Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

Don't get your hopes up about AVX-512. Apparently, full support is not coming until Cannonlake: http://www.kitguru.net/components/cpu/anton-shilov/intel-sky...


That's only true for Intel Skylake mobile/desktop CPUs. I find desktop and mobile offerings too unreliable for anything serious, due to lack of ECC support. Data corruption on non-ECC systems has bitten me more than once. I've learned my lesson.

Skylake Xeons — the only Intel chips I'm interested in anyways — on the other hand will support AVX-512. The software I'm writing already has rudimentary support for AVX-512. The prospect of processing 64 of 16 bit elements per clock cycle per core is pretty exciting. 32x per instruction (512/16) and 2x comes hopefully from dual issue.

Besides, we're talking about AMD chips here, aren't we?


How did you verify you were bitten my memory corruption?


By analyzing the end result. Apparently single bit corruption is not that uncommon. It's striking what it does to a CBC encrypted stream...


It might be (more) common if: your BIOS has a wonky config (memory timings), not enough cooling or not a great PSU

I wouldn't do 'hard work' on a machine that doesn't pass a Prime95 and Memtest (or similar programs) run.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: