mrb

31 ms·

391.
by mrb 4y ago
Of course, server GPUs remain kings of the hill on FP64 workloads, as they have extra FP64 units. I focus on FP32 performance because it is representative of general compute performance on all other workloads, such as FP32, FP16, FP8, or in
392.
by mrb 4y ago
« are something inherent in the system that can’t be avoided » Firstly they aren't inherent. Losses can be minimized. But you don't seem to care, even though losses account for orders of magnitude more energy waste than mining. Yo
393.
by mrb 4y ago
So the 4090 is 38% faster (in FP32 FLOPS) than the current top server GPU (H100) and 105% faster than the current top desktop GPU (3090 Ti). And it's also more than twice as efficient (in FLOPS per watt) as all current top GPUs, even c
394.
by mrb 4y ago
Ah, i must not have compiled it with libreadline-dev... And I didn't know about rlwrap!
395.
by mrb 4y ago
The built-in calc sucks: when the keyboard is displayed to edit the current expression, the app can only display ONE previous expression at a time. Sometimes I calculate things and need to refer to 2 or more previous results. I would have t
396.
by mrb 4y ago
Very cool ! For units conversion I use GNU Units. But qalc seems to combine the benefits of both the Python REPL and GNU Units. I just tried the command-line version and noticed it doesn't support recalling the history (UP arrow) and e
397.
by mrb 4y ago
Interesting. Not exactly what I described though (sorry I did not explain very well.) What I meant, you type 2 calculations: > 1+1 2 > 10+10 20 Then if you tap the first line ("1+1") my ideal REPL would add it t
398.
by mrb 4y ago
What I want is a phone app calculator that works like the Python REPL/console. Save ALL previous calculations. Display the last dozen or so. Can scroll through them. Tap on one to recall it, edit one number, recompute. Done. This would
399.
by mrb 4y ago
On standard Android phones you can do it. I just tried split screen (GMail on top, Google Calendar at the bottom) and you can create or edit a calendar event while having an open email on top. Looks like this on my phone (Android 13): http
400.
by mrb 4y ago
The thief would just observe, from a distance, and wait for someone trying to scan the fake qr code. When he sees the victim, he pushes the button and steals the bike.
401.
by mrb 4y ago
You care about 1 in 1000 ? Did you know that 660 out of 1000 watts of primary energy used to create electricity are wasted by the time the electricity arrives at the customer meter? (source: https://www.enerdynamics.com/Ener
402.
by mrb 4y ago
The value you try to compute is currently exactly $79, including fees (which are about $1). If the number of transactions were to double (or halve) tomorrow, that value would halve (or double). You are just dividing fixed ongoing rewards by
403.
by mrb 4y ago
« I haven't computed it, but you need to take into account the energy it takes to mine a block » No, because the energy consumption is the same regardless if a block contains zero or a thousand transactions. Therefore it is misleading
404.
by mrb 4y ago
« its only valid use cases were illicit transactions and speculation » Crytocurrencies have many valid use cases. For one, they are the only system that allows sending money truly anywhere in the world, to anyone, anytime, and near-instantl
405.
by mrb 4y ago
« Shouldn't that be observable somehow? » No, because of two reasons. First, many miners simply pointed their hardware to mine other cryptocurrencies such as ETHW or ETC. For example we have evidence that about a quarter of Ethereum&#x
406.
by mrb 4y ago
« Because every bitcoin transaction costs $60 in electricity. » A common misconception. Only a fraction (about 1% at the moment) of Bitcoin's electricity consumption is derived from transactions, because transaction fees account for on
407.
by mrb 4y ago
As an early Bitcoin adopter, I wouldn't say so. I have never seen anyone among the people I knew in the Bitcoin community in the last 12 years who believed it would "make tradfi obsolete". For example, I expect Bitcoin to lik
408.
by mrb 4y ago
« I once got a request from the Netherlands for half a million floppy disks » I want to know (1) who are they, and (2) why do they need them?
409.
by mrb 4y ago
For those wondering: setDisplayMessage() { switch(true) { case(this.difference <=1000): return 'Looks like you basically have no idea how long 5.5 seconds is.'; case(this.difference <=2000): return 'I
410.
by mrb 4y ago
Mirror: https://web.archive.org/web/20220906155028/https://austingil...
411.
by mrb 4y ago
Big spammers own multiple ranges of IPs just as easily as they own multiple domains. When you block only 1 of their range, they can re-use their same domains to spam you from other ranges. If you blocked by domain, this would be effective a
412.
by mrb 4y ago
Everything you said applies to IP-reputation as well. There is nothing that makes domain-reputation inherently harder than IP-reputation. Consider this: every legit IP address once started with the problem of having zero reputation. How you
413.
by mrb 4y ago
« Most of the spammers have valid SPF and DKIM » What porbelm means is that when SPF and DKIM are valid, reputation check should be based on the sending domain, not on the sending IP. This way spammers still get thwarted (because of their d
414.
by mrb 4y ago
TOPS isn't indicative of actual performance https://semiengineering.com/lies-damn-lies-and-tops-watt/ Also, this IMG chip is vaporware and they can quote an arbitrarily high TOPS/W by trading voltage & f
415.
by mrb 4y ago
FP16 AI workloads are very important to AMD and Nvidia (AMD MI250X and Nvidia A100/H100 were really designed for this), and yet Tesla leapfrogged them with a more than 2x reduction in die area, and more features (eg. out of order exec)
416.
by mrb 4y ago
In terms of absolute compute performance per chip, perf per watt, and perf per die area, it looks like Dojo matches or surpasses the best GPUs of today: « Tesla claims a die with 354 Dojo cores can hit 362 BF16 TFLOPS at 2 GHz » For compari
417.
by mrb 4y ago
But Carlos did not do "all that". He said he tried VPS servers, not dedicated servers . Maybe that's his problem. VPS servers tend to have IP addresses with low reputation, whereas dedicated servers tend to have higher reput
418.
by mrb 4y ago
« So, starting today, the MX records of my personal domain no longer point to the IP of my personal server. They now point to one of the Big Email Providers. » I would have kept the MX records pointing to my personal server, and I would hav
419.
by mrb 4y ago
Right, I understand you now.
420.
by mrb 4y ago
There is no need to do "table lookups of hop count" or to "decrement hop counts". The IP TTL value is just a field that can be read from the IP header, which is trivial since the Python would get the entire IP header fro

More ›