XMR network hashrate 60 : 2,001.841 MH/s

last 60 blocks ^

Did someone hack Azure again?

so this azure hacking has been going on since april according to their blog post

and they still don’t properly monitor it? lol

what blog post?

oh my... XMR network hashrate 60 : 2,115.315 MH/s

on the other hand, there are only 10 unknown in the last 100 blocks and pools don't show any increased hashrate

maybe just a strike of good luck?

Damn, lazy fuckers downloaded xmrig straight from github: microsoft.com/security/blog/wp-cont…020/06/Misconfigured-Kubeflow-1.png

Monero PoW on Hacker News

sort of funny the people that discuss the algo, not taking into consideration the context it is deployed (though i suppose that happens a lot in life generally)

the only meaningful comment there is from thesz but he's totally wrong in his estimates

all the optimizations that he mentions fall within 2-3x possible efficiency speedup

meh. once he said it was unnecessarily complicated, it was clear he doesn't know what he's talking about

e.g.: using fixed point to emulate floating point is ridiculously slow

but most of all, pointing to optimizations to a braindead simple algo (FFT) as proof that you can accelerate an ASIC with a codesgin tool is utterly missing the point

just as an aside, when I worked at JPL, our hardware lab had a couple decades of experience building custom floating point array processors for FFTs

in the course of the time I worked there (3 years) all of that got retired in favor of massively parallel supercomputers

Intel i860 based, SPARC based, many varieties

I worked on the i860 compiler and optimizers that made a lot of that possible...

fundamentally it's a memory-hard problem, it's easy to pipeline the CPUs so they're 100% busy, aside from waiting for memory stalls

restated - the actual computation isn't the slow part

hyc we should start a counter how many times "RandomX is dead" is said

like they have a counter for "Bitcoin is dead"

lol

this guy now claims 10x would be possible news.ycombinator.com/item?id=23530196

sounds more realistic already compared to his 100x in the first comment

still nonsense to compare any fixed algorithm to a randomly generated one

FFT requires no dynamic scheduling at all

except RandomX doesn't need "more shifters than the CPU", the instruction mix aready matches common CPUs

wider issue: not possible because of isn dependencies, only OoO would work

nor can you do ahead of time preprocessing

in parallel, since every instance starts from a different nonce. the scheduling demands of each parallel thread will clash.

IIRC the simulations showed that VLIW over 4-wide had little to no performance gain

two of these guys claimed they eval'd for their work. sounds like their employers got a bad deal.

I think one of the guys was from X41

the other one, I'm not sure

maybe TrailOfBits

So he has a secret knowledge how to make CPUs 10 times faster?

My CUDA code has VLIW interpreter. I don't remember exact numbers, but it achieves around 3 RandomX instructions in a single VLIW instruction

Memory dependencies and branches make it hard to go wider

considering we paid X41 & ToB to report any potential falws, and they didn't include these kinds of comments in their final reports, I'd say if that was true then someone over there really f#cked up

Demand moneyback?

i.e. withholding conclusions that we paid for

but I'd want to know for certain that those are their employers

this seems to be his "employer", or at least he checked in some PoW related code to one of their repos: github.com/hexresearch

I think "q3k" is from X41

Found "thesz" after some googling

would be great to ditch the term asic resistance and replace with asic equivalence

yes q3k is listed in the X41 audit

yikes lol

q3k just repeated their VLIW "issue" from audit report, no?

hm, could be

I don't see q3k in the x41 report

he wrote the hardware part

Serge Bazanski

ah I see

in that case fine, already reported and dismissed