IRC channel logs
2023-11-01.log
back to list of logs
<stikonas>anyway, I'm not sure if it's much faster for single core workload <oriansj>ekaitz: hmm, 64 cores with support for ECC but board doesn't support ECC???? either incorrect information on page or they don't realize the average bit corruption rate of DDR4 per GB <ekaitz>i have no idea about hardware mate <oriansj>fair enough; most people don't know it happens and even fewer actually read the google study on the subject. <ekaitz>but i'm sure i'm able to understand it if you summarize it ;) <oriansj>imagine solar or csmic background radiation flips a bit in your memory; ECC would be able to detect and correct it <ekaitz>i even studies that kind of error correction stuff <ekaitz>i didn't know that memories needed that but it makes perfect sense <oriansj>and as the lithography process uses smaller and smaller features it becomes more sensitive to lower and lower energy radiation. higher odds of a hit as the lithography gets denser too and the more RAM you have the greater the odds of getting hit (it scales approximately linearly with memory capacity) <oriansj>a good rule of thumb is 1 bit corruption per GB per day (probably good bit higher now) <oriansj>so 128GB => about a corruption every 10 minutes <stikonas>also most memory is never written back to disk <stikonas>oriansj: we have a mess with mescc-tools <stikonas>oriansj: I didn't know that you tagged it on savannah <stikonas>but we probably want to include that extra kaem fix... <stikonas>so I think the options are either silently move tag on savannah or bump version, prepeare new 1.5.1 tag... <oriansj>stikonas: I pulled your tag and pushed it to savannah <stikonas>oriansj: this still seems to show the old tag... <stikonas>though maybe there is some mirroring delay <muurkha>radiation from the chip packaging is also a contributor <muurkha>chips are packaged in epoxy, epoxy is largely carbon, separating carbon-12 and carbon-13 from carbon-14 is not literally impossible but far too expensive, and carbon-14 is radioactive <vagrantc>although i only talk about reproducible builds in the abstract ... looking at my slides bootstrappable builds might take up more of the talk :) <ekaitz>vagrantc: is it going to be recorded? <vagrantc>they are doing a hybrid conference this year ... with both local and remote talks <vagrantc>will be a pretty introductory talk ... kind of like a trimmed down version of my fossy talk <vagrantc>although ... i am finding it hard to trim it down as much as i had hoped ... hopefully i will fit it all in the timeslot :) <stikonas>yeah, but it might still be able to reach some people that we can't in other ways <stikonas>vagrantc: also, if I recall early boot chain was a bit misordered in fossy talk, if you still have that, I can help you fix that <stikonas>ok, so simplified chain without kaem shell and rebuild of original seeds is: <vagrantc>i think i extracted the chain i had from what guix actually uses <vagrantc>but i may have missed something, certainly :) <stikonas>yeah, but even in guix it's the same order <vagrantc>there were some limits to what i could include in a single slide reasonably too ... but best not to be *wrong* :) <stikonas>hex0->hex1->hex2->M0->cc_x86->M2-Planet->mescc-tools ->Mes ->Tinycc <ekaitz>i'm not the only one that doesn't have a clear understanding of the whole chain... good good <stikonas>with nicer interface and a few more features <stikonas>cc_x86 is C compiler written in M0 language <ekaitz>that was going to be my next question <ekaitz>and M2 planet is a platform neutral thingie <ekaitz>written in that C subset, but what does it provide? <vagrantc>i dropped the non-x86 architectures slide for this talk <ekaitz>it's not platform specific in this case i guess, but it has to have several backends... or is it like an interpreter or something? <stikonas>but new features in M2 are structs on stack <stikonas>it supports quite a few more things than cc_* <stikonas>been awhile since I looked at source of cc_* <ekaitz>stikonas: what you explained is enough for me :) i just wanted to understand the high level <stikonas>I'll try to make stage0-posix release today or tomorrow <stikonas>anyway, M2-Planet is written in maybe 60% smaller subset than it supports