This posting is on-topic because any serious AI effort will,
necessarily, need to configure and purchase a machine capable of running
it.
Here's a reply I got from the Spaun mailing list:
#######################
Spaun is a rather hefty beast. It requires at least 24GB of ram to run
the model. The latency timings of the ram doesn't matter so much, it
would just affect the running speed of the model itself.
Spaun also requires about 2.5 hours of real world time to run 1 second
of simulated time. Note that this run time estimate is based on Spaun
running on a 8 core (16 hyperthreaded cores) 2.6GHz machine.
######################
First, D-ram itself; it sucks.
http://en.wikipedia.org/wiki/Dynamic_random-access_memory
Basically it has to go through an 8-step process to detect the charge on
a 2-3 pf capacitor. The entire process takes, according to a Kingston
datasheet, around 48 nanoseconds (minimum). (equivalent to roughly 20
mhz) One consistent rule though is that the smaller/denser you make the
cells, the SLOWER it is for the circuitry to read the increasingly faint
signals from each cell.
High datarates are achieved by the aggressive use of bank interleaving.
A die might have 8 or more separate banks in order to support high data
rates, but only for burst reads/writes.
A D-ram cell does require several wires. On the horizon is memristor
based memory that will have a cell-pitch of only 10nm and will be
stackable in three dimensions.
There is a new memory interface on the horizon that will continue the
trend towards using multiple packet-switched links over traditional
buses. But you can't get those yet.
The existing technology is DDR3. There are two closely related flavors
of this standard, DDR3 which is used in consumer desktops and RDDR3
which is used in servers.
DDR3 is faster but it has a strict limitation of two DIMMS per channel.
So a dual-channel board such as mine will have 4 slots. A 3-channel
machine will have six slots. This is a strict limitation caused by
electrical issues. Practically speaking, an AMD machine can have up to
32 GB and an intel machine up to 48GB of ram. [does fact checking,
chipzilla (intel) now has a quad-channel chip. I am terrified of an
intel monopoly and won't use their chips unless under duress] This is
why upgrading from where my system is right now is not trivial. I'm at
the point where any change I make would be a trade-off. I would be
sacrificing latency for capacity.
A server, using RDDR3 can have 4 dimms/channel. It can do this by adding
a "register"/buffer (amplifier) to the dimm that will add a clock or two
to the latency but make it possible for the controller to drive
additional slots. An AMD socket G34 processor has four channels, so you
can have up to 128GB per socket. G34 machines can have up to 4 sockets
for compute nodes, or 2 sockets for IO/GPGPU nodes. So you can get up to
half a TB of ram in a server.
Some interesting boards are:
http://tyan.com/product_SKU_spec.aspx?ProductType=MB&pid=696&SKU=600000215
(supports up to 4 GPUs, but 2 sockets, marketed as a GPGPU platform)
http://www.supermicro.com/Aplus/motherboard/Opteron6000/SR56x0/H8QGL-6F.cfm
(Supports up to 3 GPUs, 4 sockets)
http://www.supermicro.com/Aplus/motherboard/Opteron6000/SR56x0/H8QGi-F.cfm
(supports 1TB of ram but only 2 GPUs)
At this point, I would tend to value GPGPU capacity over the number of
CPU cores because GPGPUs can be up to 100 times faster than the host
processors.
--
E T F
N H E
D E D
Powers are not rights.
-------------------------------------------
AGI
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424
Modify Your Subscription:
https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657
Powered by Listbox: http://www.listbox.com
