Don’t need Flash, if the new plotting software is set up correctly.
That was one of the early selling points for NoSSD.
However, not being able to RAM plot on REASONABLY INEXPENSIVE hardware is a game killer for the whole idea of forcing K35 size plots as a mininum, and DIRECTLY CONTRADICTS the stated Chia objectives.
2 Likes
Go read the post fully. The goal is to make plotting harder not easy.
When you make it SO MUCH HARDER that only big farmers can afford the hardware to plot at all, that’s a TOTALLY INVALID goal.
THAT is the point I’ve been trying to make.
1 Like
You can always plot on the cheapest hardware you can get a hold onto it. It would just take longer. Obviously a plotting machine that cost 10x would be expected to plot quite faster. That will never change.
Plotting was never unaffordable. I always plotted on hardware costed a few hundred bucks and so on. For some reason, everyone thought they need to get the latest threadripper and fastest nvme and they paid for them.
For Chia to get back to “Proof Of Space”, instead of also being “Proof Of Work”, then the harvester’s role needs to be increased in difficulty. That could be accomplished by increasing the minimum supported “K” size value.
My understanding is that when you go from a k32 compressed plot, to a k33 compressed plot (for a given compression level), you need ~double the VRAM in your GPU. That is, it takes double the VRAM for the GPU to complete its task, when the “K” size increments by 1.
So if k34 plots became the minimum size recognized by Chia, then that would do away with countless GPU based farms, and bring us back to standard, non compressed plots. Ergo, proof of space.
And k34 plots are not that difficult to create.
Creating k35 plots is where you would need north of 4TB of temp space, which would be a problem for a lot of people.
But the above would return Chia to Proof Of Space, until a time that GPUs can handle the larger “K” sizes and be affordable, which would bring us back to a combined proof of space and proof of work.
But a mandatory “K” size increase, combined with the scheduled plot filter halving, would result in proof of work becoming a daunting task, where proof of space would likely come out on top.
1 Like
And as k sizes increase, computation required to plot increases as well. I don’t know the formula but muxh more than exponentially.
I think 4tb flash is reasonably affordable, the real challenge would be the upfront cost of computation.
1 Like
Creating a non compressed k34 plot takes a little over 4x the time compared to creating a non compressed k32 plot (and the k34 file will be a little over 4x the size of the k32 file). It also takes more RAM. 64 GB is enough, along with a 2TB SSD (k34 requires approximately 1.25 TB of temp space).
k35 and k36 are also doable with 64 GB of RAM. But k35 requires more than 4 TB temp space, and k36 is probably close to 9 TB of temp space.
For kicks, I created a k36 plot:
I did not have enough RAM to try a k37, and it would probably have taken a fortnight to complete.
And although there are 8 TB SSDs (even 64 TB SSDs), they are prohibitively expensive. And I have no idea if anyone makes a fast one, that remains fast, for more than a small portion of the drive’s cache.
1 Like
This is the idea with higher k levels, it will take longer because it cannot be parallelized due to memory constraints.
You were able to get k34 with linear increase since it still fits into gpu memory and can be parallel processed as more cores available.
4tb flash easily affordable and you can get one for a couple of hundred dollars or just by 2x 2tb. I am pretty sure you can spend a lot more money though to increase the plotting rate by buying faster flash or more of them.
When Chia first came out, plotting was very slow. It would take several hours for a k32 to plot. It took months for netspace to reach to top and it would have taken longer if there was no Madmax. This barrier of entry helped early adopters who have been plotting since testnet. New comers had to wait for several weeks before reaching to peak. This is exactly what I am proposing to reward farmers for their risk taking and keep blockchain grain by high barrier to entry.
Ok you know everything - tastes salty to me. Be humbled to be early. If you think we are at the end of the road than well good riddance.
Seems like you are trolling. have fun
A 4 TB NVMe drive will not allow you to create any “k” value higher than a 2 TB NVMe drive will allow, which is k34.
When you go to k35, you need over 4 TB of temp space. That is where it gets expensive. That is where the unknown speed of the native NAND cells come in to play. All (to my knowledge) consumer level SSDs are comprised of fast NAND cells that initially get written to (acts as fast cache). When that cache gets filled, the SSD slows down. Depending on the SSD, that slowdown might be minimal, or might be drastically slower.
So to create a non compressed k35 plot, which 8 TB NVMe drive will not slow to a crawl? Which one will continue to write at 2+ GB/s throughout the entire plot creation?
That was before the madmax plotter. I believe that the initial, standard Chia plotter was single threaded. But due to being single threaded, it used less RAM and less CPU cycles. However, you were able to execute multiple, simultaneous, plot creations, giving you an overall, relatively fast plotting experience. If I am remembering correctly, I was running 8 concurrent k32 plotting jobs, in order to bottleneck my 5950X plotting computer, effectively reducing the average time for creating k32 plots to approximately 1/8th the time (since 8 ran simultaneously, without taking much longer to run).
Non compressed, k35 plot creation adds a new hurdle, due to the over 4 TB of temp space that is required, and we do not have a list (none that I know of) of 8 TB NVMe drives that are known to work speedily for the entire k35 plot creation operation. And 8 TB NVMe drives start north of $800.
It would do away with the common consumer based GPU farms, but it would open the door to SXM style GPU clusters which have the high bandwidth memory interconnects to continue doing tomorrow whats being done today if the original plot format remains, even with larger K sizes.
So while I get the concept you are going for, in the end equipment already exists to take advantage of that and would force out the small to medium (heck even many larger) operations, it would only now be in the hands of the large scale asic style farms you see in other coins that take a sizeable investment to participate in the network, I don’t think thats what you truly want.
This is why a new plot format is necessary, to correct the oversight in the first few plot tables when it comes to GPU compute and the methods currently used to take advantage of that. Currently its much easier to drop the beginning tables because they are weighted different in compute needed, by correcting that and making it difficult the entire way through the plot, advantages are mitigated.
1 Like
It seems that no matter the hurdles, someone is going to use a 256 core, 40 TB RAM, 192 PCIe slots (all at 16 GB/s), z16, IBM mainframe.
1 Like
yes someone may spend $100K and create plots at higher rate but that is the whole point, $100K is pretty high barrier of entry.
Plotting should not be easy so people who have already plotted has the advantage. If plotting is easy, netspace will be always close to the high point.
The crypto rat race has always been a competition of how can you allow consumer devices for wide participation without the immense advantages of specialty equipment few can afford, think of the evolution from one cpu one vote to GPU, FPGA, ASIC.
One of the ways was to use memory intense tricks so you needed a lot of memory and/or bandwidth, some variations range from algos like RandomX that need little memory but insane amounts of bandwidth and low latency ala L3 cache to favor CPUs, to various algos that want higher (sometimes increasing) memory size requirements plus bandwidth to favor GPUs.
The delicate balance becomes how do you create a method to utilize something that doesn’t “waste power” as its primary means of security, is something unlikely to have specialty equipment, and be what consumers can afford or already have (the “mining trilemma” if you will), thats where storage enters the picture to tackle all of those with a better balance overall.
The first plot format had an area that was vulnerable to one of these things, so now with that knowledge added and working with some of the people who were able to apply it, a new format will get flushed out so that we can continue using storage as the primary means to keep participation rates among the average consumer who ultimately is the best form of decentralization.
So keeping the smaller plot size (32 over 35) is important, using common consumer parts is important, keeping power costs down is important… in the future as nand takes over the power costs will be reduced (6 watt idle to 60 milliwatt idle) so the last thing you want is to be heavily GPU/ASIC focused by that point, so that the “waste power” is mitigated as much as possible in the future.
1 Like
power efficiency does not matter so much for total power consumption of block chain, right? ASICs are extremely efficient over GPUs for BTC mining but total power consumption got much higher than GPU era.
Total power consumption depends on two things: CapEx and OpEx. Keeping the CapEx (hardware, initiation and set up) high is the only factor to limit to OpEx , primarily driven by power cost.
This is all kind of an “it depends” thing with what the end goal is and who you want to participate.
Using everyday storage has the benefit of not just being accessible to the average consumer (compared to a $10,000 modern asic using 3000 watts each) but it weights towards capex rather than opex, opex will improve over time (with nand) opening up participation to more people long-term.
The wide gap in electrical costs can often be the difference to participation, you don’t see ASIC farms in high cost areas paying 0.40/kwh (compared to 0.04/kwh some have), but some storage farms cling on because the weight was on capex, a shift to nand storage weights to capex even more.
The last thing I want is for opex to remain important after nand because GPUs are competitively required, this will erode the future participation pool and leave a large segment of people unable to.
It’s not the Power Consumption or Equipment price that primarily matters. It’s where these market forces (ergo brainpower, innovation, capital) go. If demand for farming is high it would directly facilitate innovation in Storage which benefits everyone! A better BTC Asic benefits no one because we know how to make them better and more efficient - just ask chia. To make them better you just need a shitload more capital and at that point it becomes a gamble.
Not really trolling 
love the brainpower you put to it but asking myself the following: Do you really believe Bram hasn’t thought about ANYTHING u come up with (at least when it comes to something so close to his priorities) while having a toilet visit 6 months ago? I mean serious? Can u imagine all the programmatical, algorithmic and mathematical implications every decision takes? Is he trolling us? Sure he probably knew (or had an educated guess) what would happen with GPU farming and was more curious than worried. “Bram should … , chia should …” etc. is nearly always a self defeating sentence because u tell very smart fuckers something they already know. Thus it’s hard to take this serious. As a basic discussion about pro and cons why not. I just get allergic every time I read from people that seem to have all the answers.
I don’t make suggestion to Bram. I actually don’t even know making “k35” is the right way to achieve what I am proposing. I have so little understanding of Chia and PoST; I am not in a position to make any technical argument.
But I am proposing a direction to reward early adopters. This is of course Bram and anyone in XCH can think of but will not because they are not incentivized to support this. Their goal is to build a healthy network and they think this means increased participation.
The angle I am pitching to CNI, if you make the network to easy to join then it won’t be green which CNI also puts out as goal. If CNI wants to create a green network, it needs to make the network hard to join:
- Reward early adopters and primarily motivated by selfish reasons. I don’t want to be in this rat race as a farmer, constantly burning money to keep the farm running with no upside. I want the farm to return higher profits for a while if/when XCH goes up.
- Influence CNI and community to be truly a green network and don’t become another blockchain consuming power equivalent to Netherland, a highly developed country of 18M.
Perhaps K35 is a bad idea to achieve these goals:
My proposal is to erode future participation to pool so it rewards early adopters and stay green. As long as barrier to entry is low, network power consumption will be directly related to XCH value and its emission rate through mining.
1 Like