On Sat, Sep 25, 2021 at 01:20:57PM +0200, Borislav Petkov wrote: > On Fri, Sep 24, 2021 at 07:46:10PM +0000, Yazen Ghannam wrote: > > I agree with you in general. But this device isn't really a GPU. And > > users of this device seem to want to count *every* error, at least for > > now. > > Aha, so something accelerator-y where they do general purpose computation. > > So what's the big picture here: they count all the errors and when they > reach a certain amount, they decide to replace the GPUs just in case? > > Or wait until they become uncorrectable? But then it doesn't matter > because we will handle it properly by excluding the VRAM range from > further use. > > Or do they wanna see *when* they had the correctable errors so that they > can restart the computation, just in case. > > Dunno, it would be a lot helpful if we had some RAS strategy for those > things... > I completely agree. The system integrators have their own policies for error tracking, part replacement, etc. I expect they'll propose kernel changes if they want any. Though I think general strategies will become apparent once these sort of devices are in wider use. Thanks, Yazen
