 22”
“WARN: Built-in watchdog has been disabled!”
“WARN: NVML: can’t get fan speed for GPU #1, error code 999”

Played with overclocks a lot to find any solution. Even hashed at 80% of potential, still crashes with LA increase.

Any tips from admins?

I found it hash with T-Rex miner
core clock -100
memorry clock +1000
Fan 0 default/automatic i keep on 85
PL watts 260
don’t forget during summer and in hot countries like africa it will perform not as good as in north pole, ambient room temperature plays big role, open windows/door keep fresh wind breze
also keep drivers up to date here vid how to
https://www.youtube.com/watch?v=jwWaez-CSKY
cheerrs

Did you find a solution???

I think the problem is the CPU.

Maybe i would try with an Intel Core i5 (it has more cores)
and Power Limit 305W for the GPUs

Disable nvml, exemple whit t-rex miner “no” : “nvml”.

Is this an actual solution to high LA?

This topic was automatically closed 416 days after the last reply. New replies are no longer allowed.