Seeing so many people struggle with this problem I wanted to share what I discovered by accident in case it saves you time. I had several abrupt shut downs under sustained load, capped clock to 1800 as a sticking plaster then ran partnerdiag and got a fail on GPU stress after 3m18s. I had my RMA request ready but when I moved my machine, I noticed it was dusty - it seems our beautiful gold front is a great dust trap. After a few 60 second sessions with a vacuum cleaner, the idle temp fell from high 40s C to mid-30s C and the partnerdiag is now a pass. Since that is a short test, I also ran a 1 hour soak test with no shut down. Worth a try, even if your work area is less dusty than mine!
You might want to have in mind, that Spark’s cooling profile is apparently driven mainly by the temperature sensors on the GPU side of the chip. When you underclock the GPU it would stay cooler and consume less energy but this wouldn’t apply necessarily as much for the CPU side which in certain circumstances might be literally cooking itself. So, either underclock the CPU as well or consider other strategies for improving the overall cooling for all temperature-sensitive hardware in the box.
For my part, running with stock clocks and cooling profile settings I’ve seen situations with sustained CPU temperature above 86 °C while the GPU is not processing anything and the fans are essentially quiet…
As a side note, I’ve experienced somewhat similar issues in the past with a completely different hardware setup (1u sun fire server, after motherboard replacement and subsequent better application of thermal paste on the CPUs than the factory one before that, all available cooling profiles were leaving fans at a too low speed and that lead to (registered!) RAM errors caught by running memtest – the solution was to run the fans always at max speed, regardless of the relatively cooler CPUs and this was achievable only by cutting wires on that system… ). So, the lesson is: consider all possible consequences of a given change at a system level.
I solved the problem as the same way, but i cap it at 2000. I never had any other shutdown for months. On ascent, is the one i use,the coolers are interconnected, if gpu deliver many heat it make cpu be hottest. But using models we are using intense gpu and not cpu, for this reason i think is not necessary cap cpu. Btw, i tried time ago and i dint found how, system says its changed but cpu is still using max hz. Anyway, capping gpu is a good workaround. Also it doesn’t affect on token production since the bottleneck is ram and not gpu. Maybe a lil less prompt processing.