Ok, let's try to evaluate the hardware problem on a more abstract level
to try to get a better handle on what the missing pieces are.
A computing system has three fundamental functions:
1. memory -> how much information can the computer contain?
2. Processing -> how can the computer manipulate what's in its memory
and how fast.
3. IO and networking -> how can it interract with other computers and
the rest of the world.
Networking is important because whatever you need to achieve human
equivalency, the very next day you will want to add 100x that to the
network and the day after that 10,000x the original...
I can't comment too much about network theroy or current best practices
in network design. I'll point out that high-end motherboards support
10gb ethernet and 40 and 100gb switches are available.
http://www.hpctoday.com/best-practices/10-things-to-know-before-deploying-10-gigabit-ethernet/
http://www.mellanox.com/page/products_dyn?product_family=280&mtag=sn3000_label
With conventional hardware there are two types of IO bandwidth, there is
bandwidth to the network, and there is bandwidth to secondary processors
such as GPUs, FPGAs, and ASICs.
Everyone should be familiar with commodity hardware. It's cheap.
Specialized hardware costs 10x as much as commodity hardware,
and custom hardware costs 10x or more as much as specialized hardware.
So therefore you'd better have a damn good reason for looking beyond
commodity hardware.
AM4 processors are pretty good for thin edge nodes, they don't have many
IO channels but they're extremely fast. In order to run dual video
cards, the existing IO channels need to be multiplexed by the 570 chipset.
The TR4 (workstation) processors have twice as much memory bandwidth and
64 (iirc) IO channels, which provides ample connectivity for accelerator
boards.
There is also a SP3 platform (iirc), for the EPYC server processors
which are designed for energy efficiency so aren't as fast in
single-threaded performance but are massively parallel and support 8
memory channels and HUGE IO.
Pending further information about what is actually required, I think the
TR4 platform is a very interesting choice. It is reasonably priced and
has the memory capacity and IO capability to support a variety of
accelerator hardware regardless of what it may turn out to be.
I expect the 3000 series threadripper to be available around January of
2020 and hope to get one for myself somehow.
--
Clowns feed off of funny money;
Funny money comes from the FED
so NO FED -> NO CLOWNS!!!
Powers are not rights.
------------------------------------------
Artificial General Intelligence List: AGI
Permalink:
https://agi.topicbox.com/groups/agi/Tdffdc158bb409445-M4e0e18b9b0299f575ba5001e
Delivery options: https://agi.topicbox.com/groups/agi/subscription
