archive snowblossom slack general
prev 2018-10-28 next
2018-10-28 00:02:35
noogie
re SUB - dont forget https://handshake.org/ forging the way of the new decentralized interwebs Decentralized certificate authority and naming
2018-10-28 00:07:14
mjay
Also this one, even older than bitcoin: https://freenetproject.org/
2018-10-28 00:07:58
Fireduck
I remember messing with that and not being impressed
2018-10-28 00:08:36
mjay
it is very slow, forgets data quickly, but has the focus on anonymity
2018-10-28 00:08:50
mjay
a different use case
2018-10-28 00:13:12
noogie
for the sole reason of not accepting a host key again, I think HNS will succeed
2018-10-28 00:17:48
mjay
what is HNS?
2018-10-28 00:36:26
noogie
handshake
2018-10-28 00:37:00
noogie
faucet for verified FOSS devs = fucking rad distribution model
2018-10-28 05:43:22
alistar
how to config the Arktika's config files?
2018-10-28 05:43:24
alistar
threads?
2018-10-28 07:06:41
alistar
got
2018-10-28 11:34:51
finex
made an update @ snowblossom.satoshis.guru, node now working on the lastest release
2018-10-28 15:20:00
Fireduck
wow, almost 2GH/s
2018-10-28 15:24:45
alexaabb
will be pushed to 8:grinning:
2018-10-28 16:23:17
offmenu
people getting PoW fomo
2018-10-28 16:23:46
Fireduck
I fear tigers
2018-10-28 17:01:22
mjay
Are there stats on which miners are using field 8 already?
2018-10-28 17:01:32
mjay
I checked some blocks manually, all 7
2018-10-28 17:02:22
Fireduck
hum, I could write a script to check but I don't have a good tool at the moment
2018-10-28 17:02:35
mjay
I´ll write something then
2018-10-28 17:02:51
mjay
Its not too difficult I guess
2018-10-28 17:12:48
Fireduck
if you take the previous version of RichList.java, it iterated through all blocks
2018-10-28 17:12:53
Fireduck
you can just copy that
2018-10-28 17:13:05
Fireduck
in the most recent version I changed it to use utxo instead
2018-10-28 17:13:31
mjay
I´m using VoteTracker now, just the last 1k blocks are enough
2018-10-28 17:21:48
Fireduck
cool, good idea
2018-10-28 17:36:59
mjay
Field 7 997 Field 9 3
2018-10-28 17:37:54
Fireduck
Field 9!!
2018-10-28 17:39:47
Joko
Is that 128gb?
2018-10-28 17:39:53
mjay
512gb
2018-10-28 17:39:59
Joko
No way
2018-10-28 17:40:19
Joko
I still remember ram mining lol
2018-10-28 17:40:36
mjay
days of ram-mining are almost over
2018-10-28 17:43:27
Joko
Vps with 512gb are expensive af
2018-10-28 17:43:32
mjay
you could buy some fancy server hardware, but that gets more expensive than ssd´s
2018-10-28 17:44:03
Joko
I was one of the first 5 people on the network :grin:
2018-10-28 17:44:10
mjay
nice :smile:
2018-10-28 17:44:19
mjay
your first block?
2018-10-28 17:44:39
mjay
mine was around 6k, 64GB still
2018-10-28 17:45:55
Joko
Not sure which was my first block
2018-10-28 17:46:16
Joko
I remember 3 weeks later someone put 100x what I had on the network
2018-10-28 19:31:32
Rotonen
mainframes, i say
2018-10-28 19:33:59
Rotonen
i applaud whomever set their block remark to ´POOL´
2018-10-28 19:36:02
Clueless
ha
2018-10-28 19:39:21
Clueless
@Rotonen hey uh. at about `2 GH/s`, if the average person nets about `500,000 H/s`, doesn't that equate to like `4000 people` ?
2018-10-28 19:39:57
mjay
500kh/s are quite cheap on aws
2018-10-28 19:40:16
Clueless
how cheap?
2018-10-28 19:40:49
mjay
2$/day
2018-10-28 19:40:53
mjay
about
2018-10-28 19:46:11
Clueless
~@Rana Waleed~@mjay link me?
2018-10-28 19:46:44
Clueless
I haven't seen anyone that allows you to use that sort of bandwidth, io, cpu for that cheap
2018-10-28 19:50:56
Rana Waleed
hey whatsup @Clueless
2018-10-28 19:51:06
Clueless
sup
2018-10-28 19:51:09
Rotonen
what do you need bandwidth for?
2018-10-28 19:51:30
Rana Waleed
were you asking me something @Clueless
2018-10-28 19:51:49
Clueless
oh sorry
2018-10-28 20:13:43
Rotonen
@Clueless i’ll show you the opposite end http://www.cirrascale.com/pricing_power8BM.php
2018-10-28 20:15:02
Clueless
Holy hell.
2018-10-28 20:16:23
Rotonen
well, there’s the deep end too https://www.ibm.com/cloud/bare-metal-servers Bare metal servers are dedicated, IBM high-performance cloud servers configurable in hourly and monthly options.
2018-10-28 23:16:01
alistar
can we mining,with,gpu?
2018-10-28 23:17:39
Rotonen
not enough memory in gpus
2018-10-28 23:26:02
Fireduck
Maybe. How fast is system memory to gps bus?
2018-10-28 23:26:25
Fireduck
Maybe an arktika like solution where CPU bundles chunks to GPU
2018-10-28 23:26:41
Fireduck
And do the hashing on GPU
2018-10-28 23:29:57
mjay
If all conditions are perfect (Pcie-3.0, 16x connector, CPU supports it) its ~13GB/s
2018-10-28 23:31:02
Fireduck
There is probably some room to do something there
2018-10-28 23:33:05
Fireduck
You saturate the system memory bus, assuming the cpu can move words around fast enough
2018-10-28 23:33:12
mjay
this would only accelerate memory mining
2018-10-28 23:33:19
mjay
unless gpu memory is used as well
2018-10-28 23:33:22
Fireduck
you flood the gpu with as much data as it can hash
2018-10-28 23:33:35
Fireduck
sure, memory mining is cpu bound
2018-10-28 23:33:37
Fireduck
it moves that bound
2018-10-28 23:33:54
mjay
the GPU will outperform any memory bus
2018-10-28 23:34:07
mjay
its a whole different league
2018-10-28 23:34:17
Fireduck
So, probably pretty fast
2018-10-28 23:36:51
mjay
The usual mining rigs have their GPU´s connected with extenders, limiting the bandwidth to pcie 1x
2018-10-28 23:37:00
mjay
like 700mb/s at best
2018-10-28 23:38:29
Fireduck
Well, you can get that fast with network on a 10gb link
2018-10-28 23:38:56
Rotonen
you're approaching what omnipath is doing in gpu clusters
2018-10-28 23:38:57
Fireduck
so the question is, if you put the entire field in memory, what is the max speed you could read that memory on a cpu?
2018-10-28 23:39:12
Fireduck
from there, the cpu can sling over network or to gpu or both
2018-10-28 23:39:21
Rotonen
ctrl-f gpu
2018-10-28 23:39:49
Rotonen
all that stuff has been solved for years
2018-10-28 23:39:59
Rotonen
just the backing hardware is not gonna be viable
2018-10-28 23:40:08
Rotonen
way too expensive infra-wise
2018-10-28 23:40:47
Rotonen
see use cases, case 4
2018-10-28 23:41:45
Rotonen
~about what you're musing about above with memory moves being cpu bound when they're not on the same bus
2018-10-28 23:42:46
Rotonen
keeping taps on these people might be insightful: https://www.openfabrics.org/
2018-10-28 23:43:03
mjay
with enough GPU memory this could be all GPU, add NVLink-connections and it will outperform everything
2018-10-28 23:43:22
Rotonen
that's nvidia dgx territory
2018-10-28 23:43:29
Rotonen
a mainframe is cheaper per hash
2018-10-28 23:44:08
Fireduck
I actually don't know anything. Which can read from system ram faster, CPU or GPU?
2018-10-28 23:44:11
Rotonen
hopefully, for non-snowblossom reasons, i'm wrong in a generation or two of hardware
2018-10-28 23:44:19
Rotonen
cpu
2018-10-28 23:44:39
Fireduck
So no exotic hardware needed
2018-10-28 23:44:50
Fireduck
CPU bundles for GPU hashing
2018-10-28 23:44:55
alistar
gpu have its own mem, why need system mem
2018-10-28 23:44:58
Rotonen
well, there is pci-e dma stuff, like BAR
2018-10-28 23:45:08
mjay
gpu´s dont have enough
2018-10-28 23:45:13
mjay
@alistar
2018-10-28 23:45:52
alistar
thx
2018-10-28 23:46:04
Rotonen
there is an access diagram which should tell you of how that dance goes - stuff still hits the system memory
2018-10-28 23:46:21
Rotonen
less, but snowblossom is a very evil problem where everything counts
2018-10-28 23:49:03
Rotonen
@alistar also not all gpu execution nodes actually see all of the memory, there's a lot of shuffling within a gpu, it's basically a highly branching tree - http://www.ce.jhu.edu/dalrymple/classes/602/Class13.pdf the first image here is useful for getting at why it's very difficult for what snowblossom is doing
2018-10-28 23:49:17
mjay
I see one problem with this kind of effort .. once the miner is ready and public, the hashrate will climb a lot, causing some snowstorms until the point where memory mining is not possible anymore
2018-10-28 23:50:01
Rotonen
that's not a problem, that's an inevitability - also why i'm saying 100kH/s NVMe miners are sustainable
2018-10-28 23:51:05
Fireduck
Makes sense
2018-10-28 23:51:26
Fireduck
That was my plan overall but getting there is weirder than I thought
2018-10-28 23:51:44
Rotonen
@mjay the development would only be too quick if GPUs are wide enough to have very high per executor unit miss rates not matter, which they're not in regards to width vs. memory bulk
2018-10-28 23:52:14
Rotonen
@Fireduck i'm still waiting for someone to make a *very* wide layered spinny disk raid
2018-10-28 23:52:46
Rotonen
@Fireduck that's the only way i can imagine anyone botnetting this one, but that'd DDoS all the regional ISPs in between while at it
2018-10-28 23:53:40
Rotonen
@mjay a useful way to imagine a GPU is 'could this be solved by duct taping 2^11 pentiums together'
2018-10-28 23:54:11
mjay
You would have to cluster the data required by the GPU anyway, there is not much overhead/memory usage on the GPU
2018-10-28 23:54:41
Rotonen
but each cuda executor would only see like a few tens of megs of memory
2018-10-28 23:54:55
Rotonen
and you cannot really orchestrate sideways within a gpu
2018-10-28 23:55:20
mjay
of cause you can, it´s slow however
2018-10-28 23:55:33
Rotonen
or if you can, that's some higher order side effect of the new fused multiply add they did, i've not yet seen
2018-10-28 23:55:45
Rotonen
yeah, way slower than meaningful
2018-10-28 23:55:47
mjay
if all the data is in CPU memory, why does it need more than a few kb?
2018-10-28 23:55:54
Rotonen
unless there's something missed so far
2018-10-28 23:56:20
Rotonen
if all the data is not on the gpu, why'd you crunch anything on the gpu?
2018-10-28 23:56:32
Rotonen
moving stuff in our out of there is clunky
2018-10-28 23:56:41
mjay
The plan was to move the actual hashing to the GPU
2018-10-28 23:56:49
Rotonen
that's not a bottle neck
2018-10-28 23:57:27
Rotonen
a cuda core has a few tens of megs of very fast seek space, but i have no ideas on how to orchestrate that miss fest efficiently
2018-10-28 23:58:01
mjay
I can do some real-life-testing on this one. Move 10GB of random data to GPU memory, and try to access random 16 byte chunks form GPU memory as fast as possible
2018-10-28 23:58:08
Rotonen
perhaps some of the 'how to pack multidimensional message spaces as efficiently as possible' -approaches could yield something for the known-stable packings, but i have no idea where to being looking into that yet
2018-10-28 23:58:42
mjay
my guess is: Pascal - reasonably fast, Volta - very fast
2018-10-28 23:58:43
Rotonen
@mjay please do, you do actually have a point that what i call 'clunky' might result in better latencies for small seeks
2018-10-28 23:58:56
Rotonen
aka streams vs. blocks
2018-10-28 23:59:35
Rotonen
i keep forgetting @Fireduck was an evil smartypants and made the seek size silly