"tevador: because nobody is hashing for about 4 seconds every 2048 blocks"

What's the most optimal way to use next_seed_hash?

total miners 32385

d'you have a scheduled script posting that count for you? :P

I'm waiting for "total miners: -32645" in a few hundred...

lol

i am the bot

i am the script

all your base are belong to us

i only post if its higher than the last i can see in the chat scrollback

someday someone that cares could parse their irc logs for the history. though my scrollback doesn't show day stamps, just time stamps. so, could require some fancy bashing

awk '/Day changed/{D=$5" "$6" "$7}/ginger.*total miners/{print D, $0}'

(should work with irssi, "Day changed" it what it says)

hm yeah, finch doesn't show me when days change either

aren't we coming up on 3 months soon? ASICs ought to be visible soon

/e

/s

I heard there's a company out of Santa Clara working on one

good luck to them

more Sillycon Valley VC idiocy

Their CEO is an MIT graduate electrical engineer, so I think it would be dangerous to dismiss her at this stage

Still nonsense. If she knows any cool tricks, she'd make more money poaching on Intel's turf.

I think that's what they're currently doing ;)

as I said before - good luck to them ;)

Intel themselves are quite vulnerable at the moment, of course semiaccurate.com/2020/02/06/intel-officially-craters-xeon-pricing

Actually I really wish Mill or REX computing would get off their collective behinds and make some RandomX processors

It would be an excellent way to demo their ideas

is anyone actually building Mill designs?

AFAIK no

didn't think so

I think they're falling in the researcher's trap, designing ideas which are 10 years out rather than bringing 3-years-out ideas to mass production

Something I really appreciate about Tesla is they're not fooling around with solid state batteries, instead they're scaling up mass production of current technology

yes, Tesla was quite pragmatic

You can sink a lot of time and money into ideas which are "just 10 years away"

thouh now they've built up enough capital they can afford to churn out gigafactories and do some novel battery research

I think they're playing a cynical game of letting the other auto manufacturers fund solid state research and they're just going to license it once it's ready and then focus on mass production IP

which other manufacturers are researching it? Why spend the money when they can just license Tesla's tech for free?

> Major automakers like Ford, Hyundai, Nissan, Toyota, and Volkswagen are investing in solid-state battery research, as are some Chinese EV companies and Fisker.

sounds like severe low-temperature downsides

I think the other auto manufacturers feel like current-gen battery tech is already tapped out so the only place to have impact is on next-gen... It's a bad look to buy your hardware from Tesla, investors are going to start asking "why not just invest in Tesla?"

also previous research into lithium-metal batteries has usually ended in large explosions

Tesla is probably looking at this thinking "one of these projects will win, that one we'll license and the other 9 will fail and we get that research for free"

the more practical improvements I see for near-term are replacing carbon with silicon in the anodes

yup, most definitely

And meanwhile Tesla is developing autonamous car technology so that even in a worst case scenario where someone discovers excellent battery technology but makes impossible demands to license it, Tesla will continue with current technology and upset the entire auto market with autonamous vehicles

What they have that Uber doesn't have is cameras on all of the cars, so just by driving a Tesla you're helping train the AI

i hope at some point the demand for cpu efficiency will start pushing the 4-5 GHz limits.

at some point parallelization doesn't get you what you need. mainly, realtime stuff for AI etc

well, perhaps not efficiency, but performance

yeah, higher clocks will trash power efficiency

Intel already learned that with the pentium4 netburst

"I heard there's a company out of Santa Clara working on one" " Their CEO is an MIT graduate electrical engineer, so I think it would be dangerous to dismiss her at this stage"

lol, that was very subtle

Lisa Su?

Advanced Micro Devices, Inc. (AMD) is an American multinational semiconductor company based in Santa Clara County, California

there's plenty of small semico companies in silicon valley. that's literally why it's named that

At the height of the GPU mining craze, AMD could easily have jumped in themselves. They didn't, they just continued churning out GPUs as always

I want to see a fully async processor with a inward-feeding (reverse of the mill) operand belt

s/ a / an /

Anyway, one reason why I'm highly positive toward Monero and the RandomX algorithm is because it acts as a challenge grant to push microprocessor research forward

15:48 < gingeropolous> i hope at some point the demand for cpu efficiency will start pushing the 4-5 GHz limits. <-- You can reduce heat output a bit by going into different processes (Gallium arsenide), but really if we just write better code, lots of things can in fact be done at the same time

lol. yes, writing better code can get us at least an order of magnitude improvement already

IMO real limitation right now is the assembly language, it's designed for a single core that just stalls every time it needs a memory access

Compilers are not thinking that way, and processors aren't either, but it's the way we communicate

WDYmean? You can't hide or avoid memory stalls

Sparc T1000 tried too

processors are reordering instructions pretty hard to hide them

and compilers are trying to make instructions in such a way as to help processors achieve that

.... X86 mainly is reordering. ARM OOOE is pretty limited

X86 needs to because its ISA sucks so bad

reordering isn't really a critical requirement in a well-designed ISA

although I suppose, as you say, the model only works well for single-thread

compilers can't be relied on for this, with multiple threads the compiler can't account for the CPU state

this was the too-late realization of the Itanium designers

Well, I think Itanium fell into the 10-years-out trap

or maybe 20 years out at that time

lol

compilers weren't there yet

compilers can never account for the current realtime state of the CPU running multiple threads/processes

it's impossible, and was foolish to ever rely on that.

right indeed

My estimation is that the best thing to do is cut the code up into pieces which do not depend on memory access

difficult to do except in very specific algorithms

So a memory access is an async operation and the next chunk of code will only run after the memory comes back

we wrestled with this quite a lot at JPL. Large FFTs across big frame buffers.

Well you can do it mechanically with any code, you're just going to get a lot of very small pieces

you basically had to hand-tune the algo for each CPU's cache size, and break up the frame into tiles

FFTs are a fairly straightforward algorithm but most problems don't slice'n'dice so conveniently

But if you start thinking about LOAD as an async operation, then the CPU can interleave quite a large number of threads

however branch ops throw a wrench in the works

I'm told that in the GPU space, this type of interleaving is already happening

But I don't know that well myself so I just go on what I've been told

like I said, in scientific computing, we've been doing this for at least 20+ years

*nod*

I worked on optimized FFTs for intel i860, same story - load-to-use latency, reorder instructions so they don't try to operate before operands arrive

yup

but that's static reordering in the assembly code. not dynamic inside the CPU

cjd memory load instructions on AMD GPUs return immediately, you have to explicitly issue s_wait_vmcnt(N) instruction to use their results

and you can insert as many other instructions in between as you want

hmm, that's clever

I actually use it in RandomX implementation for AMD GPUs

reorder memory loads up, s_wait* down in the code flow

*nod*

You're not using cache, IIRC

So basically every scratchpad hit is a memory access

I don't use it, and maybe putting first 16 KB into local memory can give better performance

but it complicates every memory access to scratchpad

yeah, hard to beat straightline perf, adding special cases tends to make lots of things worse

they claim RandomX is copyright Dero lol

I get a 500

502

bad gateway

it’s a Go rewrite but still AFAIK you can’t change copyright like this?

hmm it worked 1 minute ago

they can claim copyright on their Go implementation, but they have to credit tevador

how about taking square roots of negative numbers? xmrig/xmrig #1553

LOL

17:43 < sech1> I don't use it, and maybe putting first 16 KB into local memory can give better performance <-- so you can interleave quite a large number of computations, I suppose

they only provided x86 JIT, not arm64

sounds like they didn't test adequately, would've uncovered that problem

on different platforms

actually on Navi the GPU core is underutilized, it can run at half the speed (900 MHz) and with the same hashrate

so maybe using local memory there to reduce memory bandwidth can really help

ahh ok, to its memory bandwidth bottlenecked

the thing is, I don't want to invest time in this known it can only boost hashrate probably 2x

still not very competitive

*knowing

wtf, a PR with a message that's a marketing pitch. Fucking people...

sigh

Gotta admire that hustle

this is such crap. I feel we've probably discussed it before medium.com/deroproject/analysis-of-randomx-dde9dfe9bbc6

"An important piece in the security of RandomX depends on the CFROUND instruction. It is also hardware dependent, which means that a chip manufacturer (Intel, AMD, ARM, etc.) can alter or remove their implementation, resulting in unexpected behavior that could lead to forcing invalid proofs on the chain."

eh actually no, rounding is part of IEEE-754 spec, no vendors can arbitrarily choose to withdraw it

can?

nobody will get anywhere today selling a chip without 754-compliant FP math

that might have worked 50 years ago. not today.

That's kind of sad though, because Posits can be more efficient

But I agree with your analysis

Posits have their own edge cases where they're much worse than ieee-754 floats

hyc: they are doing this all the time. Last time they copied bulletproofs, started to call them Rocket Bulltproofs and claimed their version is 10x without any proof.

they also had a huge premine. I’d say most of the shills on Twitter are bought or sockpuppets.

not surprising

I wonder how big their team is. Probably more brain cells in my cat than in the entire Dero Foundation.

at least Kristy kinda called them out. Curious how they’ll spin it now :P

yeah, she was pretty incensed

sure, more efficient code could do it. but im still .... slack jawed with the idea of using live machine learning to control the tokamak for fusion reactors

and i think thats ultimately gonna require the next level of computation

AI controlled nuclear reactor, hm

<AI> I have a question, can you raise my allowance?

i mean, its just a matter of putting just the right amount of power into electromagnets at just the right time.... i don't think we'll get full blown AI from that :)

i mean, you gotta get the input from the sensors, run the calcs, and then pump the proper current to the proper electromagnet in less time than it takes for one of them magnetic fields to snap

<hyc> I wonder how big their team is. Probably more brain cells in my cat than in the entire Dero Foundation. <--- their website mentions 3 fulltime guys at the beginning, and that the "original developers" are still working (not surprising since this is a premined)

From their website "Dero Atlantis uses pre-compute tables to convert 64*2 Base Scalar multiplications into doublings and additions." I don't know what they mean by pre-compute, but otherwise it just seems to describe how all multiplications are done in ECC no? Ie, irrelevant gibberish

yah it all sounds like gibberish

I've had to stop responding. no point giving them any more oxygen.

always wonder if its better to not respond at all 🤷‍♂️

it strikes me as harmful to allow single wrong statements to go unchallenged. but also harmful to get dragged into a long pointless dialog

I don't think there is any universal right answer to that question

If one adopts a rule of never answering incorrect statements then FUDsters will abuse them by making them appear to be afraid of the truth

If one adopts a rule of always answering incorrect statements then spammers will use them to get free advertizing

yeah

Being unpredictable is the only robust solution

it generally seems to me that the world is stacked against being the good guy

i think the issue is the medium and location. if they want to spout off in their own corner its best to ignore but if they come to your "house" then response seem warranted.

This concept of "good guy" makes me uneasy