Meet the plotting experts

okay everybody hello and welcome to uh with uh cocktails for the plotting experts whiskey with wales it's uh february 19th at four o'clock pacific in order to take about an hour's worth of questions uh both here on zoom and in the test net channel uh also want to make sure everybody's aware that rc2 is out uh please do upgrade to it because i bet that the beta 27 chain will die a horrible death very shortly but with that uh let's open it up um how about a quick uh intro to the level you're comfortable obviously not a whole lot necessary but you know just how'd you get involved what do you do what's your you know what's your background sort of yeah i guess i can kick it off um so many of you guys know me in the chat is storage jm my name is john michael hands i'm a product manager at intel um so my day job is managing the product line for intel ssds so obviously when i heard about a storage cryptocurrency i'm like this is interesting i need to look into this um you know unfortunately when i first looked into it like in you know 2017 or 2018 was like the first article about chia you know i waited a long long time i kept trying to see if it was launched yet i finally caved and downloaded the key base uh i want to say like in july and installed the alpha and i ran my first plot and i was blown away i said you know i measure endurance for ssds you know in lab and stuff and i looked at the plotting uh you know the plotting uh workload i'm like wait did that really just write six terabytes of data to write to make 100 gigabytes fly i was just floored i've never seen like you know i've never seen anything like that in my career so i was just so interested first first couple weeks i was taking kernel traces just digging into the plotting workload kind of looking at it like a black box and like i was trying to deconstruct it i'm like i'm gonna figure this out and crack the code and you know figure out the most efficient way to plot and so i did it the way i know how to do it which is kind of like deconstruct it from the storage standpoint cool michael's rocks yeah my name is michelle erp i go by zorinox for forever whatever you search on the net that's under my name i do i'm a lot of things but i i'm a biologist and a computer scientist and i've all had always had jobs in that space doing computer science software engineering for for labs supporting science research i got involved with chia a little bit before with failing getting anything out of burst uh i learned about shia and then i came back to the chia website like once a month to see if there's anything posted that you can try and test uh you happen to be on your website uh early last year i think you just got rid of mongodb from the stack installed it and started to i joined the community and started to ask questions and yeah started plotting and there was a person that originally made me aware of those terabytes that are been written when you're plotting and that was early spring it was a bit of a shock for everyone it was different yeah derek yes it has improved a lot yeah it has yeah oh yeah um can you hear me well yes great yeah i'm uh on the forum uh farming 101 um my background's got absolutely nothing to do with uh computers or computer science whatsoever i specialize in uh in process development for for a company called bhp all my life i was the manager for the technology development manager for um either methodology and chemical engineering so yeah i always my life's work was about environmental processes designing uh and implementing more environmental friendly mining practices and i've done six or more patents on that and 20 or more international papers so really nothing to do with um with computers and so on and i i stumbled upon chia basically as a way to to add value to uh to energy what i wanted to try to do was to run a like a house or a small business you know basically costing no money and the only way i could do that was by adding a value-add component to a renewable energy installation in the house and i tried the mining process but the energy balance didn't work it didn't scale very well so i stumbled upon the poc cryptos which burst was not a success i i didn't i couldn't latch on to the project and then stumbled up john cheer 2018 and then just started with it then i was the first minor first farmer on alpha one outside the development developers great amish my name's hamish uh better known as h or q h um i'm a convert over from the burst bhd mining over to chia got involved in disk based farming mining because it seemed like a much better way of doing it than proof of work and i've been kind of caught the bug and been doing it ever since just a note we are at capacity as i said i tried to get the youtube live stream to work but we unfortunately have not gotten quick enough response from zoom's team uh we'll try to keep this uh you know uh flowing across and maybe uh side of flake and others can uh you know shuttle questions back and forth anyway and uh this is one other person who just showed up but who's this you're muted brim are we recording yes we are okay good i was i was teasing you uh were you just the guy asking questions okay so let's get into it um what are the primary bottlenecks to effectively plot chia plots everything yeah this is the this is the most fun about chia they're like i'm going to change one thing and it's like well you got you need cpus you need to scale the amount of cpu cores for the number of threads and processes you need a certain amount of fixed capacity for temporary space and it needs to also be fast and by the way if you have a big drive your ssd has to also support a high bandwidth because you need to support a bunch of concurrent bandwidth you also need a bunch of dram because the drm scales per process so it's really just this multi-dimensional puzzle and we're looking at like you know different cpu types and different cpu instructions for core and frequency you know different types of storage mediums and we've got a lot of questions in the chat about the status or sas for cnvme and there uh there's so many different ways to attack the problem which is interesting so it's fun you know as you've seen uh zornox and myself and folks here you know built just about everything i have you know i work at intel so i have desktops you know stacked to the roof in the garage about every type of intel cpu and we've got servers we've got all kinds of different stuff to test so there's it's cool to see lots of different approaches been successful um i think we'll talk a little bit about that today yeah i think that's actually a perfect lead in the next kind of question which is i think each one of you have somewhat come at the problem slightly differently so i'd love to hear kind of each of your takes on you know how you went about it possibly it just is you know my thoughts are more background than anything else here um the algorithmic problem of plotting is very it's a fun thing to work on it's very very different from the normal uh things you worry about in um in algorithms in computer science uh there's a few notable little quirks at the edges of it uh the biggest thing is it's optimizing over number of passes over the data uh and we're actually using some techniques that i suspect like aren't in hadoop and would probably make hadoop run a lot better if it was um and we're doing a whole bunch of sort passes and actually we're doing this like sort uh this like bucket sort in memory which even our non-paralyzed implementation of it is a lot faster than the normal sort functions uh which is weird but it only works on really really nice smooth data uh if the data is at all not smooth it just fails miserably but our data being cryptographically generated is uh perfect for this but that's sort of an interesting thing to know there might be many real world data processing things where actually bucket sort would perform better uh depending on you know distribution of keys and such um uh we still haven't finished our optimizations you know we did all these uh we had this long sorted history where we had a competition for uh improving our implementation of it and we actually put together a bunch of notes about how to um about how you could go about improving it and that competition mostly failed uh that um uh well in a few ways so people did manage to get a little bit of speed improvement but they didn't really do any algorithmic optimizations to number of passes over the data and they didn't really um and they didn't change the format at all so there's optimizations that can be done to the format to compactify things more uh that we're expecting to be implemented reasonably in the future but we're not working on right now there's basically two of those worth thinking about um but then for the just minimizing the number of passes over the data which is this really super weird and interesting thing that when we have gone and done optimizations and we got like basically like a factor of three speed up from all that i think uh after we did it ourselves um we're getting to the point now where the optimizations like that are a lot harder to do and their trade-offs are more ambiguous it's like it might help under these circumstances it might hurt under those circumstances it's complicated um but we have more of those likely in the pipeline i'm not expecting another factor of two improvement in speed from algorithmic improvements in it um uh and i'm not going to be the one april but to get probably after we launch arvid is going to go back to working on those uh sorts of things uh yeah but the the it it is at the point uh uh like you said that um there's all kinds of complicated trade-offs that happen you're really trying to make it be that you're limited by bandwidth but you can also be limited by cpu and you have issues with how well you parallelize across all of your cpus and then you have to worry about whether you're destroying your ssd by doing more rights than you have to to it and um and there's also just like cost versus latency on this stuff right like you can do plotting on a physical hd it's slow but it'll work so you know if you have enough of them uh you could do that as well so it gets it gets all complicated uh but we we will have more um algorithmic improvements in the future uh not nearly as dramatic as the ones before that they we got it working well enough that that kind of stuff got lowered in our priority list so again let's go back to you know what each of you kind of approached it with because there are some different kind of overall approaches that you're taking well for me i realized early on that sort of nvme was going to be the fastest but to get to the scale i wanted in vme just wasn't an option for me in the work i do i do a lot of work with servers so i went with a server-based approach sas ssds multiple parallel go wide rather than go fast yeah it's interesting like you know my my gut feeling said that you know a desktop would outperform you know i have you know done like the top of the line intel you know i9 10 uh you know like you know 10 18 core you know max cpu clock speed and those you know produce really fast results but then it's much harder to attach a lot of io to a desktop and that goes back to a server where a server has been designed to have you know serviceable accessible storage it's designed to have a lot of it and you know going back to like practicality i had my fastest plotter was my i9 desktop but every time i put a new component in there it i had to reset the cmos and i had to reset the entire os and i had it which is such a pain in the ass servers are meant to have you know a bunch of different cards attached a bunch of different you know storage and i o performance things so yeah i you know i went from desktop to server and i have some desktops in the house that are plotting right now but uh at the end of the day you know servers are designed to be running 24 7 be running sustained workloads be running at max cpu utilization at all cores be running thermally efficient they have fans that are designed to you know you cool all the components and you know you can't beat with these servers that have 24 two and a half inch bays for trying to you know support a lot of drives uh so yeah there's going to be a bunch of different approaches work especially depending on people's budgets um yes yeah i have early early in in when i realized that um there are a lot of uh temperatures i i i also decided that the nvmes is not the right thing if you want to do more plots than a couple drives full because you basically for every couple drives you buy you have to buy a new nvme drive unless a consumer grade at least so i've looked for alternatives and i have i've actually benchmarked uh and compared how much faster is the nvme drive versus a zeta drive a set of flash drive and there wasn't that much of a difference so uh for a single process in the single plot at a time obviously with an nvme drive that has very good sustained i o you can scale multiple processes on it but it's still once you want to have four you already need a large nume drive which is pretty expensive uh for the same amount of money you can buy uh used um data's integrate sata flash drives uh it's just a completely different build so um i have done a lot of work in the hpc space and when i really saw how much i o how much compute and processing is happening it's really it's a high performance compute application if you want to do this beyond a couple thousands of terabytes so i have designed my desktops in ways that i can fully utilize every single pcr lane to the max and have for example not purchased 12 or 16 core cpus i i stayed with eight because then i i know i only have that many pci lanes i use them all and i buy another computer now if you're if you're in the service segment you could do the same maybe you have a dual socket system this is a way do you have to decide also considering budget and space i have an or i'm in an apartment i i can't have a server rack anywhere here really i can you you have you have to decide how much do you scale up or how much do you scale out and and i took the scale out approach so i have many machines that are built similarly with up to 24 sata drives and gets really really great results eric what was your take yeah the first machine that i i bought you know was uh was the i9 900k and you know the similar to what you guys said is at the end of the day that desktop is a gaming machine it's fast single core but the io is restricted so as a non-computer person i didn't know anything about servers so i took a completely different approach i i built a lots of small computers four core and six core so with that you know they were cheap you could build a machine for 150 dollars right and it's mostly nine one hundred and pentium uh five four hundred six four hundreds the single core is pretty good for the price you pay on it so i just built 16 of those they can run four plots each and there is no restriction on the i o you don't need to to try to to balance the plots or wait you know try to synchronize stuff you just pretty much press play and the thing can plot 16 terabytes a day and you don't have to worry about it they are still the same as when i started they're still running never been down and uh from the you know from the addition of the intel uh enterprise drives that made it even better right because now you had um endurance to that and yeah those things are great each one has got a little screen on it and i can just go for it we don't mean this to be an intel ad but it is and it should be uh i just want to add one thing for people who are just tuning this to replay or just tune in right now you know i'm the shrimp in the room and i've just got a three-year-old imac and a nice usb 3.2 external and i get nine 10-hour plot times and i string it together and let it run forever and slowly but really fill up splats so it's not like you have to go here but if you want to get high throughput this is the way to do it uh let's talk about how much you used factored in i think you know that was kind of came up what else are you guys thinking you about used how has used hardware been part of what you guys are doing well i can tell you it brings joy to my eyes seeing all these hilarious you know all the drives i worked on six seven years ago that are showing up on ebay the people in the chat are asking me about and people are dming me with like this is some weird firmware i see on this drive i'm like oh my god i remember this like weird bug like seven years ago you know uh but no i mean you know they um the thing about ssds that i think people forget was when flash every year like flash was really good every year they tried to reduce the cost so you know every year ssds get cheaper and the challenges of flash moving from you know 2d nand to 3d nano then scaling the layers you know sometimes in that big transition came at a trade-off you know when they went from mlc to tlc and tlc to qlc you know consumers just don't need a lot of endurance and a lot of performance the cost was king and so in the last you know 10 years that's mostly what the ssd industry has been optimizing for now the ssds are getting faster as well and then i'll talk a little bit about nvme and some other stuff but those drives that are like five or six years old the first nvme drives that launched in 2014 are absolutely beasts in endurance you know they just don't make drives that are 25 and 20 drive rights per day anymore it's like you know those were just way overkill for the workloads and you know people have gotten much more sophisticated and be able to model endurance and think about actual drive wear out in the beginning of the industry people were so scared that ssds were going to wear out that they just they overbought endurance and so some of these old drives are actually great it's funny people buying old drives that are five years old with like 80 90 endurance left on them you know anybody else on the use side so uh let's flip to you know uh some people walk up and go hey i'm gonna use nvme and we've kind of talked about it but what should people be aware of when they're thinking about using ssd as temporary plotting storage so let's be clear i mean sata is dead um you know the last sata speed improvement was sat at 6 gigabits per second it was i think 10 11 years ago and the entire industry you know meeting the people that make ssds and hard drives got together and said okay is there going to be a sat at 12 gigabits per second and the answer was no it would one it was you had sas which is for enterprise and that's one of the questions we get a lot in the chat about sas for sata you know sas is specifically multi-host multi-device it was meant for these large topologies of jbods and cascaded jbods many disks you know being cascaded from servers where sata is very single host single device type of application um but 10 years ago the industry decided let's go to nvme and we'll use pcie to go you know drive that performance scalability year over year and then nvme was the new protocol for ssds so now that the price of nvme drives and sata are basically the same you know buying a new drive no consumer would really be interested in sata uh besides the fact that it's very easy to hook up you know because it shares the same interface as hard drives you have a 50 cent cable you can hook up a bunch of drives on a desktop extremely easy easily and cheap um you know i have like a you know a pcie cable here for a desktop right this is like a 15 cable on amazon you have to connect a u.2 drive you'd have to nvme to a desktop and it's 15 versus a sata cable that's 50 cents so that's the whole point of sata is it's cheap easy everybody understands it but you know there is no there is no future of sata it's it's a dead road map great so uh can somebody talk about the ratio of ram to number of cores and how they relate to number of buckets in parallel versus cube anything yeah michelle go ahead go ahead i just wanted to say that one i think the one guy that figured this all out is actually not on the panel you know pyl he has posted a there's a very good description and i'm still just following that it's in it's one of the discussions uh the title might be misleading it's the something like the crazy new plotter or something like this but it's basic i have been following this forever and i have not played much around with it but i think he has some more details yeah a caveat about any numbers here is our implementation still probably kind of sucks so there probably are like if you see something where like adding more cores past a certain point just isn't helping entirely plausible that like that's just a software issue and if something was changed that could improve that and there's all kinds of weird deep voodoo when things are bottlenecked on throughput there you always want like the cpu running while throughput is happening and you don't want them to sequentially wait for each other getting that to work right is deep magic and we probably haven't done it super well in lots of places in our plotting software yeah i have commented on that many times this is this is very very very hard this is the core business this is the core business of any database vendor or open source project is how do you get the course play perfectly well with io yeah and generally like in in many databases like a mysql instance you know they're very dependent on dram for the i o so the fact that in this case in plotting you're doing so much writing you know you can't really take all the advantages of some of the modern database performance which is actually tuned for read and analytics and all that stuff that is mostly read heavy um yeah it's really interesting we you know for the the deer when i took a kernel trace of the plotting very early you know basically the way it creates the temp files and then you know from it dumps those onto the drive from dram and it happens basically all at once so there's a big spike at the q depth of the ssds uh that are basically writing all those temporary files to disk and then when those are done the plotting process cleans up those temp files which happens on the ssd it sends trim commands to those you know to the ssd the corresponds to the lbas of the temp files some ssds handle that really well some some do not right that's why people see a lot of variability when you have these massive amount of file deletes happening all at once you know some of the older drive models the firmware does not like that you know so you know there's a huge variation again people using all kinds of different types of hardware but yeah it's interesting because we could compensate for that in software by like not deleting files um by using sections of the same file instead yeah yeah that works too like um there's a like enterprise well i would say hyperscale database like rocksdb that like facebook uses for all the graph databases of facebook they actually rate limit the trim going to the drives because they know about this i o you know discard you know if the more trims you send to the drive the higher the latency is going to be now the thing in the plotting algorithms is you don't really care about latency you're just really caring about disc bandwidth you're not you don't care how fast you receive the ios in the temp file you're not caring about a quality of service for latency you just really care about how fast can you get those to disk um so you're really optimizing for something a little slightly different but yeah well it's all weird because you are you don't care about the latency well you really don't care about the latency on your rights uh you you mostly don't care about the latency on your reads as long as they're coming in as fast as they can because you can kind of say here's a whole bunch of things i want there are places in the plotting process where you're like reading from one or two files simultaneously and writing out to another one at the same time so so the drive is going to have to be flipping between reads and writes there and some some drives like the data center drives that are mixed they're tuned for mixed workflow performance will handle that much better than like consumer drives which are much more read oriented and have caching algorithms that are not meant for like heavy sustained right workloads yeah yeah the other thing you know for to take advantage of ssd performance um ssds are inherently parallel you have a ssd controller you have a bunch of dye from the nandi that are basically in packages on the board so the way ssds get faster is you do reads to multiple man die at once so the higher the q depth the more parallelism within the drive the higher the performance and this is why you see when you have increased plotting you know processes you know when you have one you know you're not going to see a huge difference between a sata drive and an nvme when you you throw 15 plotting processes at a drive you know if you had a high capacity sata drive it would be severely bottlenecked by the interface where an nvme drive you must see how much higher bandwidth because it scales by the number of q depths so yeah there there's definitely you know in that question you know that's another question we get a lot about set of versus nvme is all about the size of the drive like as xornax said you know on a 480 gigabyte drive it's not gonna matter right you're doing one plot you know the plotting algorithm is not you know i o bound but on if you're doing 16 flats to a drive and it's an eight terabyte drive you're absolutely going to be i o bound if you're on a sata interface there's other weird hinting that like onedrive won't let you do but if you have multiple drives because we have this like multiple passes thing we could implement it so that you the workload would oh we it would plot the balance between two drives and the workload would be it's always doing a whole lot of reading from one and a whole lot of writing to another at the same time uh that could potentially optimize better maybe yeah yeah the other thing you know is the whole staggering thing you know because it's only phase one that's multi-threaded you know people have tried to figure out the best ways to basically get maxi utilization of the proc you know the plotting process by running high number of threads on phase one when that's done starting off another process um and then kind of escalating from there until you have it filled up but it's really that's really not the best way to approach it right if you had other threads multi-threaded then you would not need as much temporary storage you need a much more smaller capacity ssd if you had other phases multi-threaded um yeah we don't really know how to make our other phases multi-threaded there's less brute computational work they're doing but they are doing like a lot of sorting and stuff one of the biggest issues is that bucket sort thing i said it's hard to parallelize that but that thing running mono threaded is faster than multi-threaded normal sort so so that's what we're doing so let's talk a little bit about economics um just for simplicity's sake let's kind of throw out that you've got hardware that exists and i'd pause and say start with the hardware that exists when you're trying to learn to plot don't don't go out and buy stuff yet but you know give people a sense of like how much do you need to spend to get to what like terabyte output per day we if you guys are following the chat i i just built up so i just moved into my new house so i i finally have a garage where i can build stuff again i had this city in my garage which was a just a i a 10th gen core i7 an i7 10 700 i cheated because i get it from intel for half price so i so i a little cheating but you know say this is like 250 bucks on micro center um this thing you know with just a four terabyte ssd can do you know three terabytes a day of plotting and that was fully optimized that's and that's pretty good that's like a sub thousand dollar machine um my highest rated machine if you go to the the wiki page where you posted all our machines you know i have a dual 18 core server with 256 gigabytes of dram and you know massive amount of ssds in there that's doing you know seven terabytes a day so it's hard right like you can get better economics by scaling a lot of desktops but servers they're in a rack they're nicely organized they can all be automated you can you know remote desktop into them it's much easier to manage you know servers you ran 24 7. um you know you can get very good economics on a desktop it's just harder to manage like many of them and if you're going for servers used servers makes so much sense yeah you get a used uh dell 730 or a gen 9 hp and 24 plots going in the time 48 plots going at a time it's you're not as fast as anyone else's but you get that i mean yeah just so uh to iterate on that you know that the server refresh cycle is happening soon intel and amd are releasing you know new server cpus this year and a lot of people will be getting rid of their existing hardware so there'll be a bunch of you know xeon v3 and xeon v4 flooded in the market this year from people selling them there i mean you can already get a you can get a 14 core um you know xeon v4 on ebay for 120 bucks and so if you get an older server and you're comfortable building you know know an older server you can build a very high performance body machine out of an old server um you know either by you know buying a used l1 or you know building your own based off super micro or other components but yeah there's really not one size fits all solution for this plotting problem yeah i mean even the company we're buying used enterprise pull out hard work because it's like we want to have a test net support of 500 terabytes but why spend a bunch of money well that 14 core cpu when it launched was three thousand dollars and now it's 130 bucks on ebay or you know whatever it is right uh we converted you to the colton super michael microgene yeah we're going use super micro just you know straight up uh storage storage jm hooked this up so uh we're awesome yeah yes good i'm gonna be racking hardware too you know i have servers that i bought like the intel you know intel reference platform servers and i have ones that i bought from super micro and built up and you know when you're building servers just things are weird like you have to see the cpus right and sometimes you know the dram slots on my micro board like you know two of the slots didn't work with one vram but they liked this other one and you know so it's just you know desktop a lot of people are comfortable building desktops so in the chat i'm like don't go out and buy a server if you've never built a desktop like if you're comfortable building desktops and you can you know build desktops back and forth like maybe you know that'd be time to go build a server from scratch but it is a little trickier you know for sure so i have two quick questions and then a much more involved question that's related question number one is graham everybody wants to know what you're drinking oh uh you know i don't have well i don't drink very often so i don't have a lot of alcohol around here so i i was a minute late because i was scrounging for stuff so i uh right now i'm drinking uh rye with vitamin water as a mixer so okay and re not related but related why would you plot larger than k32 [Music] um i hate you'll answer if you'd like yeah yeah yeah yeah so a lot of what has to do with uh support for the future that we have formats that are supposed to last until the end of time so i could envision a situation where at some point in the future you have like a three-dimensional crystal storage uh on your home machine that has i don't know a zettabyte of capacity or something like that and the sheer number of uh reads that are going to have to happen uh if that thing is all split into k32s is going to kill it um so uh i i'm not particularly advocating anyone do greater than k32s uh just yet we did manage with the new uh consensus algorithm to really really limit the need for anything larger than that um but you know decades or centuries in the future who knows and so that actually reads directly into the last question which was like how and why would we make a decision to not have k32s oh um yeah that basically if i if grinding becomes economical right so there's kind of everything is supposed to bottleneck on uh on cpu or not cpu on a disk that you're just supposed to have you know normal hard drives sitting around so things we don't want to have happen are people getting a whole bunch of tape drives and and using those uh because those are just those wouldn't really be very useful for anything else which is not what we want to have happen um and then uh we also don't want uh someone having a rig which instead of having a whole bunch of plots sitting on drives uh regenerates a whole bunch of plots really really fast whenever uh challenges come in now we're quite a few orders of magnitude away from that being economical right now which is a good thing um uh but in principle if that does become economical uh we might uh crank up the minimum k uh in order to stop that from happening uh at the moment given the way everything has planned out we're basically anticipating unless someone has a differing opinion among you guys my expectation is people just kind of stick with k32s for the next couple of decades my big question is like so i think the way i understand it is you have to complete all of phase one before you know if a plot is going to pass the filter yes right yeah okay so basically you're you're going to your bottleneck is going to be how many cores can you throw at phase one and if you can complete that signage point well no no no no no no you know no no you know no there's this factor of 512 filter which you can check immediately up front so so that was a direct trade-off that's a direct trade-off that um they uh it's a linear thing where the fewer lookups you have to do the more advantage someone trying to grind is right so we had to trade that yeah but then there's just that factor but if you want to know the actual quality uh of your thing where if it does pass that filter that just gets you a factor of like 512. if it does pass that filter you then need to basically create the whole plot all of phase one yeah to find out if the actual responses you get are of any reasonable quality before the signage point completes yes the thirty seconds okay i get it yeah it's really before not before the signage point completes before the infusion point happens so that's like a you know 30 40 second okay yeah really it's a really hard computational problem to crack if you're trying to attack if you're trying to be an attacker yeah yeah yeah we feel pretty good about it right now and that's partially because i spent most of last year on our new consensus this was one of the things we were optimizing for a lot of really really complicated trade-offs there i think we've hit a pretty good sweet spot for them okay so uh here's a fun one uh what's the difference between a tib and a tb oh man tebby biter everyone yeah so yeah it's funny people thought as simple as like oh and you know an eight terabyte drive you know is you know you know like actually eight terabytes but it's not the way the way that these device manufacturers do the label there's two specs there's one called idema but basically they try to make the number of lbas 4k aligned the file systems don't just throw up and so it's not exactly a terabyte but the big change is you know device manufacturers list the drives in terabytes and you'll see a mac os and other operating systems they use you know in you know decimal form for the uh terminology but in many operating systems like linux and windows they they use gib and tib so gibby bytes and tebby bytes uh which is 10 you know 1024 divided by not by a thousand it's not buying it not but not decimal it's binary and so i have a nice little calculator i put up on the uh the wiki page um that has for every capacity drive what the actual capacity is going to be in the operating system how many plots you can fit on it for temporary storage we've done all the math it's it's somewhere up there so uh you know one of the questions i want to ask all of y'all but first i'll preface it with everyone here placed very highly in multiple of the farming competitions so they have proven that they have that space kind of indirectly um who here is a little sad they overspent on plotting and didn't buy enough storage well now that we're racing to mainnet uh you know i'm happy i have the extra plotting power but yeah i already had the storage so it wasn't a problem well we'll just say i was we were all planning for a december launch so we were had everything set up perfectly and like the number of terabytes and rack i had my rack perfectly fit where every single dry was filled full and everything dialed in and i'm like oh now i'm gonna have to rip whole thing apart but no we're all we're all upset i do have the t-shirt jean somewhere somewhere packed up in my boxes here when i'm in the mood for my farming og t-shirt yeah sorry everyone we didn't manage to launch mainnet in december uh in the interest of getting mainnet out as quickly as possible we're actually going to be launching um well we just to get liquidity in the market we had already decided we were going to have a period of time uh with no transactions uh what we're doing actually is making it so that there's um so that the rules for transactions aren't going to be in yet when mainnet launches uh and we're going to soft fork in uh the rules for how chilis works uh uh before transactions start uh so it's not that we are going to start work on that we're almost done with work on that but it's not finished finished yet um so that's kind of buying us some time and allowing us to get uh actual farming started earlier so we're trying to get that as quickly as we can yeah now we're actually a pretty good place because we've made that decision we uh you know lowers the total risk that we have to launch with everything bolted and all the things we already know the work so we do have a few more things you can always look on trello which is on the uh github project tab and you can see exactly what we're looking at to get things done and going well you guys got a whole lot of things in that list a lot of exciting new stuff for chia and in the works some of those things are like three or one so it's not as bad as it looks but uh let me ask this what are you plotter guys a lot of guys have to ask for bram and i yeah i mean the one the one thing for you know we kind of touched on it was like the the plotting format you know just as being somebody who's you know dealt with high performance ssds and stuff it seems like most people storing high performance storage you know do it in a database format so how did you guys end up on the final like plot file universe doing in a database versus you know complexity yeah so it's it's a flat file right it's a well it's a database and since it has its own designated format it's all about getting it as compact as you possibly can um and a lot of the work on it i mean this might sound a little silly especially given what we were just saying about k32s but um a lot of the work really before our first before we did our uh plotting our plotting implementation competition was on getting the format good uh and there's this subtlety where there's like back pointers between tables and those were like using up like eight bits per entry and we got that down to basically half a bit per entry by algorithmic stuff the um uh yeah what we have and that actually impacts consensus because there's a question of how much credit you get for the k that you're using and that should be fair across different k's and so that value impacts that so it turns out it's k plus one-half is the value that's used there anyway a long time spent on that formula um uh the um it really is just as compact as it can possibly be it's nothing even vaguely resembling like some general purpose format or something it's like super bit level squeezed down into nothing uh kind of stuff and like we've had notes about how well if you did blah you could make these more compact to the tune of like a tenth of a percent there's a few notes that we put in about things like that um the one big one would be if you did a helmet attack on it which could probably basically eliminate your first table um from it it just comes compactify your your first table down to nothing from the finished format um at the expense of let's say running an basically doubling your plot time probably something vaguely like that uh but that will be doable it'll be possible to do that for a plot file that for a plot that you've already finished so if you finished a plot today and uh you want to compact a fight further when someone's implemented helmet attacks uh then you'll be able to do it then uh we actually did implement helmet attacks but the way we implemented them uh totally fails um unless you have like an amount of memory equal to like the entirety of uh table one basically um so more algorithmic magic needs to be done in order to do that is like not multiple best so they recreate some of the magic in table one with the other with the other tables or something no no it's just table one uh it just uh compactifies table one by itself there's this trade-off where in exchange for using more computational power you can have less space so like doing something where you have to do like a thousand operations instead of one in your cpu and getting like a factor of a thousand in table one that's clearly worth it right you just eliminated your table one um but as soon as you go for more tables that trade-off becomes much much less favorable it's now compactifying more tables but it algorithmically doesn't work so well uh so basically the optimal way of using it in practice is probably going to be somewhere between one and three tables that you do it on um [Music] uh definitely you could use it to almost completely get rid of table one and that's a real optimization um two and three it's dubious the first four is just not gonna happen on no way um uh yeah so so that's something we'll work on you know eventually uh the next up on the how could you improve the format to get it even more compact thing is uh there are these things where sometimes you have something in one table that has children but no grandchildren and so it's silly to just put the child in the next table you can just inline it into the table after that uh that can get you like two percent uh in principle sort of that's actually a lot more work than even a helmet attack to implement so that's way down the list of stuff for us to work on um eventually um but in the vein of these things the format is uh basically compactified almost as well as we could possibly make it be compactified uh we had gave you know a whole list of known potential uh ways of compressing it further and one or two theoretical we're unsure of whether there is actually a way of getting this last half a percent even in principle using this known thing uh in a few places in it um given what a complete brick wall this has been to everybody else maybe we've been a little bit obsessing on it ahead of the game um uh but yes uh it it's all about how can we get this squeezed into as small of a space as it possibly can so everyone's on an even footing here and no one has an advantage just because they have better algorithms for doing this and it's still not at its terminal point we know that at a minimum there's uh doing doing helmet attacking on it uh but uh yeah so just just fyi at some point in the future you might want to take like most of your plots and do this other operation to compactify them further might be a thing that happens and so if you ever wanted to wonder why it took a little bit of time for this to launch there's a hell of a lot of thought that had to get done and tested and actually implemented uh if you're slowing up an ico it doesn't have much work in it you get it done in a few months and if you're trying to figure out how to make sure that you're actually future proofed against the entirety of the pc architecture infrastructure well you got to think about it a little um there are a bunch of great questions about stuff that's not plotting i apologize if i don't get to those we'll get to those in other cocktail hours we're trying to stick to the plotting i think a good question came in on plotting um you know what's a good starter resource for server grade configurations uh you know is there something like basement data center for dummies the plotting hardware channel i know we we've been trying to put as much as we can in the wiki but you know there are some good resources like uh you know there's like some youtube channels like serve the home and stuff that have like good resources on servers and how-to guys and one-on-one there's one called like craft computing or something like that on youtube there's some good stuff but yeah it's funny i see all these questions about hey can i take this brand new nvme drive and put on an eight-year-old server and it's like well the first nvme drive launched in 2014 so no it's probably not going to work you know you know i could as somebody who ran validation and testing for intel you know i can tell you probability that that works is not not good you know so i would start out with like uh you know if you're buying a server generally these server guys qualify like certain components in the server specifically for compatibility when you put it in it's going to work like dell spends a meticulous amount of time doing that like when you buy a dell ssd or something like that you put an intel server it gets all the functionality and it works so um yeah mix and match you know when you start doing that stuff you're gonna see bugs and you'll see issues and weird stuff but that's how you learn you just you start with something and you start playing around then you'll you'll get better as you start hacking away at the hardware you know i mean i try to just stay away from rebranded stuff i bought of ebay so i wouldn't buy a rebranded intel ssd rebranded by dell rebounded by hp it's like that they'll rebrand at once you only can upgrade the firmware if it's installing a dell server and things like this it's just very annoying so kiwi yeah people ask me about the the vendors and whatnot you know i'll give you my my real answer which is go with the most popular guys you know i i can you know vouch for intel stuff because i i know it but you know samsung and other guys you know the ones who are major major deployers of flash you know generally by them versus some like little tiny guy i can tell you just based on the you know the comments and the firmware about the terabytes written and some of the some of the way these consumer guys manage their firmware it's just abysmally bad like i mean your 3710 is just the best sata drive ever made idiot michael's a a big proponent of my uh one of the one of the products i worked on back you can't destroy it i mean i think farming i fried one on i one i have a question related to actually who's doing uh fluorocarbon cooling on their things like are people doing that no no yeah well gene you're we're doing it so why don't you tell them that we're heading there but that's what time war is not for plotting oh that's right time right yeah the uh the plotting process is pretty frequency uh sensitive right so if you have a you know a 10 year old cpu that's plotting at two gigahertz versus you know a brand new amd or intel that's body you know 4.5 to 5 gigahertz you know you're going to see a massive difference in plotting speed but that's i think that that's the thing that a lot of users don't realize which is there is the core clock and then there is the boost clock and if you're if you're plotting one plot you're plotting that at the boost clock all the time and if you start to load your system you're plotting at core clock which is often 25 or or even below the boost clock and that is what's causing a lot of the slow down and you can't just blame it all on your desks it's really the cpu it's running a lot slower so i was going to say h and it's somewhat related uh you know almost everybody thinks you got to use ssd you know we talked about a little bit but can you tell us more about like what you're getting away with with not using ssd yeah i've got a couple of plotters uh one runs just a standard two and a half inch cssds um it's obviously not as fast as sorry hdds uh they're not as fast as the ssds but 24 plots of the time and uh they carry on nicely because this i find it a very interesting topic so i had rented the server a couple weeks ago cloud server by uh used to be packet now it's a equinix metal there's a i think there were eight or six i think eight eight hard disks in there you get that you get away loading a system fully with 30 000 seconds per plot on hdd sata hdd that's not bad yeah you guys that that 3x speed up that you guys did on the plotting algorithm you know basically enabled some low end hard drives like before you know in the alpha there was no plotting on hard drives it would have been you know 10x 20x the the times for plotting you know so it's it's good that the work that you guys have done has really helped you know low-cost hardware guys at least get something to be able to get plotted you know obviously ssd is going to be faster but you know being able to at least do some plots on hard drive is very pretty good yeah i think there was a moment where you got treated you got rid of millions of seeks i believe that's that was really uh very good for the hard drives oh it wasn't it wasn't it wasn't seeks its number of passes over the data that it's doing uh it was dramatically lowered uh well from a monitoring point of view i i saw sikhs millions of sikhs yeah because it was doing the read again right so he's taking a pass writing yeah it wouldn't that that's interesting uh that you say that i mean we did a number of uh improvements uh mariano got rid of those i remember the conversation that we had yeah if miranda's asleep otherwise you'd probably be on here right now [Laughter] uh let's see um let's talk about the uh the flags what's bitfield what's buckets what's stripe size what should i do is always 128 do not change it always set it stripe size is always 64 whatever it is yeah actually mariano is that how you pronounce his name yeah we're talking about the bucket so ssds are are good at random i o right that's the uh it's it's when you in flash is when you you know the core bottleneck of flash is to write two ssds you have to erase the block before you write it you program in pages racing blocks and there's this whole garbage collection process and stuff but you know in a race block is you know anywhere from like you know we'll say 50 to 100 megabytes and these these writes the bucket sizes you know 128 bucket sizes like seven megabyte files so the difference between like three megabyte file and a seven megabyte file to an ssd is nothing it means it's literally nothing we write in 16k pages it doesn't matter you know so the more buckets actually the better and uh for ssds you know the smaller you break the files into the i o and the less dram you use um so basically when you scale the buckets up you scale the dram use down and on ssds most mainstream ssds can handle almost any type of low block size like all the way down to 4k you know four kilobyte transactions so there's almost no reason as kiwi said to not use 120 buckets which is the max and the software today yeah you use the 128 brackets it's the minimum amount of ram you need you can run more plots in parallel and one of the things i found doing some testing is there's the optimal amount of memory you need to do all quick sort but you can go even lower than that if you want to do more plots in parallel and the penalty performance if you get right down to bare minimum ram it's like a 10 percent performance difference so if you really want to go lots and lots parallel yeah you can do a lot of quick sort you can basically double the number of plots you're doing at once so bram i think it's probably worthwhile to explain uh bit field versus non-bitfield yeah you know it was interesting that you said about ssds so ssds like when you write a small section has to like read the whole region it's in and then rewrite out that whole thing again with that small section change is that it's full right so it's like a parking lot when you know uh so if the drive is empty the algorithm is very easy the ssd finds the next free you know erase block and it writes the race the pages right there if the ssd is full that's where it gets really complicated it has to do a read modify write on every single write that comes in and in ssd world we call that right amplification and the higher the right amplification the worse the endurance because you write more to the flash and wear it out faster yeah it doesn't have good defrag from our nice clean just like right out whole sequential thing destroy the whole block of things at once so that's good so the more the more sequential the workload the lower the right amp is and this is another one where in the chat people often look at this terabyte written on the drive and say okay this is the math of when my drive's gonna die and that enterprise drives that's not how it works they actually you have a number of program erase cycles for every nandi that you can program and erase the cell before it wears out and depending on your workload whether it's sequential or random and how full the drive is for whether you have trim or trim trim enabled or not enabled that determines the right amp factor which determines your endurance which determines that the wear out number of plots you can get per drive before it wears out uh okay so that's really interesting a lot yet yet even more deep magic what's going to determine performance here um uh so the question of what does bitfield do so uh basically um the format from an algorithmic standpoint what's going on with the format is each table uh each entry into it has like two pointers into the table before it and so when a challenge comes in the response to the challenge is actually 64 things so it like doubles every table that it goes through uh the stuff that it's getting uh sometimes uh there are things that have no children and there are sometimes things that have children but no grandchildren or grandchildren but no great grandchildren obviously the probability of this goes down with each generation you get to um but you definitely get um the the plotting process in a nutshell is it generates everything going forward in its first pass it then back propagates the non-existence of children all the way up to the very first generation in table one uh and then finally it does a compression step running forwards across the whole thing and some aspects of these are kind of munged together uh what it's doing on a bit field is instead of like sorting all the things to be in order of what actually gets used it just generates a bit field of everything from the previous table that was ever accessed so that's really really tiny because it's one bit per entry instead of you know k bits per entry um or more than k bits per entry um and uh yeah so that can generally be kept in memory uh depending on stuff uh and if you can keep it in memory it saves you a pass over a whole bunch of stuff uh and gets you some performance assuming that your random accesses to check the bits are fast which they should be if it's in memory but that's also complicated because relative to cpu memory lookups are expensive so again all kinds of crazy complicated stuff going on but that's what that optimization does it saves you passes over disk it's not a huge optimization just because it's not it turns out it's only a small fraction of all the operations it's doing that it impacts at all uh so it doesn't make a huge difference there so zero knocks what's that ended up looking like in practical day-to-day yeah i i see so on my testing i did with that with that rented service using https only it's a good 10 15 percent better oh wow using bait field yeah so it's like it's like a 34 i i can i can post if i find time to to to present this data nicely i can post a summary but i think if i don't mess it up in my head it's like 34 ish thousand versus thirty two thousand seconds per plot yeah our our implementation is probably doing a much better job of optimizing for hd than for sd um just because that's kind of the mental model i was working off of and sort of the most important target that i've been thinking about is just some consumer person wants to set up a plot on their hard drive and they run it so that that's a lot of what the stuff is optimized for it's also the easiest to think about yeah i also want to add on the practicalities that um you'll find that using bit field sort is going to take a little more memory than using non-bit filled sort and so the trade-off and parallelism may be very different uh i think kiwi wanted to add that yeah that field requires about 75 percent more ram so if the trade-offs worth it go for it uh if you want to go for lots of parallel yeah turn off midfield yeah it actually requires an amount of ram that's linear on the size of the plot but since everyone's using k32 this is a constant so i think the thing that most of us i think saw is that enabling bit field using any flash drive it makes it a little slower yeah but it might it might not be really a bit filled but i think this is an area to look into yeah it didn't adjust ram with iso memory it's about a 20 to 30 percent worse you're plotting time uh on ssd with midfield enabled versus without oh it's also going to reduce wear on your sd if you use bit fields yeah it's about 12 less total rights yeah that's slow yep so there's trade-offs everywhere we're not really sure like crazy things happen when you're doing sd because um sometimes the bottleneck will be cpu rather than disk um if you're doing it so it wouldn't shock me if it turns out that bitfield is worse for sd because you're doing random access lookups in memory and random x's lookups in memory are actually slower than continuous like throughput on sequential things with your sd that's entirely possible yep so we're kind of running out of time i'm gonna go a little over everybody's okay with that um there's three different things i'd like to do the first one is bram can you share you know why we ask why we say someone might want to delete a plot and how we're thinking about that right now oh oh yeah so this is a little bit it's really ridiculous the skull duck array that are is part of the everyday when you're designing consensus algorithms you really think about like insane crazy things um uh so there is this in principle attack where uh if you have like a network where it's like so dominated by a small number of whales that there are surprisingly long periods of like hours where if you knew like all the keys of all those whales you could do a foliage reorg of that so uh in bitcoin uh when you make um uh when you make a block once it's buried it's really buried uh when you're doing proofs of space there's really two kinds of being buried um on the one hand something can be buried uh in that it's impractical to [Music] um to change who the plot winners are going that far back uh but then uh in the interest of stopping grinding all the malleable things aren't factored into the upcoming challenges they're signed off on by the people who won the previous ones so if you get a cabal of these signers who all get together for the past whatever they can just make up a new version of the history up to the point where they don't have the keys anymore uh so it's not unreasonable to have a policy to avoid anyone getting any funny ideas about putting a gun to your head and trying to make this happen now how ridiculous and crazy this is and whether this sounds like me being paranoid i don't know but uh if you delete the plot if you delete winning plots they're just gone they're they're toast they're they no longer exist um and uh so particularly for plots that win a transaction if you're a big enough whale it's kind of a i'm saying best practice it's nice it's polite it's a good idea um uh if you have like thousands or millions of plots that if a plot wins uh you then delete it and replace it with another plot uh afterwards to make sure that the the keys for that are well and truly deleted and there's just no way anyone in the universe could possibly reorg uh past that depth uh in the foliage if you're not a huge whale don't worry about it even today if you are a huge whale don't worry about it too much but it's it's polite it's a good idea at some point in the future to do this so with that uh let's run around the the panelists you know is there anything you think we didn't cover well enough i know for example derek that there's questions a little bit more about your uh your strategy so uh let's start with h and go to derek and then zoranox and then storage jam um i've got nothing off hand um mental blank all good jack what was the question about my strategy yeah people really like to know more about your kind of you know small cheap multiple computer strategy yeah okay so i've actually built my first server standing next to me i had to try it right it's an old 2014 dual xeon ddr3 it's it's here so i'm i'm gonna try it but you know when when chia started off in the early days there was no there was no real possibility of doing all these um parallel plotting and all that the early days plotting was was was really tough right so you were just grinding out k34 k35 k36 and uh so that i i stuck to that hardware but then for me i just found it i just find it easier to use than struggling with large machine that is plotting a lot of things in parallel i i get less i don't plot very fast but it's very consistent and it's very user friendly i have i don't have any errors i don't have any downtime you know i can plot less i can plot more you know so i can manage the the the energy usage much better i don't have to run you know a large machine large cpu uh energy consuming server i can manage it that way so my computer is very simple they are four port to six core processors each each computer they are not in they are open open frame mount there's no boxes or nothing like that each one has got a small little uh bay with that it's got a 35 lsi 924 9240 the rate card on it and the machine boots for one terabyte ssd micron and then on the right card i have uh two three seven ten intel enterprise drives and four four terabyte sas drives so what i found early on is there's a lot of time that's been wasted you know and actually moving the file to the destination so i plot two to three files and then the destination is a raid zero of four sas drives right this gives a total of 16 terabytes so the file copies much faster it sounds insignificant but when you plot 10 000 plots if you accumulate all the time that the thing takes to copy it 100 megabytes per second it actually a lot of time lost there so mine moves into that and then i would take a 16 terabyte you know farm drive and i will copy the files over to that while the thing is plotting and move that drive then manually into the farmers and that's why that's how i do it and i can plot 16 to 20 terabytes a day doing that my farm my farmers are straight rippers x399 uh with uh 9 240 cards flashed with um the 920 firmware so it's just normal hba no doesn't have to mega rate on it and it's just plug and play all the farmers are custom built right they don't i i'm sitting in the room now there's 1.4 petabytes us running here and there's no noise they're all passively cooled they are basically noise free that's awesome so we're knocks anything we thought we didn't catch on don't go and just buy hardware just really unless you unless you don't have just anything um i i know from a user i want his closest name because i haven't discussed this but he's we have chatted a lot and he has uh he has also done a very similar strategy to you derek ended up with dozens of small machines so if you're gonna buy i i've started my first plotting on intel knox that i still had like four or five year old knocks there was already a a pci version 2 nvme drive in there and it was it was it was it was fine it was almost a terabyte maybe 0.7 terabytes a day with the latest versions a lot slower obviously with early versions um i have an old 4770k that was a really i think that was maybe the best intel cpu that ever was that i bought in 2013. it really didn't get much faster since then from a single core point of view if you buy a cpu worry less about the cores per core you need to have io so it doesn't help you if you have 16 cores but only three nvme flash drives in there you can't utilize all these cores uh worry more about the best single core performance that you can get i have a core cpus mostly i'm much happier with that they clock a lot uh higher uh the the most decent 58x that i used to crack the 12 000 k seconds uh mark it's a little below that you can over i thought i've overclocked that cpu if we have regular corsair uh liquid cooler to uh what was it 4.7 gigahertz all cores much happier with this if your desktop and you know what that computer if you're done plotting you can use it as a gaming machine or office computer or you can actually resell it pretty well i think that is maybe one of the most important reasons for me to not buy server gear i have a bit of a fear if i buy five-year-old server hardware or so i'm not going to be able to resell this if i don't use it anymore yeah john michael yeah no we see a lot of questions in the in the chat and key base of like okay the people are trying to get like the perfect config before they start plotting like like these guys said just like grab anything and start plotting and learn the process learn the tool you know you don't have to have like high-end stuff to start it and then once you have an understanding how it all works you can start putting it together like theoretically the best plotter would be you have super high frequency single cores you have lots of cores you have lots of dram lots of temp space and you have complete parallel processes and you know we've tried to do a good job of posting all of our machines on the wikis we have a couple google cheats pages with all the machines and times you know we're trying to make it very open to like you know figure out you know what we're all doing just so everybody kind of knows what to build and where we're at and so yeah feel free we're all up on key base pretty awesome i mean maybe maybe one one last comment here as fast as plotting is it's still pretty slow so it's better to start plotting slow versus just waiting around not plotting because the one thing you cannot buy is time that's right well also a lot of plotting is about cost and you know the cheapest stuff is what you've already got sitting around so yeah yeah so for everybody to know when you're watching the replay this especially uh we're almost all here somebody's one of us is usually there 24 7 365 effectively and the few people who aren't on this call are awake on different time zones as well so you've almost always got a plotting expert somewhere around key base um so everybody knows uh hamish is kiwi h uh derek is farming 101. michelle is john michael hands is storage jm and you know me and brown um and i've got my uh penultimate question and i told everybody i told bram about this as we were pulling together and he's like oh i have questions i told all the panelists that and they went that's not scary so bram do you have any other questions yeah i do well my main thing is this sort of uh uh i mean people have said a whole bunch of really interesting indicative things about um performance on sds and stuff i actually everyone has observed i get really excited when i talk about uh plotting optimizations but um but uh so all kinds of exciting stuff there although unfortunately i'm not going to be working on that again anytime soon the big uh question observation kind of thing i have is i notice everyone's here everyone here is talking about uh like really acquiring and plotting on uh drives uh and no one is talking about but no one is or no one is owning up to uh doing it on storage media that they have for some other reason we do know there are many many enterprise places that uh have lots and lots of over provision storage which might if they start getting in the game that could really you know wreck the work difficulty on our network we we make no promises about what's going to happen in the future um so a series of questions related to that do people know of that happening or have any idea of places that might do that or why we're not talking we're not seeing anyone talking about doing that now and if that might happen might some of people with plotting rigs uh have plans to just like bring their plotting rig over to someone else's thing and plug it into a wall and spit out uh uh the pots there i mean i could definitely see um places there being places who have over provision storage either shall we say officially or unofficially starting doing farming on a whole bunch of that curious how that might start happening and kiwi you're completely allowed to do an advertisement here no no i've got nothing to advertise yes service does exist that is legit and you can securely offer plotting as a service and i do yeah and remember the key piece to plotting as a service is that the ups truck has tremendous bandwidth oh yes um yeah and we you know honestly we probably would have seen more people asking questions in the chat if there were other large you know people that already have storage trying to try to do some plotting i think it's very easy once the economics happen once the mainnet launches once it's trading on an exchange and people can say oh i can plot this much and make this much money people think in very total cost of ownership like ways they're going to say okay here's my investment of cpu and i can this i'm going to get it back in coins that's great let's do it but right now it's so you know people are in the early investment stage so it's hard to like materialize some of that you know excitement you know which is exciting for us the early adopters but you know the the people that are you know haven't heard about it yet are going to be much more exciting when they have a price tag on it you know so we've gone about 20 minutes over and i want to leave it a little short just because the replay is something i think i want to make make easy for people to go watch and catch and some of the best content unfortunately is here at the end but my ultimate question for everybody in the room is and you know i will say it this way everyone here has um better bayesian priors than others how much net space will there be 24 hours after maintenance launch and i'll tell you that we're thinking about launching main at about 6 a.m pacific time so they can be pretty fair around the globe and everybody's you know able to be awake we're sorry australia that it'll be uam but so uh we'll start with h we'll go then to derrick then to zernox and then john michael my instrument has always been a hundred terabytes plus i was 50 to 60 products 350 350. wow there's under no understanding below 50 you you have fixed the puck that estimated 2x the amount so i think um very short comment people throw around petabytes like terabytes it's not like that it's a thousand times more okay so i i apologize what what is your actual prediction again 50 less than 50 petabyte less than 50 better right it's great and storage oh uh i'll take my my best sports betting like i'll take the over on 60. uh i'll take 80. so so i admit i'm still thinking it's like 60-ish yeah i i have no idea i don't think in constant values um the uh the thing i'm really curious about is whether there's an immediate bump on mainnet launch if if people are going to have some amount of storage that they're just not attaching to test net that immediately shows up when mainnet launches i think we can predict that that one's definitely true some amount right but yeah you know they're it's risky to just not test it so yeah we we've been all monitoring every every single launch to try to see what the spikes are and meticulously try to guess what yeah and it looks like the blockchain has now got no point where it should be moving at normal speed so farming right now it makes life easier for us if uh there isn't a big bump on uh mainnet launch our our plan is to make it so that the initial work difficulty for me net launch is set at you know somewhere between 1 and 2x what testnet has stabilized on at that point and let it settle out reasonably quickly we'll see how that goes partially a bit of an uh comparison i use is the the the 10 900k the the intel cpu right uh it does four-ish terabytes a day so it takes it takes 250 days for that cpu to put the petabyte it to put 30 more petabytes you need 30 of these machines doing this for 250 days and um it's a lot of compute it's really a lot of compute you can have 10 20 of these you're gonna deal with heat problems in your house that some users can talk about um it's not easy yeah well as a comparison i used to do cold coins bhd i could plot a hundred terabytes in a day uh the bhd network had 800 petabytes yeah that's uh yep well gentlemen look thank you go ahead sorry i i think that the despite i think the surprise is going to come from not the the whales or the large players gonna come from the hundred terabyte cars yeah possibly yes there's a lot of one terabyte guys that's not farming on the on the on the destination i plan on my 40 showing up hey gentlemen thank you all for taking some time out on this afternoon we'll do this again on a different topic but it's been a blast uh it's time for a refill yes oh it is yep again it's uh february 19th uh it's about 5 25. any question we didn't get to happy to answer them over on testnet if they're uh support related plotting related plotting hardware is kind of the advanced channel for plotting uh if it's more about energy consumption or what do you think about the project general is actually the right place for those kinds of questions and uh many of us will be online through the evening so thanks everybody and thank you guys for helping out graham gene ellis thanks a lot graeme thanks everybody take it easy

Channel: Gene Hoffman Jr

Views: 20,649

Rating: 4.9778395 out of 5

Keywords:

Id: wDVsZMDlQYw

Length: 82min 40sec (4960 seconds)

Published: Fri Feb 19 2021