►
Description
You’re ready to start providing storage to the Filecoin network but need help getting started - join expert Filecoin Storage Provider Benjamin Hoesjbo for a long awaited update to his Solo Storage Providers video.
For more information on Filecoin
- visit the project website: https://filecoin.io/
- or follow Filecoin on Twitter: https://twitter.com/Filecoin
Chapters
00:00 - Introduction
04:35 - Moving to real data
15:00 - Balancing the sealing workers
23:24 - Reference architecture
23:37 - Ref: Daemon, Miner, Market node
34:45 - Ref: Combined Worker
41:04 - Ref: Backup system
47:18 - Rightsizing Combined Workers
58:53 - Reference hardware list: Full 1PiB RAW
1:18:05 - Reference hardware list: Minimum 320TiB RAW
1:26:30 - Bill of Materials: Full 1PiB RAW
1:31:25 - Bill of Materials: Full 320TiB RAW
1:33:13 - Pictures from the test lab
1:46:50 - Where to go next
Links from the last slide:
https://filecoin.io/slack/
https://github.com/filecoin-project/community#forums
https://www.web3espa.io/
A
Hi
everyone
and
welcome
to
the
second
round
of
becoming
a
falcon
solo
storage
provider.
My
name
is
benjamin
and
I'm
the
bpr
founded
at
picnic
and
company,
which
is
a
free
storage
provider
in
north
america.
So
this
video
is
intended
for
people
who
are
considering
to
becoming
a
storage
provider
on
the
falcon
network,
but
it
could
also
be
for
current
storage
providers
who
are
looking
for
optimizations
to
the
hardware
or
it
could
be
for
current
storage
providers
who
are
looking
for
transitioning
into
storing
actual
data.
A
So
this
video
is
not
intended
as
an
investment
advice
and
we
are
obviously
still
mining
and
talking
a
bit
about
that.
But
as
we
move
into
storing
real
data,
that
is
the
main
purpose
here,
whereas
mining
in
itself
is
more
like
what
what
we're
using
to
make
sure
that
we
can
that
we
can
control
incentives
and
also
for
the
basically
the
vehicle
for
transferring
funds.
A
A
So,
in
part
one
we
went
through
how
to
seal
data
on
unmodified
bluetooth
code,
and
that
is
still
the
principle
that
we're
using
today.
So
basically,
we
want
to
try
not
to
change
the
bluetooth
code,
as
is
you
can
get
optimizations
out
of
doing
that.
But
for
me
I
think
it's
very
important
that
we,
instead
of
doing
our
own
modifications,
we
try
to
push
back
to
the
community
and
therefore
change
the
bluetooth
code
in
itself
from
from
there.
A
A
Pip
paper
bite
raw
storage
provider
setup,
which
will
be
able
to
to
seal
around
three
tips
a
day,
and
I
think
it's
getting
more
more
common
that
this
is
probably
around
what
you
can
do
per
setup
and
then
you
could
obviously
from
there
scale
out
if
you
need
more
capacity.
But
that's
basically
what
we're
going
into
another
important
point
here
is
that
I'm
going
to
show
and
prove
that
it's
now
totally
possible
to
do
this,
both
on
amd
and
insole
hardware.
A
So
in
the
last
video
it
was
really
focused
on
amd
only
for
a
lot
of
the
stuff
here,
but
now
it's
perfectly
possible
to
do
the
whole
stack
on
intel
hardware
without
much
penalty
and
we're
gonna
talk
about
the
storage
market
and
how
to
see
the
real
data.
What
that
looks
like
and
how
we
do
a
backup
and
I'm
not
a
big
fan
of
the
idea
that
storage
providers
need
to
do
backup.
A
Basically,
we
should
be
able
to
just
lose
the
data
and
because
the
client
has
saved
that
data
with
x
number
of
different
storage
providers.
It's
okay!
That
one
loses
the
data
and
needs
to
to
rebuild,
but
the
thing
is
right
now,
as
this
network
is
still
early,
the
collateral,
so
the
price
that
we
have
to
pay
for
each
sector
of
data
that
we
store
the
coordinator
is
is
quite
high,
meaning
that
basically
for
each
storage
providers,
it
would
be
pretty
high
risk
not
to
have
a
good
backup
in
place.
A
A
A
If
you
can
say
that
is
symbol
here,
but
but
trying
to
make
it
simple
here,
this
is
only
about
doing
committed
capacity,
so
committed
capacity
is
not
real
data
that
is
kind
of
random
data
that
we
seal,
but
the
reason
why
we
do
that
is
we
committed
to
the
network,
meaning
that
this
will
now
be
available
for
storing
real
data.
The
important
thing
about
doing
that
is
that
it's
it's
a
good
way
to
make
sure
that
the
storage
providers
are
actually
able
to
to
onboard
data
when
the
real
data
is
there.
A
But
what
happens
here
is
that
we
have
an
internet
connection
out
here
and
basically,
all
we
need
to
to
mine
and
to
store
committed
capacity
is
we
need
our
daemon
running
over
at
the
very
left
side?
We
need
that
to
have
an
internet
connection,
so
it
can
sync
to
the
chain
and
basically
do
the
messaging
needed
for
sealing
data.
A
A
Then
it
also
does
the
pre-commit
two,
which
is
the
second
stage,
and
what
I'm
going
to
show
today
is
like.
I
really
like
the
science
where
we
try
to
keep
this
on
the
same
physical
box,
because
when
that
happens,
then
we
basically
don't
need
all
the
high
capacity
and
networking
and
we
don't
need
two
places
where
we
store
the
same
data.
A
The
reason
why
I
don't
do
it
and
this
all
in
the
same
box
in
this
setup
is
because
I
find
it
a
little
bit
difficult
when
we
go
into
the
hardware,
how
to
balance
it
correctly
on
one
machine
and
that's
basically
because
of
when
you
get
up
to
these
sizes
and
memory
stars
become
really
expensive.
If
you
run
want
to
go
really
high
density.
A
So
this
is
a
setup.
Quite
simple:
you
see
the
the
the
light
the
the
yeah,
so
so
the
lighter
color
here
going
from
the
ap
pc1
to
pc2,
then
it
actually
from
there
is.
Is
you
could
say
that
the
c1
and
c2
process
comes
after
that?
But
the
thing
is
it's
not
really
that
much
data
that
is
transferred
to
do
the
to
do
the
snark
proof,
because
that's
basically
just
a
challenge
on
the
data
that
you're
doing
so
data
flow.
A
A
A
This
is
adding
the
the
the
the
market
note
to
the
system
and
basically
it
also
needs
the
connection
to
the
internet
to
do
storage
dealing.
So
that's
basically
telling
or
working
with
the
client
to
figure
out
what
is
the
price
and
everything,
and
then
we
also
have
the
physical
data
transfer
going
on.
This
can
be
offline
and
can
be
online.
So
there's
two
different
ways
of
doing
it.
Basically
online
means
that
it
goes
directly
into
them.
A
The
market
note,
but
the
offline
version
means
that
you
can
just
transfer
it
like
a
normal
transfer
file,
transfer
and
then
afterwards,
on
the
market.
Node
you
can,
you
can
import
the
data
there
and
the
the
further
we
go.
It
seems
like
online
dealing
seems
to
work
better
and
better,
but
there's
still
some
development
to
be
done
there.
So
for
really
large
batches
of
data,
real
data,
then
offline
dealing
seems
still
to
be
the
preferable
way
of
going.
A
It
could
also
be
that
the
offline
dealing
was
simply
because
you
would
get
the
data
in
in
physical
hard
drives
to
to
import
into
the
system,
but
this
is
kind
of
focusing
on
the
online
part,
also
because
retrieval
would
probably
also
be
online,
and
what
happens
is
that
you
see
that
now
the
data
flows
from
the
internet,
it
goes
into
our
market
node.
I
also
put
boost
in
there
because
that's
probably
going
to
be
the
next
implementation
of
the
market
and
meaning
that
that
will
take
over.
A
But
data
goes
into
here
and
then
it
will
be
moved
from
there
into
the
the
app
piece
worker
to
the
pc-1
worker
to
the
pc2
worker.
Then
you
will
have
the
c1
c2
snug
proof,
and
then
it
will
move
into
the
miner
again
once
it's
stored
there,
we
would
normally
have
both
an
unsealed
copy
and
a
sealed
copy.
The
sealed
copy
is
used
for
proving
that
we
actually
are
storing
the
data.
A
A
So
this
is
basically
all
that
happens
when
we
need
to
move
into
storing
real
data.
One
important
thing
here
is
that
you
see
that
the
market
node
here
is
actually
not
located
at
this
same
physical
box
as
the
miner
and
the
daemon
and
reason.
Why
is
because
importing
data,
meaning
that
it
also
needs
to
be
validated,
that's
actually
quite
challenging
for
the
for
the
server,
and
you
could
do
it
in
one
box
and
separate
everything.
But
you
need
to
be
really
aware
of
the
of
the
hardware
consumed
by
the
market.
A
Note
because
it
can
use
a
lot
of
cpu
cores
when
it's
importing
and
basically
what
can
happen
if
you
have
everything
under
the
same
box.
Is
that
if
you
have
a
lot
of
incoming
deals
at
the
exact
same
time,
when
you
need
to
prove
your
sectors,
it
could
be
that
your
cpu
is
actually
over
utilized
and
you
start
seeing
really
slow
proofs,
and
it
might
actually
mean
in
the
end
that
you
could
miss
a
window
post
which
would
be
very
bad.
A
And
the
other
thing
also
seen
is
that
if
you
import
a
lot
of
data,
it
can
also
be
that
your
daemon
is
not
able
to
sync
up.
Basically
it
just.
You
know,
it
gets
starved
and
stops
thinking.
And
if
you
don't
have
your
bluetooth
demon
sync
well,
you
can.
Obviously,
you
know
send
the
window
post
messages
as
well,
and
the
same
goes
for
mining
blocks,
winning
post.
Also,
you
need
to
keep
your
pipe
clean,
so
you're
able
to
actually
provide
those
when
it
happens.
A
So
that's
why
you
would
see
that
I
tried
to
put
this
on
different
servers,
because,
obviously,
if
you
do
this,
then
there
will
be
no
impact
of
whatever
goes
on
on
the
market.
Side
will
have
no
impact
on
your
ability
to
mine
blocks,
to
do
provings
of
your
deadlines
and
also
just
to
keep
your
chain
in
sync,
and
actually
the
same
goes
for
the
ceiling.
A
A
So
next
up
is
back
to
talking
about
the
balancing
of
the
ceiling
workers,
so
this
is
basically
where
we
left
off
from
from
last
time.
So
up
in
the
corner,
you
will
see
one
of
the
slides
that
I
did
on
the
last
video,
which
was
trying
to
see
trying
to
visualize.
How
are
we
actually
balancing
stuff
and
it
makes
really
good
sense
to
try
to
to
put
it
up
like
that
and
see
you
know
what
is
what
is
actually
the
timing?
A
How
much
capacity
do
I
need
between
each
of
these
different
kind
of
workers,
but
that's
happened.
A
lot
of
things
still
since
last
video,
which
is
in
october
2020..
So
at
that
time
it
took
about
10
to
15
minutes
to
do
an
and
that
running
single
core,
so
it
was
very
predictable.
That's
that
was
fine.
It
was
not
the
limiting
factor,
but
today
it's
running
multi
multi-core,
and
so
now
we
can
be
down
to
one
to
three
minutes.
A
Normally,
it
takes
a
little
bit
longer
when
you
have
real
data
compared
to
just
having
a
cc,
but
the
speed
up
is
significant,
pc
one
one
of
the
really
important
things
here.
It
went
from
about
almost
six
hours.
When
we
only
had
a
single
core
sdr.
Now
we
can
do
multi-core
sdr
and
basically
we
have
almost
doubled
the
speed.
So
that's
very
awesome.
It
uses
some
more
cores,
but
we
can
basically
control
exactly
how
that
operates.
A
So
that
is
a
really
good
advantage.
I
don't
think
there
is
many
designs
today
that
would
benefit
from
running
the
old
single
core
sdr.
There
are
some
high
performance
amd
cpus.
That
would
only
be
able
to
do
a
single
core
sdr,
so
you
would
be
locked
on
that,
but
I
don't
think
it's
an
advantage
today,
because,
basically,
you
can
get
cheaper
cpus
that
can
leverage
on
the
multi-core
sdr
and
your
total
aggregated
throughput
would
probably
be
just
as
good
or
even
better
so
on
the
on
the
pre-commit
2.
A
A
So
you
have
some
more
cuda
specific
optimizations
in
the
bluetooth
code
and
obviously
yeah
that
works
really
great.
If
you
have
an
nvidia
gpu,
so
that's
kind
of
the
downside
of
it.
It's
like
this
is
meaning
that
if
you're
running
unmodified
bluetooth
well,
you
are
pretty
locked
into
running
it
on
a
on
a
nvidia
gpu
that
can
do
cuda.
If
you
want
these
kind
of
speeds,
then
on
commit
2,
which
is
also
a
gpu
intensive.
A
A
One
thing
to
think
about
here
is
that
with
these
optimizations
it
also
means
that
the
gpu
is
running
much
harder
than
it
did
before
we
actually
in
the
or
after
the
first
video,
but
before
this,
where
we
started
actually
running
two
processes
in
parallel
on
each
gpu
just
to
utilize
it
better.
So
if
you
had
a
gpu
which
had
20
gigabytes
of
memory
or
more
so
that
could
be
the
3090
which
has
24
gigs
of
memory.
You
would
be
perfectly
fine
at
running
two
c2
processors.
A
At
the
same
time,
you
can
still
do
that.
So
that's
not
a
problem
to
run
two
processes
on
the
same
gpu.
The
thing
is
that
now
that
it's
using
the
crude
optimizations,
it's
running
more
heavy
on
the
gpu,
so
the
kind
of
the
benefit
of
running
two
processes
is
not
the
same
as
it
was
before,
and
it
might
be
that
in
the
future
we
have
so
much
optimization
that
basically,
it
makes
no
sense
to
run
to
process
anymore.
A
The
good
thing
about
doing
that
is
also
when
you're
running
two
processes.
You
also
need
the
memory,
the
system,
memory
of
two
processors,
so
if
you
could
run
one
process
and
run
it
faster,
that
would
actually
be
more
beneficial
than
having
to
hit
it
with
two
processors,
at
the
same
time,
just
to
get
the
maximum
performance
out
of
it,
but
nev
definitely
be
going
in
that
direction.
So
that's
really
good
and
obviously
coming
from
an
hour
and
down
to
12
minutes.
A
So
the
same
thing
goes
on
with
the
cuda
optimizations
for
window
post.
This
is
the
this
is
the
the
gpu
part
specifically
so
in
here
I
don't
take
into
consideration
the
part
about
actually
loading
the
data,
because
that
can
be
very
different.
Compare
depending
on
what
kind
of
storage
system
you
have.
A
But
if
you
need
to
scale
to
a
bigger
storage
capacity,
obviously
you
need
to
distribute
it
better,
and
if
you
do
that,
then
you
will
have
to
take
into
concern
or
consideration
how
how
the
speeds
are
for
actually
loading
the
data,
because
that
can
take
some
time.
But
this
is
the
gpu
part
of
it
and
it
is
really
really
fast
now
so
back
from
10
minutes,
you
have
30
minutes
to
to
complete
your
deadline.
A
That's
a
window
and
basically,
if
you
used
10
minutes
just
on
the
gpu,
well,
you
actually,
you
didn't
have
that
many
tries.
So
if
something
did
go
wrong,
you
need
to
restart
the
minor
anything
like
that.
You
know
you
needed
to
do
it
first
thing
in
the
in
the
deadline
so
coming
down
to
only
3.5
minutes.
It
means
that
you
can
do
like
many
proofs
in
the
same
deadline.
A
So
now
it's
just
like
the
it's
not
as
demanding
and
challenging
as
it
was
before
for
the
winner
post,
sorry,
the
winning
post
it
it's
just
really
really
fast
now,
and
we
can
get
it
under
one
second,
which
is
good,
because,
basically
we
want
to
make
sure
that
we,
if
we
do
get
a
winning
ticket
and
are
eligible
to
do
the
winning
post.
We
want
our
message
out
there
on
the
network
really
fast
to
be
sure
that
we're
not
getting
and
any
if
issues
on
yeah
and
getting
the
block
out
there.
A
A
The
other
part
of
it
is
because
basically,
the
low
tooth
code
still
has
some
optimizations
left
in
there
that
can
be
utilized,
and
there
will
also
be
future
optimizations,
of
course,
so
this
will
only
go
faster
and
faster,
but
the
main
point
here
is
that
being
able
to
balance
like
we
did
in
the
first
video
being
able
to
calculate
so
what
is
my
throughput
for
each
of
the
stages
of
the
ceiling?
That
is
still
the
same
concept,
and
it
remains
the
same.
So
we'll
still
have
to
think
about
that.
A
A
We
will
actually
start
to
look
into
the
reference
architecture
and
I'm
gonna
start
with
some
different
ideas
of
what
you
could
use
and
let's
just
get
on
to
it.
So
this
is
for
the
the
daemon,
the
note,
the
miner
and
for
the
market
as
well,
not
because
it
needs
to
be
in
the
same
box,
but
this
is
kind
of
the
hardware
that
I
would
use
for
those
kind
of
of
yeah
systems.
So,
first
of
all,
I
have
some
different
components
here
that
I
like
that.
A
A
Obviously,
you
could
also
buy
real
rack
servers
like
integrated
servers,
where
you
have
everything
in
there
it
could
be
a
bare
bone
or
it
could
be.
Obviously
just
you
know
a
fully
inspect
server
here,
I'm
using
super
micro
as
reference,
but
it
could
be
any
of
the
big
brands
because
they
all
provide
this
kind
of
hardware.
A
This
hardware
looks
very
much
like
what
you
would
buy
if
you
do
a
ai
machine
learning
stuff
like
that,
so
so
it
is
common
hardware.
But
what
I'm
trying
to
do
here
is
more
set
a
reference
for
what
kind
of
sizing
would
you
go
for,
so
starting
with
that?
The
super
micro
chassis
here
is
a
24
bay
chassis.
A
I
like
it
very
much
because
it
has
those
24
bays
which
can
hold
our
long-term
storage,
because
this
is
about
doing
storage
right,
but
in
the
back
you
also
have
two
little
drives
which
can
hold
some.
Basically,
that
would
hold
the
os
drives
in
rate
one
and
that's
cool,
because
it's
also
hot
swap.
So
you
don't
need
to
take
the
note
and
everything
offline
if
you
do
to
have
a
hardware
failure
on
that.
A
So
on
the
motherboard,
I
kind
of
put
up
two
different
examples,
one
being
the
the
amd
based
the
other
one
being
intel
based
now
for
intel.
Also
on
the
node
and
the
miner.
It
makes
really
good
sense
to
have
a
a
cpu
that
can
do
the
that
has
the
shar
extensions,
because
we
also
need
to
do
some
crypto
calculations
on
there.
So
there
is
a
speed
up
on
the
note.
A
A
That's
on
the
third
generation,
so
that
would
be
iceland,
and
you
could
also
obviously
use
consumer
hardware
here.
So
that
could
be
the
new
leg
and
there's
also
some
of
the
others
that
can
do
this,
but
you
will
pretty
soon
get
into
a
memory
limitation
and
that's
basically,
why
you're
looking
at
server
versions
here
on
the
amd
part,
it
would
be
the
epic
cpu.
I
would
choose.
I
know
a
lot
has
been
running
the
threadripper
and
the
threadburper
pro
is
out
there,
which
is
really
competitive
to
to
running
an
epic
for
this.
A
The
concern
I
have
about
running
epics
is
that
I
are
not
the
epic
sorry.
The
threadripper
is
that
it's
the
gdp
on
these
cpus
are
quite
high,
because
they're
clocked
high.
We
don't
really
need
that.
So
it's
like
it's
not
really
efficient
in
terms
of
how
much
power
it's
actually
consuming
to
do
the
task
and
the
other
thing
is
like
most
of
the
thread
rippers.
They
would
be
limited
to
not
that
much
memory,
maybe
256
gigs
of
memory.
A
That's
okay,
on
these
systems
right
here,
but
I
would
prefer,
if
you
were
running
that
that
at
least
it's
a
ecc
memory
to
make
sure
that
we
are
not
getting
hit
by
any
corruption
of
data
or
anything
like
that
from
the
memory
part.
Another
thing
to
look
out
for
I'll,
actually
go
back
to
them.
When
we
see
some
pictures.
A
Is
that
the
reason
why
I
like
the
the
epics
and
the
installations
is
because
mostly
the
motherboards
would
be
lined
up
for
a
server
setup
where
you
have
decent
airflow
and
that's
quite
important
when
we
start
sealing
really
hard
so
I'll
go
back
to
an
example
on
that,
and
that
could
also
be
why
I
don't
really
like
using
the
consumer
stuff
like
a
threadripper
and
so
on.
It
is
cheaper.
A
A
I
really
like
the
a4000,
because
it
has
such
a
low
footprint.
It
uses
only
a
single
pc
connector
for
the
power,
a
single
slot.
It
has
a
decent
amount
of
memory
there,
and
actually
it
has
a
really
large
amount
of
cuda
cores,
which
is
not
running
fast,
but
we
don't
need
that.
We
still
get
those
blasting
fast
speeds.
A
An
alternative
to
running
these
two
a4000
and
the
reason
why
I
would
have
two
in
there
is
basically
because
right
now
we
have
the
winning
post
and
the
winner
post
running
on
the
same
process,
but
very
soon
we'll
have
those
in
two
separate
processes,
and
the
thing
is
like
when
you
do
a
window
post
for
those
three
and
a
half
minutes.
When
you
do
that,
and
if
you
have
many
deadlines
every
time
you
do
that
the
miner
is
actually
not
capable
of
providing
a
winning
post.
A
So
if
you
get
a
winning
ticket
at
the
same
moment
as
you're
doing
your
winner
post,
you
have
the
ticket,
but
you
can't
provide
the
the
proof.
So
that's
really
annoying,
because
that
does
happen
sometimes,
especially
as
you
move
to
bigger
sizes
of
mine,
it
does
actually
happen
and
that's
super
annoying,
because
you
just
see
your
blog
reward,
just
get
lost
there.
A
But
the
thing
is
like
in
the
future:
we'll
have
a
separate
worker
for
doing
the
winner,
post
and
winning
post,
meaning
that
we
can
separate
them
and
we
shoot
both
at
the
same
gpu.
We
just
need
to
make
sure
that
you
will
run
both
and
not
just
say
hey.
You
need
to
wait
for
my
window
post
to
be
done
before
I
will
do
your
winning
post,
because
that
time
it
will
be
totally
over.
You
know
we'll
be
at
the.
A
I
don't
know
how
many
next
epochs
will
be
there,
but
but
you
know
that's
that's
done
so
just
going
back
to
these
two
motherboards.
So
what
I
like
about
these
motherboards,
if
you
are
kind
of
the
home
builder
style,
is
that
they
do
have
some
integrated
stuff
on
there.
One
thing
is
like
they
already
got
the
10
gig
networking
on
there.
So
that's
on
board,
so
you
don't
need
to
buy
a
separate
pc
express
card
for
that
and
the
other
thing
is.
A
It
also
has
a
sas
controller
and
I
actually
ended
up
using
sas
for
my
os
drives
and
also
for
my
long-term
storage.
But
I
really
like
it
on
the
os
drives
because
it's
it's
it's
kind
of
a
rate
rate
one,
but
it's
kind
of
out
of
band,
so
I'm
not
using
linux
or
anything
integrated
in
there
to
try
to
make
a
raid.
A
A
I've
ended
up
just
using
two
two
three
drives,
so
it's
not
on
the
nvme
protocol,
but
obviously
it's
an
nvme
drive
but
running
on.
This
has
a
protocol
instead,
because
then
I
have
my
my
perfect
rate
one
and
I
also
have
the
kind
of
the
speeds
from
from
nvme
and
that's
pretty
good
when
you
have
the
the
note
so
the
chain
data
syncing
up
on
there
and
so
on.
A
So,
let's
move
on
to
a
combined
pc1
pc2.
This
could
also
do
c2.
But,
as
I
said
before,
the
c2
is
more
like.
That's
challenge
doesn't
really
need
that
much
data
where
the
problem
about
is
having
pc1
and
pc2
separate.
Is
that
if
you
run
a
32
gig
miner,
then
basically
that
expands
into
11
layers
so
11
times
that?
And
then
you
also
have
additional
data
on
top
of
that.
A
So
in
the
end
you
end
up
having
like
almost
500
gigs
of
data,
which
is
expanded
from
pc1
and
then
needs
to
be
moved
into
the
pc2
process
to
finalize
the
job.
And
when
you
keep
that
on
the
same
server,
then
you
don't
have
the
get
process,
so
you
get
rid
of
that
and
basically
you
don't
need
to
wait
before
the
ps2
can
be
executed
on.
So
the
waiting
time
is
not
the
problem
in
itself
because
that's
not
taking
up
any
compute
resources,
but
it's
definitely
taking
up
a
lot
of
storage
resources.
A
Obviously,
these
cpus
are
more
expensive,
so
it
really
comes
down
to
what
prices
can
you
get,
but
in
the
end,
the
reason
why
I
came
up
with
two
different
designs
here
is
also
because,
right
now,
with
all
the
scarcity
and
hardware,
it's
you
know
it
really
comes
down
to
what
you
can
actually
get
your
hands
on,
and
that's
why
I
said
you
know
if
you
can't
get
a
single,
a
single
cpu
board
like
this.
You
could
always.
A
If
you
go
with
a
dual
cpu
design,
then
you
probably
can
do
with
the
with
eight
and
you
would
also
do
with
cpus
being
on
a
lower
like
lower
performance,
cpu
slower
cores,
but
still
these
are
some
different
examples
of
the
motherboards
and
I
just
mainly
just
use
the
super
micro
in
here
the
asus.
There
is
actually
one
of
them
that
I've
been
running
since
space
race,
it's
kind
of
an
odd
design
it
has,
but
it
works
quite
well.
A
A
If
you
are
doing
a
bare
bone
set
up
or
like
a
full
spec
server,
you
might
go
with
like
a
dual
cpu
server,
which
has
16
dimms
each
so
32
dimms
and
then
basically,
what
you
have
is
just
the
double
amount
of
my
single
setup
here.
So
I'm
not
saying
that
this
is
like
you
should
go
in
exactly
this
direction.
I'm
just
trying
to
put
out
some
principles
for
what
it
could
look
like,
and
this
is
probably
a
good
way
to
go
again.
A
A
The
benefit
of
doing
that
is,
like
you
can
run
even
more
processes
at
the
same
time,
but
it's
just
like
for
a
building
a
simple
setup
like
we're
trying
to
do
here.
This
would
probably
be
the
the
preferable
setup
and
again
reason
why
I'm
going
with
dcp
are
these
gpus
is
because
they
have
a
lower
energy
footprint,
but
also
because
they
will
actually
fit
very
well
into
most
servers
and
if
you
are
into
using
gaming
gpus
for
your
first
rig,
that's
fine,
but
be
aware
when
this
needs
to
be
moved
into
a
server
setup.
A
As
you
start
to
scale,
those
gpus
will
probably
not
fit
and
they
will
be
using
too
much
power
to
get
rid
of,
and
they
will
basically
just
have
a
physical
layout
which
is
not
compatible
with
the
server.
So
that's
why
I
would
go
with
something
like
this,
although
if
you
compare
an
a5000
to
the
3090,
which
I'm
also
running
in
the
test,
it's
it's
not
as
performant,
but
then
again,
it's
still
okay,
because,
as
we
just
saw,
the
the
speeds
on
the
gpu
part
of
the
proving
and
the
ceiling
is
much
faster
now.
A
So
this
is
basically
all
I
had
to
say
about
this
setup,
we're
still
using
the
same
chassis
here,
because
we
still
need
to
have
the
long
term
storage
in
front
and
then
basically,
we
have
the
sealer
running
in
the
back
up
there
and
how
we
put
this
together.
It
comes
back
to
that
original
idea
of
how
we
balance
the
different
workers
righty.
So
I
also
included
a
backup
note
here
and
the
reason
why
I
did
that
was
back
to
the
thing
about
that.
A
A
You
now
have
to
pay
10
times
as
much
collateral
for
that.
So,
actually,
you
have
quite
a
lot
of
dollars
in
collateral
for
that
piece
of
data.
You
basically
need
to
secure
it
and
make
sure
that
you
don't
lose
that,
because
that
would
be
pretty
bad
if
you
lose
it
and
not
being
able
to
to
prove
it
will
get
terminated
in
the
end
and
you'll
basically
lose
that
collateral
so
right
now,
incentives
is
pointing
out
that
you
need
to
have
some
kind
of
backup
solution.
A
So
what
I'm
proposing
for
this
setup
is
a
backup
miner
that
I'm
using
myself,
which
is
located
at
a
at
a
totally
different
physical.
It's
an
another
location,
basically
and
reason
why
I
would
do.
That
is
because
my
backup
system
here
is
not
only
backup
of
the
data
in
itself,
but
basically
this
backup
node
is
set
up
to
be
able
to
prove
my
data
so
say
that
I
had
a
major
disaster
at
my
main
site.
A
I
would
basically
be
able
to
use
my
backup
server
to
prove
my
data
to
to
run
through
the
deadlines
for
every
day
until
I
have
my
main
site
working
again.
So
this
is
the
idea.
Obviously,
you
could
also
just
store
the
data
and
say
yeah.
I
just
have
a
backup
my
data.
You
could
even
have
a
like
a
cold
backup
of
your
data
saying
like.
If,
if
happens,
then
I
will
go
and
copy
back.
A
You
should
just
be
aware
that
if
you
are
in
like
the
one
pip
size,
it
will
take
you
a
lot
of
time
to
actually
copy
data
back.
So
this
is
why
it
makes
really
good
sense
to
have
something
that
can
actually
prove
from
where
your
data
is
stored.
The
backup
is
stored
and
then
you
can
do
the
copy
job
at
the
same
time,
maybe
stop
it
while
you're
approving,
but
basically
you
need
to
be
able
to
be
proving,
because
otherwise
it
could
take.
A
What
I
found
is
like
I'm
using
a
a
little
t4,
tester
gpu,
which
is
quite
expensive,
but
also
quite
performant,
and
that's
actually
quite
funny
with
these
kind
of
gpus
is
like
these.
These
very
small
gpus
like
this
is
like
a
single
slot.
It's
low
profile,
it's
very
tiny.
It's
only
using
70
watts
of
power,
but
it's
still
perfectly
capable
of
of
running
the
window,
post
and
stuff
like
that.
A
So
it's
like
you,
don't
really
need
that
much
to
go,
and
this
could
also
be
something
that
could
be
sitting
in
your
your
main
miner,
but
again,
the
reason
why
I
would
probably
not
go
with
something
that
is
because
basically,
it's
just
too
too
expensive
compared
to
what
you
actually
need,
but
for
something
like
this,
you
would
probably
need
to
go
with
something
like
this.
An
alternative
which
is
coming
up
now
would
be
the
the
nvidia
quattro
a
2000.
A
The
2000
is
an
awesome,
little
low
profile,
gpu
it's
using
two
slots,
so
it's
not
the
official
low
profile,
but
it's
it
can
totally
fit
in
there.
But
you
need
the
12
gig
version,
which
just
came
out
and
reason.
Why
is
because
when
you
do
your
window
post,
it
is
still
consuming
up
to
almost
10
gigs
of
video
memory.
So
so
I
would
prefer
to
be
using
it
on
a
12
gig
and
not
do
any
fallbacks
or
anything
like
that.
So
you
have
the
full
performance.
A
A
A
So,
basically
you
could
do
with
the
500
megabits
or
less,
but
you
should
be
aware
that
if
you're
sealing
with
three
tips
a
day,
you
know
it
also
takes
time
to
actually
copy
that
data
over.
So
you
could
have
a
set
up.
You
could
end
up
with
a
setup
where
you're
actually
sealing
faster
than
you're
able
to
back
up
your
data.
So
now
you
just
have
a
cluster
where
it's
not
really
backed
up
correctly,
because
you
haven't
been
able
to
move
over
the
six
or
so
so.
Keep
that
in
mind
when
you
scale
okay.
A
So
this
is
just
an
additional
slide
on
the
how
we
put
together
the
the
different
workers
on
each
of
the
machines,
and
the
thing
that
I
want
to
do
here
is
I
really
like
to
combine
my
app
piece,
my
pc
one
and
ps2
pc,
one
and
ps2.
We
talked
about
about
that
before.
Ap
is
because
that
is
where
you
kind
of
generate
the
data
and
you
pack
the
data
from
from
your
market
note,
and
once
it's
there.
A
So
what
I'm
lining
up
here
is
now
we're
actually
getting
into
some
different
versions
of
cpus
that
you'll
be
using.
Some
of
these
I
have
been
using
so
the
7702p
is
actually
running
as
one
of
my
ceiling
workers
right
now
and
I
find
it
like
for
the
last
year.
It
has
been
probably
the
most
performant
setup,
but
it's
also
a
really
expensive
cpu.
So
it's
like,
of
course,
if
you
can
only
get
it
for
ten
thousand
dollars-
and
I
probably
wouldn't
go
with
that.
A
But
if
you
could
get
a
bargain
like
secondhand
or
anything
like
that,
it
would
be
an
awesome
cpu
to
use
for
a
setup
like
this.
Basically
because
it's
a
single
cpu,
but
it
can
do
15,
pc
ones
in
parallel.
So
that's
give
you
that
gives
you
a
lot
of
a
throughput
on
the
pc
one,
which
can
be
the
limiting
factor
right
now
to
really
get
some.
You
know
some
some
throughput
on.
So
that's
really
awesome
and
I'm
running
at
that
and
that
now
with
one
tips
of
memory.
A
A
A
Each
of
these
cpus
has
eight
compute
complexes
and
the
compute
complexes
is
pretty
important
when
you
do
multi-core
sdr,
which
is
how
we
get
the
high
performance.
Basically,
because
when
we
do
multi-core,
these
multiple
cores
need
to
be
in
the
same
compute
complex
because
they
need
to
run
out
of
the
same
layer,
3
cache.
So
what
happens
is
like
on
a
cpu
like
this?
You
have
a
24
cores,
but
it
actually,
it
has
eight
compute
complexes
and
it
has
three
cores
in
each.
A
I
would
actually
just
run
it
normally,
which
would
be
having
like
one
master
and
then
three
producers,
but
you
can
actually
control
this
yourself.
So
the
reason
why
I
bring
it
up
is
because
this
is
really
important.
If
you
want
to
go
into
some
of
these
systems
and
using
multi-core
sdr
is
you
need
to
be
aware
of
your
cp,
your
cpu
topology?
A
I
would
say
it's
not
it's
like
it's
going
really
deep
into
the
technical
part
of
a
cpu
to
figure
out
exactly
what
it
looks
like.
It's
not
really
rocket
science.
It's
just
like.
Is
you
really
need
to
go
really
nerdy
into
exactly?
What
does
the
cpu
look
like,
but
I'm
imagining
that
this
is
just
because
this
is
where
we
are
today.
So
probably
we'll
see
in
the
future
that
bluetooth
can
do
more
of
this
in
the
future
and
basically
say
okay,
you
got
a
cpu
looking
like
that.
A
I
would
probably
seal
like
this
and
you
don't
have
to
care
that
much
about
it,
but
for
now
having
a
setup
like
that,
you
would
have
to
go
in
and
say:
hey
you
only.
You
can
only
use
two
producers
per
job
because
otherwise
you
would
like
you
will
start
getting
over
each
compute
complex
that
wouldn't
work.
A
The
seons
on
the
other
side
is
very
much
different
from
an
amd.
They
basically
have
one
shared
a
layer,
3
cache,
meaning
that
now
you
have
one
compute
complex.
That's,
basically
how
we
look
at
it.
So
you're
not
limited
by
figuring
out
how
you
can
fit
these
different
processes
into
each
compute
context,
because
there's
only
one
there's
only
one
shared
layer,
3
cache,
it
makes
it
kind
of
easier.
But
then,
on
the
other
hand,
most
of
the
scenes
does
not
have
the
same
amount
of
cores,
so
you
might
have
to
reduce
it
even
further.
A
A
So
that's
of
course,
not
so
beneficial,
but
it's
also
very
fast,
but
it's
one
of
the
cpus
that
I
have
actually
been
testing
this
on,
and
I
know
that
this
works
really
well
and
one
of
the
key
findings
from
that
is
that
I
can
run
it
with
two
two
producers,
but
I
could
actually
also
go
down
and
only
have
one
producer.
So
now,
I'm
only
using
two
physical
cores
for
each
pc
one
job
and
I
did
it
at
almost
no
penalty,
so
we
are
still
in
the
room
of
three
hours.
30
minutes.
A
If
I
do
like
a
ton
of
pc
ones
in
parallel.
So
now
your
limitation
will
probably
not
be
the
cpu
in
itself
because
it
has
a
lower
amount
of
cores,
but
it
will
probably
be
about.
You
know
how
much
memory
do
you
actually
have,
and
the
thing
is
like
for
each
of
these
pc1
jobs.
We
need
those
65
gigs
of
memory,
so
you
will
probably
be
limited.
A
If
you
are
into
one
tip
of
memory
on
your
your
board,
you
will
be
limited
to
only
doing
15
pc
one
jobs,
but
that
would
be
totally
doable
on
a
single
cpu
setup
like
this.
You
can
also
do
the
dual
setup.
The
reason
why
I'm
trying
to
keep
this
to
one
tip
of
memory
or
two
times,
half
a
tip
of
memory
is
because,
with
the
amount
of
dimms
that
we
have
here,
we
keep
it
at
at
64,
gig
dimms.
A
The
reason
why
I
do
that
is
because
moving
to
20
128
gig
dimms,
that
is
totally
possible
today,
but
it
might
be
that
the
price
premium,
basically,
what
you
have
to
pay
to
get
these
sticks
is
just
like
it's
a
big
step
up,
and
it's
not
times
two,
it's
more
like
times.
Three
I
know
times
are
changing.
This
will
probably
change
in
the
next
half
year
or
so,
but
for
now
it
can
be
really
expensive
still
to
buy
the
the
like.
A
The
big
capacity
sticks
and
that's
why
I'm
trying
to
size
this
after
saying,
like
yeah,
we
need
to
keep
it
at
using
the
the
64
gig
memory
sticks
that
just
makes
the
most
economical
sense
for
now,
but
obviously,
as
we
go
bigger
sizes,
we
can
go
more
dense.
We
can
do
even
more
pc
ones
in
parallel,
and
it's
really
interesting
to
see
something
on
the
insole
here
that
you
can
even
go
down
to
one
producer.
Only
on
the
amd.
A
A
A
But
what
is
important
to
also
think
about
here
is,
like
you
also
have
the
other
processors
that
need
to
run
on
here,
and
one
trick
that
I
really
like
to
use
is
to
keep
the
smt
on.
So
you
have
the
the
hyper
threading
or
the
smt
still
running
on
these
cpus.
The
multi-core
sdr
make
sure
that
it's
only
using
the
first
thread
in
each
cpu
core.
So
it's
not
using
both
threads
of
the
cpu
call.
A
That
would
be
pretty
stupid
if
you're
trying
to
do
pc1,
because
they
would
basically
need
to
go
both
at
full
speed
and
you
would
end
up
having
half
speed
for
each.
So
we
don't
want
that.
But
what
we
do
want
is
for
the
ap
process,
which
is
a
multi-core
process
and
for
the
pc
2
process,
which
is
also
a
multi-core
process.
A
We
would
like
to
kind
of
utilize
some
of
this
part
of
the
cpu
that
is
not
so
heavily
utilized
and
what
happens
is
like
when
you
run
with
the
two
producers
or
three
producers.
It's
actually
running
at
maybe
like
40
or
20,
something
in
that
room.
So
it's
like
it's
not
really
using
the
full
capacity
of
that
core.
That
is
running
the
producer
job.
A
So
if
you
have
the
producer
job
running
on
the
first
thread
in
each
cpu
core,
then
why
not
use
that
second
thread
to
do
something
else
and
that
could
be
the
ap
process.
So
that's
what
you
see
here
in
the
light
red
or
it
could
be
the
the
pc2
process,
which
is
here
in
in
light
blue
and
basically,
what
you
see
is
on
one
side.
You
have
the
the
green,
which
is
the
master.
You
have
the
the
yellow
which
it
produces.
A
A
So
where
I
have
the
pc
one
jar
calculations
going,
that
is
always
hammering
at
a
hundred
percent
load.
I
would
never
want
to
have
anything
happening
on
the
second
thread
trying
because
I
would
basically
reducing
my
throughput
of
my
pc
one
where
having
something
else
going
on
on
the
producer.
Threads
seems
to
have
very
little
impact
on
actually
still
performing
on
the
pc
one.
So
that's
one
thing
to
look
out
for,
if
you
want
to
go
into
these
kind
of
combined
systems
is
like
you
can
actually
utilize.
A
So
that's
a
lot
of
the
optimizations
on
the
actual
node
and
how
you
can
do
that,
then
we're
moving
into
the
summary-
and
I
guess
this
is
actually
where
you
could
say
that
we
are
looking
at
what
is
actually
the
reference
architecture
here.
So
I'm
just
moving
this
over
here.
So
we
can
see
what
goes
on
so
basically
in
this
setup,
which
is
the
full
one,
pip
storage
provider
setup,
you
would
have
three
of
these
24
base
servers
sitting
next
to
each
other,
and
then
you
would
have
your
backup
miner,
backup
node.
A
A
A
So
that's
why
I
want
to
keep
this
a
little
bit
open
just
to
say
that
you
have
to
be
pragmatic
about
getting
hardware
these
days,
but
at
least
24
drives.
That's
actually
in
all
of
these,
and
then
you
have
your
miner.
You
have
your
your
daemon
and
then
the
storage
in
itself
sits
on
the
on
the
first
server.
There
there's
also
a
different
specs
on
what
cpu
you
could
use
up
there.
I
use
the.
I
actually
use
the
7302p.
A
So
that's
a
16
core
epic.
It's
an
older
version
now,
but
it's
like
perfectly
fine
for
doing
that.
No
issues
there.
You
could
also
have
a
higher
like
a
higher
core
cpu,
especially
if
you
also
want
to
put
some
of
these
stuff
together,
like
having
the
market
note
on
there,
still
separating
it
and
sell
reducing
the
amount
of
cores
that
can
actually
use.
But
you
know
the
more
you
put
on
there
make
sure
that
you
have
enough
course.
Then
we
have
the
debuts
as
we
talked
about.
A
A
I
would
I
would
if,
if
possible,
I
would
never
turn
them
off.
If
I
have
a
hardware
failure
and
say
just
you
know,
power
offs,
I
would
make
sure
that
I
had
the
full
array
on
each
of
these.
I
wouldn't
split
the
array
between
each
of
these
j-pops.
Basically,
because
if
you
have
half
an
array,
you
might
have
your
your
rage,
software
there
trying
to
rebuild
or
start
to
do
something
that
it
should
not
do
so
better,
just
lose
it
all.
A
Just
have
it
offline
and
just
saying
like
when
I
turn
this
on
again,
it
will
rediscover
that.
Okay,
there
was
my
array,
so
that's
one
important
thing,
but
I
don't
really
consider
it
that
big
an
issue
from
an
operational
perspective
yeah.
It
could
be
an
issue
that,
if
I
need
to
do
service
on
one
of
the
other
nodes,
and
I
need
to
turn
it
off
then
obviously
I
would
also
need
to
turn
off
my
my
my
my
first
node
to
that.
A
A
Obviously
the
hard
drives
and
everything
like
that
would
be
replaceable
without
turning
it
off.
So
that's
hot
swappable,
the
fans
in
there
and
so
on.
There
is
some
stuff
that
is
housewarmable,
but
for
the
cpu
and
so
on.
I
would
you
know
this
just
needs
to
run,
as
is
no
changes
at
all,
so
that
was
the
the
first
machine
up
there
and
it's
running
the
demon.
It's
running
the
miner.
A
It's
it
has
these
different
roles
called
mining,
sealing
and
the
sector
storage,
I'm
not
having
the
markets
on
there,
as
I
said
before,
but
it
could
also
yeah.
It
could
also
be
on
there
talking
about
the
the
storage
here.
So
if
we
have
24
drives
on
each
of
these
machines,
I
would
probably
set
it
up
as
as
two
arrays
on
each
of
them.
A
A
You
also
spend
more
on
connell,
but
you
use
less
on
physical
hardware.
So,
basically,
your
footprint
will
be
much
smaller.
Your
physical
footprint
will
be
much
smaller,
but
your
storage
provider
operations
will
be
larger
and
you
start
probably
larger.
So
if
you
go
that
like
saying
like,
I
only
did
a
verified
deals
and
you
have
the
fast
retrieval.
So
you
have
the
two
copies,
so
you
don't
need
to
do
the
unseal.
A
A
So
if
we
move
on
from
the
the
daemon
and
miner
down
to
the
market
node
the
thing
about
that
market
node
is
that
it
actually
has
a
double
mount
memory
and
why
we
have
more
memory
on.
There
is
not
because
of
market
node,
because
that's
almost
using
no
memory,
but
it's
because
we
want
to
do
the
c2
on
that
machine.
A
Now,
as
I
said,
c2
is
kind
of
lightweight
on
the
networking,
but
obviously
it
needs
a
lot
of
resource
resources
on
the
cpu
and
also
on
the
gpu
and
basically,
what
I
propose
there
is
like
we
would
run
two
c2
c1
c2
workers
on
there
and
basically
point
at
each
gpu.
If
so
only
one
gpu
at
a
time
if
it
was
the
a4000
or
point
both
processors
at
the
same
a5000,
if
there's
only
one
gpu
in
there
and
it
needs
all
those
almost
200
gigs
of
memory
each
of
these
processors.
A
I
just
pointed
out
here
that
it
probably
take
about
20
minutes
to
to
complete
two
c2s
at
the
same
time,
even
if
it's
on
the
same
cpu
gpu-
and
that
said
we
are
at
about
the
capacity
of
about
three
and
a
half
four
and
a
half
tips
a
day.
So
we
just
need
to
make
sure
that
we
keep
that
high
enough
to
keep
up
with
the
pc
one.
That's
basically
the
main
point
here.
A
So
another
thing
is
like
you
also
see
that
this
has
some
storage
there,
which
is
a
bigger
partition
for
the
bluetooth
markets.
Currently
you
also
see
that
I'm
using
eight
tips
drives
on
that.
So
that's
obviously
a
pretty
expensive
drive
to
use,
but
the
reason
why
we
do
that
is
basically
because
online
deals
and
offline
deals.
They
would
probably
need
some
caching
and
yeah.
A
You
could
put
it
on
a
single
drive
or
just
put
a
single
nvme
in
there
just
say,
like
all
the
data
that
I'm
caching
here
for
importing,
I
just
put
it
on
there.
That's
fine!
It's
just
like
if
you
lose
that
disk,
that's
just
annoying
if
you're
working
with
real
clients
in
the
other
end,
you
say:
hey,
hey,
I
screwed
up
your
data
and
now
I
need
to
download
it
all
again.
It's
like
yeah,
okay,
that's
probably
not
so
serious.
A
It's
like
that!
Doesn't
sound
like
you're,
really
a
secure
place
to
store
your
data.
So,
although
your
long-term
storage
might
be
fine,
then
it
you
know
if
you
lose
your
data
while
doing
the
import,
it's
like
it's
just
not
annoying
if
you
have
to
go
back
so
so
that's
why
there
are
some
some
raid
1
storage
there
for
for
the
caching
as
well.
A
So
just
like
we
talked
about-
and
I
set
that
up
here
both
been
an
example
with
the
apex
and
the
cns,
both
with
the
single
cpu
and
double
cpu,
because
basically
you
can
do
all
that.
But
again,
if
you
do
the
dual
cpu
and
you
do
it
in
a
setup
like
I
proposed
here,
you
are
going
to
have
a
bigger
overhead
on
your
power
usage,
probably
having
a
dual
cpu
setup.
A
You
can
buy
a
much
less
expensive
cpu,
so
it
could
actually
be
that
on
your
purchase
price
and
the
availability,
you
would
maybe
be
better
off
doing
a
dual
cpu
setup.
Basically,
because
you
can't
get
a
good
prices
on
the
single
cpus
because
they
need
to
be
really
high-end
and
also
because
you
yeah
it's
just,
it
might
be
just
cheaper
and
more
available.
A
So
what
happens
here
is
like
if
we
say
that
we
can
do
somewhere
in
the
room
of
12
to
15
pc
ones
in
parallel
on
this
box,
and
they
will
take
about
from
three
hours
tin
up
to
maybe
three
and
a
half
hours.
You
are
looking
at
just
around
in
the
area
of
three
chips
a
day,
and
that
is
going
to
be
like
your
limiting
factor
of
the
ceiling
cluster.
So
you
could
say
if
you
said
one
cpu
setup
and
you
can
do
three
ear
tips
a
day.
A
Well,
probably
that's
about
our
benchmark
right
now
for
how
fast
you
can
go
and
if
you
do
something
optimizations
you
buy
a
higher
end,
cpu
and
so
on.
You
might
be
able
to
tweak
that
up
to
almost
four
tips
a
day,
but
that's
you
know,
that's
just
really
pushing
it
and
again,
if
you're
talking
about
verified
data
here,
you
times
it
by
10,
it's
still
a
lot
that
you
can
actually
seal.
So
if
we
say
that
our
reference
is
around
three
tips
a
day,
you
could
have
a
dual
cpu
bare
bone.
A
It
could
probably
do
six,
because
now
we
would
have
like
two
cpus
high
performance
cpus
and
you
would
have
32
dimms
in
there.
So
it's
like
per
cpu.
You
could
probably
do
around
three
today
and
maybe,
if
you
optimize
very
much
up
to
maybe
four,
but
that's
basically
what
we're
looking
at
and
if
you
want
faster
feeling
than
this,
you
know
you
could
totally
go
ahead
and
just
buy
one
more
of
these
workers.
A
You
just
need
to
make
sure
that
you
can
still
provide
the
c2
so
on
this
box,
I'm
also
doing
the
pc2,
as
I
said,
and
the
thing
is
like.
If
we
run
two
processes,
it
doesn't
use
that
much
memory
to
do
so.
It's
not
really
a
concern
the
memory
it
does
use
some
of
the
cpu.
So
you
need
the
that.
Those
additional
steps
that
I
talked
about
you
need
some
of
those
available
on
the
cpu.
A
So
if
you're
doing
an
intel
setup
with
only
one
one
master
and
one
producer
and
you're
doing
that
on
all
available
cores,
I'm
pretty
sure
you
know
you
won't
have
available
capacity
to
actually
do
your
pc2
or
you
will
see
a
decrease
in
your
pc
one,
because
if
you're
only
running
with
one
producer,
it's
going
like
more,
like
percent
on
that
load
on
that.
So
so
like.
A
But
we
see
that
if
you
do
two
pc
twos
on
the
systems
like
this,
you
probably
have
you
could
do
two
process
at
the
same
time,
on
the
c,
gpu
and
or
or,
as
you
said
before,
you
know
doing
the
a4000
just
put
one
on
each,
but
you
will
end
up
with
a
pretty
high
capacity
there.
So
again,
the
probably
the
limiting
factor
here
will
be
the
as
long
as
you
go
with
the
like
these
kind
of
size
setups.
A
So
that
will
be
probably
where
we
need
to
do
the
tweaking,
and
the
really
important
thing
here
is
like
we
need
to
make
sure
that
the
that
we
have
enough
ps2
capacity
to
always
keep
up
with
our
pc
one.
So
that's
why
actually
one
of
the
key
issues
around
having
a
combined
ceiling
worker
like
this
is
like
what
happened
if
our
throughput
of
pc1
just
doubled.
A
A
I
have
a
problem,
because
now
I
can
do
some
of
the
pc2
on
the
node
itself,
but
I
would
actually
have
to
transfer
some
of
it
like
off
the
the
server
to
do
that
and
that's
a
little
bit
annoying,
because
if
you
do
that,
it's
like
now,
you
need
the
networking.
Now
you
need
the
double
nvme
and
everything
to
do
that.
Let's
try
kind
of
what
we're
trying
not
to
do
so.
A
The
backup
node
here
is
is
basically
the
same.
It
has
more
drives
in
there
and
the
reason.
Why
is
because,
if
we
have
six
arrays
on
those
on
the
main
side,
and
each
of
these
six
arrays
has
all
the
verified
deals
and
the
fast
retrieval,
then
it
actually
has
kind
of
the
double
the
physical
storage.
So
what
we
need
is
we
need
our
backup
node
to
have
about
half
of
that,
because
we
don't
need
well.
If
we
were
really
nice,
then
obviously
we
just
had
a
full
mirror
of
the
main
side.
A
But
the
thing
is
like:
if
our
main
sign,
you
know
total
disaster
there,
we
don't
need
the
fast
retrieval
files
to
actually
prove
our
data.
So
as
a
bare
minimum,
we
need
the
the
sealed
copy
and
the
sealed
copy
is
in
the
seals,
cache,
folder
and
also
the
the
sealed
folder.
But
it's
important
to
have
both
those,
but
we
don't
need
the
unsealed
copy
and
therefore
my
backup
side
is
basically
half
the
size
of
my
main
site.
A
A
A
The
quatto
a2000
will
probably
be
a
good
fit
here
when,
when
it's
really
out,
so
that's
basically
it
so
that's
it
for
the
the
full
setup.
Now
what
I
wanted
to
touch
on
is
like
I
already
grayed
out
on
the
ceiling
worker,
that
you
know
you
might
not
have
those
24
discs
in
there
reason
why?
Because
today
I
actually
don't
have
that,
so
I
don't
have
that
full
capacity
of
this
system.
I
just
know
that
you
know
I
could.
A
I
could
do
that
if
I
liked
you
and
the
in
the
same
setup,
but
just
to
go
from
there,
and
this
is
kind
of
the
full
setup,
I
would
probably
say
that
it
unless
that
you
know
that
you're
just
gonna
hammer
it
and
just
get
that
full
one
pip,
and
then
you
just
go
on
from
there.
Obviously
you
could,
you
know,
put
in
the
orders
for
something
like
this
again.
This
is
not
a
setup
that
is
designed
to
go
from
one
pip
raw
to
tin
paper
to
100
pip
raw.
A
If
you
know
that
you're
going
to
100
pip
roll,
then
probably
you
want
to
do
like
a
network
setup
from
the
very
beginning,
because
you
know
your
storage
will
never
be
direct
attached.
That
would
be
crazy.
So,
basically
you
need
something
that
can
scale
out
really
nicely
without
any
abundance,
so
think
it
right
from
the
start.
But
if
you
want
to
be
in
the
realm
of
this,
like
one
pip
size,
maybe
less,
this
could
be
ideal.
A
This
is
a
very
efficient
setup
where
you
don't
have
more
servers
than
you
need.
I
try
to
go
by
the
kind
of
the
idea
that
I
want
the
least
amount
of
power
supplies
in
there.
The
more
power
supplies
I
have
the
more
overhead
because
you
would
have
motherboards
and
stuff
like
that,
and
it's
all
just
overhead,
and
over
and
over
here.
A
So
the
more
you
can
put
things
together
make
sure
that
you're
using
your
resources
like
the
ceiling
worker,
is
not
spending
any
resources
on
having
those
24
drives
in
the
front,
because
they
will
just
sit
on
the
assess
like
assessed
direct
attached
up
to
the
the
main
miner.
So
it
doesn't
even
touch
the
system
there.
So
there's
no
overhead
than
that,
and
that's
my
general
idea.
A
So
if
you
have
a
dedicated,
nas
server
standing
next
to
this
system,
they
have
cpus,
they
have
memory,
they
use
a
lot
of
it
when
they
transfer
data
and
so
on.
That's
just
overhead
and
overhead.
So
for
small
setups
I
I
would
definitely
go
with
some
something
like
this,
but
if
you
are
to
start
at
a
small
scale,
which
I
make
think
makes
totally
sense,
then
you
could
also
go
on
something
like
if
we
jump
to
the
next.
It
could
be
sorry.
It
could
be
something
like
this.
A
Where
now
what
we
did
was
we
got
rid
of
one
of
the
servers
from
the
main
site,
and
that
was
basically
the
the
separate
market
note
we
got
rid
of
so
just
saying.
Like
okay,
we
don't
need
that
to
be
running
on
a
separate
server,
we'll
just
separate
it
out
on
the
main
miner
and
that
that
will
be
fine,
because
we
just
upgraded
the
main
miner
a
little
bit.
If
we
do
that,
then
obviously
the
market
process
and
the
c2
process
we
had
before
well
that
needs
to
move
to
somewhere
else.
A
So
now
we
are
in
setup
where
we
only
have
like
the
the
demon,
micro,
miner
and
lotus
and
the
market
node.
Well,
so
all
the
the
the
c2
and
going
on
the
first
note.
So,
let's
just
let's
just
get
back
to
that,
so
we
basically
took
out
one
server.
We
reduced
the
amount
of
hard
drives
from
from
96,
so
that
was
like
fully
occupied
on
the
whole
on
the
whole
system
and
then
down
to
only
25.
A
A
Half
of
it
filled
up
and
also,
I
would
probably
not
have
a
tpu
in
the
in
the
in
the
backup
server,
because
it's
quite
expensive,
if
I
needed
to
prove
then
probably
I
could
get
my
hand
on
some
kind
of
gpu.
That
would
be
able
to
do
that.
So
it's
kind
of
it's
more
from
a
risk
perspective.
That's
saying
like
yeah.
Now
I'm
not
really
able
to
just
prove
my
data
like
hot.
A
If
something
happened
on
my
main
site,
but
probably
in
24
hours,
I
would
be
able
to
come
up
with
a
solution,
because
I
do
have
the
data
available
there.
So
I
would
be
able
to
come
up
with
solutions,
so
I
can
start
proving
my
data
again.
So
it's
like
saying,
like
yeah,
it's
nice,
to
have
that
gpu
on
there.
But
on
the
other
hand,
it's
also
it's
also
quite
expensive
to
do
that,
and
we
could
probably
leave
that
out
if
we
just
like
need
a
really
minimum
starting
setup.
A
But
what
I
do
want
here
is
like
what
I
also
put
here
is
like.
I
would
like
to
keep
stuff
separate,
so
I
know
that
a
lot
in
the
community
there
is
starting
with
a
single
setup,
basically
having
a
single
rig.
So
you
have
your
miner,
your
demon,
you
have
your
ceiling
work
worker
running
on
the
same
system,
it's
all
using
the
same
gpu.
A
I
find
that
really
nice.
If
you
want
to
just
get
started
with
filecoin,
but
I
would
definitely
say,
go
and
do
that
on
one
of
our
test
networks
that
could
be
calibration
or
the
butterfly
or
it's
just
don't
use
real
fill
on
that,
because,
in
my
experience,
what
happens
that,
if
you
have
a
single
setup
with
everything
on
there,
you
can
be
totally
fine
if
you
are
able
to
separate
the
processes
so
running
it
in
containers
running
it
with
the
task
set,
making
sure
that
it's
not
using
up
too
much
resources.
A
A
Unless
that
you
are
like
totally
pro
of
how
you
separate
your
processes
and
making
sure
that
they
won't
eat
up
your
demon's
ability
to
sink
the
chain
or
your
miners
ability
to
do
winner
post,
improving
your
deadlines,
if
you
can
do
that,
awesome,
do
a
one
box
setup
if
you're
not
totally
comfortable
with
setting
that
up
better,
go
with
two
boxes,
just
making
sure
that
you
have
your
demon
and
miner
running.
It
can
also
do
the
the
this
this
c1
and
c2
on
there.
A
But
if
we
do
that,
it
would
be
because
we're
running
the
two
a4000
so
we're
running
the
one
a5000
and
as
these
are
two
different
processors,
we
would
actually
be
able
to.
You
know
do
a
c2
and
at
the
same
time
do
a
winner
post,
but
it
won't
sit
there
and
say
I
have
to
wait
for
the
c2
to
be
done.
So
so
I'm
pretty
comfortable
that
this
setup
will
work
fine.
I
I'm
not
using
it
right
now,
because,
basically
I
want
to
keep
those
things
separate.
A
I
don't
think
it
worth
it
trying
to
jeopardize
your
proving
deadlines
or
generating
block
rewards
with
just
cramping
it
all
together,
not
like
this,
but
the
thing
here
is
like
we
are
offloading
our
c2
to
our
main
miner
and
saying
like
yeah,
you
can
do
that
as
a
separate
process.
You
just
need
a
double
amount
of
memory
and
it
can
do
about
probably
about
two
tips
a
day.
A
If
you
do
it
like
that,
then,
on
your
on
your
other
machine,
that
will
be
your
old
ceiling
worker,
but
it
also
has
a
market
node
on
there
again,
using
that
high
spec
cpu
or
using
the
other
designs
that
we
talked
about.
You
will
still
have
that
one
tip
memory
in
total.
So
now
you
probably
need
to
reduce
a
little
bit
on
how
many
pc
ones
you're
doing,
because
you're
probably
locked
in
by
how
many
tips
you
can
you
can
do
because
now
you
also
need
to
do
the
c2.
A
So
the
thing
here
is
like,
if
you
will
have
one
tip
of
memory
in
total.
Now
you
also
need
to
do
one
c2
on
that
same
machine.
It
eats
up
those
first
200
tips
of
memory.
So
now
you
only
have
a
little
bit
over
800
tips
of
memory
left
to
do
your
pc
once
so
now
you
can't
do
15
now
you
probably
have
to
do
11
12.
Something
like
that.
A
But
if
I'm
offloading
some
of
the
c2
to
my
my
my
miner
node
my
miner
server,
then
I
would
probably
get
up
to
the
full
two
and
a
half
tips
a
day.
So
you
can
start
to
see
here
that
it's
not
as
clean.
If
you
do
this
also,
if
the
market
node
is
running
really
heavily,
it
could
also
impact
your
ceiling
performance.
So
it's
like
it
is
kind
of
a
cheaper
setup,
but
it
should
still
be
pretty
safe
in
terms
of
that,
your
daemon
and
market
node
is
running
separately,
so
they
won't.
A
You
know
the
volatility
from
what
can
happen
from
incoming
deal.
Activity
won't
have
any
effect
on
your
ability
to
prove
and
to
sink
the
chain,
and
that's
the
most
important
here,
but
that's
basically
all
I
have
to
say
about
this
part
of
it,
because
this
is
just
to
say
like
go.
Look
at
the
the
full
one,
pip
size,
setup
and
say:
hey.
I
can
get
rid
of
some
of
this
stuff.
I
don't
need
to
fill
it
up
with
all
almost
100
drives
to
start
with.
A
I
can
just
start
with
something
lower
and
I
just
know
that
it's
all
set
up
correctly,
so
I
can
basically
just
swap
in
the
drives
and
initialize
the
arrays
and
attach
them
to
the
miner
and
now
I'm
up
to
the
full
capacity
and
that
make
that
might
make
more
sense
when
we
move
to
the
next
slide
here,
which
is
basically
just
telling
what
is
what
is
a
bit
of
materials
on
something
like
this?
So
now
I'm
getting
really
specific.
A
I
cannot
say
that
these
prices
are
like
what
you
can
get
right
now
out
there,
but
I
tried
to
google
most
of
it
just
to
get
a
reference
price
on.
What
is
you
know
what
should
be
possible
to
get
just
to
get
a
price
on
this,
and
you
see
that
if
you
go
with
a
design
like
this,
which
is
a
one
pip
setup,
I
would
call
it
probably
not
enterprise
great,
because
you
know
it
is
still.
You
know
it's
not
even
a
bare
bone.
A
It
is
still
something
that
I've
screwed
together,
but
the
hardware
components
in
there
is
pretty
enterprise-ish.
Well,
the
the
cost
will
also
be
pretty
significant.
So
we
are
up
to
almost
100
000
for
the
full
setup
here.
So
that's
not
nothing,
but
then
again,
I
think
what
we
need
to
do
here
and
that's
also
what
I
pointed
out
here
in
the
lower
left
side.
A
Is
that
basically
you
know
if,
if
you
look
at
this
100
000
on
the
hardware,
yes,
but
you
have
almost
half
a
million
dollars
on
the
collateral
that
would
go
into
these
boxes
if
you're,
actually,
stealing
only
verified
deals.
So
the
thing
is
like
yeah
hardware
is
is
expensive,
but
on
the
other
hand,
it's
like
you
are
putting
a
lot
of
trust
into
these
systems,
and
I
know
that
we
have
both
a
backup
side
now
and
we
have
our
primary
side
and
we
have
these.
A
It's
all
on
rated
storage,
but
it's
like
if
you
lose
this
data.
Well,
that
is
potentially
half
a
million
dollars
of
collateral
that
will
just
be
lost
and
that
would
be
pretty
bad.
So
it's
like,
I
can
see
why
people
would
go
with
something
like
cheaper
and
consumer
hardware
and
so
on.
But
if
we
are
in
mainnet
and
we
start
getting
into
verified
deals-
and
we
want
to
go
big-
it's
like
why
on
earth
would
you
spend
ten
thousand
dollars
on
your
mining
rig,
which
is
holding
like
a
ton
of
collateral?
A
That's
just
like
that.
Doesn't
really
make
any
sense.
It's
just
too
much
risk
if
you're
not
running
rated
drives
for
your
os,
all
that
kind
of
stuff
it's
like.
If
things
goes
bad
and
you
can't
really
recreate
you
can't
prove
you
might
end
up
with
the
lost
collateral.
It's
like
that's
not
really
worth
it,
and
I
kind
of
found
out
this
moving
from
back
in
the
days
and
calibration
that
moving
to
the
the
main
at
launch
and
doing
space
race
and
so
on.
A
I
started
out
with
having
none
of
this
in
control
like
I
was
just
setting
up
a
test
system
and
now
it
went
into
mainland
and
I
was
like
whoa.
We
probably
need
to
do
something
about
this,
because
I'm
taking
like
a
risk
that
I'm
probably
not
really
happy
with
with
doing
and
what
I'm
trying
to
do
here
is
more
saying
like
yeah
it's
getting
expensive,
but
then
again
you
know
if
the
goal
is
to
put
verified
deals
on
there
and
you
see
how
much
collateral
you
want
to
put
on
there.
A
Well,
it's
you.
You
probably
can't
take
those
kind
of
risks
on
the
hardware.
So
so
that's
it.
The
hardware
cost
percentage
I
put
out
here.
So
that's
about
70
on
this
would
be
buying
physical
hardware
versus
how
much
you
need
to
uncollateral.
Another
thing
to
point
out
here
is
like
this
will
be
using.
A
I
won't
say,
burning
because
it
is
collateral,
but
it
will
be
using
about
if
we
use
those
20
dollars
per
fill
right
now
it
would
be
spending
about
three
thousand
dollars
a
day
if
you're
sealing
at
full
capacity.
So
that's
basically
what
you
need
to
feed
it
with
the
field
to
to
onboard
data.
So
that's
kind
of
a
lot,
and
another
thing
in
here
is
like
it
will
probably
take
about
half
a
year
to
fill
up
the
storage.
A
So
if
you
just
fill
it
all
up,
so
obviously
you
can
also
do
the
math
on.
If
you
only
start
up
with
having
24
drives
in
there.
Just
having
this
set
up,
24
drives
how
long
time
would
that
take,
and
you
know
what
is
the
schedule
for
when
I
need
to
buy
new
drives,
so
I'm
ready
to
put
them
in
so
they
are
initiated
and
ready
to
go
when
I'm
out
of
storage
from
the
first
arrays.
So
that's
that's
perfectly
easy.
A
I
I've
been
you
know,
starting
out
with
the
first
12
drives
and
then
24
and
then
gone
from
there
so
yeah.
I
think
it
makes
totally
sense.
But
then
again
you
know
with
the
hardware
scarcity
and
so
on
might
make
sense
to
just
be
on
top
of
it
and
just
make
sure
that
you
have
all
the
hardware
available.
A
A
A
You
will
have
that
could
cause
a
breakdown
and
what
I
really
want
with
the
minor
note,
and
that,
like
the
physical
hardware,
there
is
that
it's
in
a
setup
where
it
just
doesn't
break
and
obviously
consumer
heart
grade
hardware
today,
is
also
pretty
good,
but
it's
like
if
it
does
break
and
you
have
a
ton
of
collateral
in
there
and
your
miner
has
no
way
to
prove
your
data.
It's
like
that's,
probably
not
a
place.
You
want
to
be
so
it's
like
you
can't
go
cheaper
than
this.
A
Obviously,
but
it's
also
a
matter
of
the
risk-
and
we
see
like
if
we
go
down
to
a
system
like
this,
and
we
only
have
those
limited
raw
capacity
because
of
the
two
arrays,
then
we
would
also
have
a
bigger
hardware,
cost
percentage,
so
obviously
yeah.
We
need
to
spend
more
hardware
and
we
get
less
amount
of
adjusted
storage
power
out
of
it.
A
A
So
that's
it
as
last
time.
I
did
bring
some
pictures
just
to
show
what
I
have
been
working
on
so
currently
the
setup.
I'm
just
gonna
move
my
my
picture
down
here,
but
up
here
in
the
left
top.
This
is
a
look
into
one
of
these
node
servers
that
I've
been
building
and
basically
what
I
like
about
these.
These
are
the
super
micros
bare
bones.
A
Obviously
I
would
have
down
time
if
I
like
prune
the
chain
or
do
a
loot
upgrades
and
so
on,
but
I
would
really
like
it
not
to
be
because
I
have
to
do
hardware
maintenance
this.
No,
this
just
needs
to
be
like
rock,
stable
and
like
never
have
to
touch
it.
Then,
next
to
it,
you
have
the
the
hba
which
is
holding
all
the
drives
on
this
system.
It's
actually
both
internal
and
external
port
hba.
A
So
it's
like
yeah,
it
is
basically
holding
the
24
drives
internally
and
then
it's
also
connected
to
48
drives
externally,
because
I
use
these
nodes
as
sda
bots.
Then,
next
to
it
you
have
the
a4000.
I
only
have
one
carbon
here
on
this
picture,
but
that's
basically
that
little
gpu
sitting
there
perfectly
doing
all
the
winning
post
and
window
post
then
down
in
the
left.
Here
we
have
one
of
these.
A
So
you
have
the
32
dimms
there
you
have
the
cpu
coolers
and
then
you
have
your
gpu
thing
there
in
the
back
and
basically
because
we're
using
the
the
the
qatar
rtx
a
5000
there,
they
sit
perfectly
in
a
server
and
if
you're,
trying
to
use
any
of
the
consumer
nvidia
gpus
in
there.
No,
I
just
can't
recommend
it.
It's
a
it's
a
bad
choice
of
design.
You
see
these
fit,
obviously
very
nice.
In
there
they
have
a
power
consumption,
which
means
that
they
don't
go
past
60
70
degrees
celsius,
it's
all
good.
A
So
so
that's
basically
what
goes
down
there
right
here
where
my
picture
is.
We
have
what
my
backup
note
looks
like
right
now.
One
funny
thing
is
like
I
couldn't
really
get
a
real,
proper
cooling
out
of
it
with
the
t4.
So
what
I
had
is,
I
have
a
3d
printed
shroud
here
for
it
with
a
normal
60
millimeter
fan
in
front
of
it
and
yeah.
A
If
you
had,
you
could
probably
also
tweak
it
somewhere
different
and
then
get
a
better
performance,
but
one
of
the
issues
having
a
test
line
here
is
like
this
is
really
meant
for
having
a
gpu
specific
server
which
can
blow
a
ton
of
air
at
it,
because
the
little
g4
there
is
a
passive
card,
so
different
ways
of
doing
it.
But
I
find
this
you
know
really
efficient,
because
I
can
actually
hold
down
to
below
60
degrees
celsius
like
that
and
over
here
at
the
cpu.
A
You
also
have
the
shroud
there,
but
I
didn't
have
any
more
of
the
passive
cpu
fan,
so
I
just
use
this
one,
but
it's
good
to
know
that
it's
basically
forced
there
by
by
the
redundant
fans
so
right
about
that.
This
is
an
example
of
what
it
looked
like
earlier
on
in
the
in
a
in
a
yeah.
In
a
rack,
these
servers
are
rack,
mountable
and
obviously
yeah.
A
A
And
now
you
have
to
rebuild
everything,
because
now
you
actually
got
into
a
size
where
you
needed
to
be
in
the
data
center
and
now
nothing
fits
also
be
aware
that
data
centers
might
have
restrictions
on
exactly
what
hardware
they
would
allow
in
there
having
a
super
micro
server
like
this
is
just
like:
that's
the
most
normal
thing
so
over
in
the
right
side.
I
have
some
of
the
more
fun
stuff,
so
that's
what
I've
been
using
here.
A
You
know
from
the
first
video
I
also
showed
a
little
bit
about
using
my
my
floor
heating
to
provide
a
cooling
for
the
server
or
put
it
the
other
way.
I'm
my
ceiling
server
is
actually
heating,
my
house,
so
that
was
that's
mostly
relevant
for
the
that
little
test
equipment
that
I've
been
running
here
at
home,
because
it's
obviously,
if
you
have
like
a
one,
pip
roll
miner
or
like
five
pips
of
storage,
powder
and
collateral,
you
would
probably
not
have
that
at
home.
A
Water,
cooling
and
exotic
everything
doesn't
matter
that
much
the
more
you
go
into
the
kind
of
the
the
minor
note
and
all
that
that
needs
to
be
online.
It's
on
mainnet,
it's
real
us
dollars
the
more
it
needs
to
look
like
an
enterprise
redundant
server
and
not
like
this
kind
of
stuff.
So
my
lesson
from
the
very
beginning
was
a
space
race.
I
basically
had
one
big
pump
that
was
recirculating
all
the
the
computer
liquids
into
the
system
that
we
see
here.
A
I
ended
up
figuring
out
that
I
was
relying
on
two
different
pumps
at
the
same
time
that
they
would
be
online
and
working
and
if
they
shut
off.
I
don't
know
exactly
what
would
happen
to
my
hardware,
but
probably
that
temperature
on
the
water
would
just
rise
like
astronomically
and
I
might
start
breaking
stuff,
and
so
when
we
went
to
main
net
that
was
kind
of
a
concern
that
that's
a
lot
of
risk.
To
take
that
you
need
x,
different
things
to
work
perfectly
or
everything
will
break,
and
that's
why.
A
I
think
that,
on
that
side,
where
you
have
like
the
refill
and
so
on,
better
go
with
some
enterprise
grade
hardware.
If
we
are
into
ceiling,
you
know
yeah,
you
could
lose
your
data
while
sealing,
so
that
would
be
a
little
bit
annoying,
but
you
could
also,
if
you're
doing,
testing
doing
cc
and
so
on.
It
doesn't
really
matter
if
it
breaks
up
and
you
have
to
clean
up
it's
like
nothing.
Nothing
really
bad
happened.
A
One
last
point
here
that
I
wanted
to
point
out
on
this
side
is
like
you
see
up
there
that
I
have.
Actually.
I
can
move
my
head
up
there
so
just
around
here
you
have
those
two
fans
sitting
right
next
to
the
the
border
cooler.
So
this
is
a
good
example
of
what
happens
if
you
try
to
run
with
something
which
is
a
little
bit
more
consumer
oriented.
A
This
is
not
consumer
hardware
at
all.
This
is
a
this
is
a
intel
single
cpu,
and
what
happens
is
that
you
have
your
memory,
dimms
they're,
sitting
right
next
to
the
cpu
16
of
them,
but
because
this
is
made
for
a
desktop
computer
case.
They
put
it
in
that
direction.
Therefore,
they're
not
taking
up
any
of
the
cpu.
Sorry
the
pc
express
lanes.
So
it's
like
you
have
all
the
availability
for
the
cards
in
there,
but
when
you
have
the
memory
sitting
like
that,
you
have
a
really
bad
airflow
and
the
problem
is
like.
A
A
That
is
really
bad,
and
this
is
some
of
the
issues
that
you
get
into.
So
that's
back
to
the
initial
question,
like
hey
man,
why
don't
you
just
build
this
on
threadripper
and
you
can
do
threadripper
pro
and
it
can
have
two
chips
of
memory.
That's
much
easier!
Well,
think
about
it,
a
threadripper
pro!
You
probably
can't
find
for
now
any
rack
mounting
solution
for
that.
You
could
take
this
a
motherboard
or
a
threadripper
pro
motherboard
like
this
and
put
it
in
one
of
these
servers,
as
I
showed
right
here.
A
But
if
you
have
your
memory
sitting,
probably
like
you
do
right
here,
you'll
have
a
issue
how
to
cool
that
stuff,
because
you
basically
relying
on
turbulence
around
them
to
cool
them.
While,
if
you
look
down
at
the
left
right,
you
see
a
setup
which
is
made
for
like
perfect
airflow,
and
this
would
no
never
have
any
issues
with
the
airflow
in
a
bare
bone
setup
like
this,
but
with
something
like
this.
A
You
know
you
probably
just
get
into
these
kind
of
solutions
where
it
was
cheap
and
you
got
a
very
efficient
setup
and
so
on.
But
you
end
up
with
something
where
you
have
like
fans
sitting
around
and
basically
you
know
this
would
have
been
better
of
just
being
in
a
desktop
case
and
trying
to
put
it
in
a
rack
case
that
just
gets
a
little
bit
annoying.
A
So
that
was
just
the
main
part
that
I
wanted
to
to
do
on
on
the
consumer
hardware.
Then.
Yes,
it
works
perfectly
fine
if
you
just
want
a
really
simple
setup.
If
you
just
want
to
try
this
out,
probably
you
want
to
do
that
on
the
test
net,
not
using
real
fill,
not
using
real
us
dollars
on
the
network,
but
if
you
are
moving
into
mainnet,
no
minimum
will
be
that
you
have
10
tips
of
adjusted
power.
A
That's
not
a
lot!
I
know,
but
10
tips
of
the
verified
power
right
now
or
sorry
of
adjusted
storage
power
will
give
you
a
one
block
reward
like
I
don't
know,
maybe
in
a
year
if
you're
lucky,
so
it's
like
you're,
not
going
anywhere.
So
if
you
go
to
mainnet,
I
would
guess
is
because
you
want
to
store
real
data
is
because
you
want
to
mine
blocks
or
something
like
that.
A
So,
if
that's
your
ambition
to
get
your
hardware
cost
back
again,
you
do
need
some
kind
of
size
and
if
you
do
need
some
kind
of
size,
you
need
to
think
about
that.
This
might
end
up
in
a
data
center
like
everyone
else
is
doing,
you
might
co-locate
to
a
data
center,
and
if
you
build
this
on
consumer
hardware,
which
is
not
really
fit
to
be
in
a
server,
probably
you
wouldn't
have
you
co-locating
a
site
liking
that
you
put
like
all
kind
of
water
cooled
rigs
in
your
in
your
rack
there.
A
So
it's
like
you
need
to
think
about
like
if
this
is
just
for
fun,
and
this
is
just
to
get
started
and
so
on
by
all
means.
Go
ahead,
do
something
it
might
not
be
exactly
enterprise
hardware,
all
of
it,
but
think
about
the
long-term
goal
here
if
it
is
about
putting
it
into
a
data
center
at
some
time.
If
you
succeed,
it's
a
little
bit
annoying
that
you
have
to
sell
off
all
this
consumer
hardware
and
then
have
to
buy
all
new,
because
now
it
needs
to
be
enterprise
great
and
believe
me.
A
I've
been
through
all
this,
so
the
only
thing
good
I
did
was
to
start
off
by
an
epic,
so
I've
been
keeping
that
for
the
whole
time,
but
most
of
the
other
hardware
here
needed
to
be
replaced,
yeah,
not
spending
too
much
time
on
it,
but
but
the
one
that
I
got
right
over
here
next
to
the
water
cooling.
That's
actually
one
of
the
test
machines
that
I
have
right
now
and
it's
really
like.
I
don't
think
you
can
go
silver
grade
more
low
key
than
this.
A
It's
I
kind
of
ripped
apart,
a
2u
chassis.
It
got
the
water
cooling
in
there.
What
you
see
right
there
in
the
side
with
lyson.
That's
actually
my
old
30
90
founders
edition,
which
is
up
here
in
the
upper
right
corner
and
yeah.
It's
water
cooled!
So
it's
it's
a
very
different
size
from
when
it
had
that
giant
fan
on,
but
right
now
it's
just
sitting
with
the
pc
express
riser
on
there.
So
it's
just
lying
on
the
side.
A
So
this
is
like
it's
a
really
really
cheap
design,
but
the
reason
why
I
have
it
sitting
like
this
and
looking
like
this-
is
because
I'm
only
using
it
to
come
up
with
these
results
that
I
showed
here
today.
So
I
really
hope
that
you
liked
this
video
and
you
found
it
interesting.
Some
of
the
points
that
I
have
tried
to
make
here.
A
A
So
what
I
wanted
to
point
out
here
on
the
last
slide
is
like.
If
you
want
to
get
more
into
this,
you
can
go
to
the
slack
community
channel,
which
is
basically
where
most
of
this
happens.
So
this
is
where
you
can
get
help
with
your
setup.
This
is
where
you
can
talk
to
the
rest
of
the
community,
about
how
they're
doing
stuff,
how
they're
setting
it
up,
obviously
there's
no
single
way
to
doing
stuff.
But
if
you
have
an
issue,
this
can
be
a
really
good
place
to
go.
A
There's
also
a
github
site
there.
Both
github
itself
for
the
falcon
project
is
really
good
with
a
lot
of
content
in
there.
A
lot
of
issue
tracking
and
everything
going
on
in
there
and
also
the
new
releases
will
be
in
there,
but
also
this
specific
site
also
holds
some
links
and
everything
to
the
different
community
sites,
also
slack
being
one
of
them.
A
A
But
if
you
have
interested
in
that,
the
the
link
is
there
and
it's
an
accelerator
which
will
take
on
you,
know
50
participants
for
now
brought
in
in
different
cohorts,
so
it's
like
open
for
application
for
the
next
cohorts,
but
I
would
also
like
to
point
out
that
there
will
be
other
accelerators
coming
up.
You
could
call
like
a
operator,
training
programs
and
right
now,
espa
is
running
in
north
america,
europe,
asia.
There
will
also
be
a
lot
of
this
training.
A
Also,
the
espa
training
material
will
be
come
out
in
the
public,
so
it
will
basically
be
available
both
for
these
other
training
courses,
but
it
will
also
be
available
directly
if
you
just
want
to
get
to
know
more,
but
I
hope
that
this
was
kind
of
a
teaser
on
what
you're
up
for,
if
you
want
to
do
real
data,
and
you
want
to
do
falcoin
if
you
are
storage
provider
out
there.
I
hope
that
this
gave
you
some
inspiration
on.
A
You
know
what
what
what
you,
what
you
could
change
and
what
you
could
look
out
for
to
optimize
your
current
setup,
and
that
was
basically
all
the
for
this
one.
Let's
see
if
there
will
be
another
round,
once
hardware
sealing
speeds
and
everything
is
updated
again
and
we
have
something
else
for
the
market
and
so
on.
I
am
enjoying
myself
very
much
on
this
network.
I
think
it's
one
of
the
most
promising
blockchain
technologies.
A
Basically,
I
think
that
we're
just
getting
started
we're
just
getting
where
you
know.
Basic
things
are
working
now
we
can
store,
we
can
retrieve
and
that's
starting
to
really
look
good
but
think
about
it
when
we
move
into
actually
also
doing
like
extensive
smart
contracts
being
able
to
control
how
we
are
storing
our
data
from
a
client
perspective
being
able
to
compute
on
those
data,
we
have
all
this
hardware
already.
We
have
tons
of
gpus
and
everything.
So
why
not
also
being
able
to
do
that?