Hmmm… Manchester is only an hour’s drive away from me, wonder if they would miss a few nodes… I could get one of those racks in the back of my car… where’s my sack barrow.
I’ve looked at their specs, it’s built around standard PC architecture using a MPI suite, master I/O terminal, nodes/ cores and Ethernet for communication/ data transfer.
The ‘neuromorphic’ moniker is a sales pitch; there is nothing about the design that lends this setup more towards neuromorphic computing than any other, don’t get me wrong it’s a nice versatile setup, but wont be any faster for neuromorphic computing than any other ‘super’ computer. And as usual the main bottleneck to this design is the network… queues, buffers and handshaking, etc are a consequence of our current level of technology and the enemy of parallel processing speed.
They started in 2011 and cost has obviously been an issue. Given the fact that they are researching neural structures and currently have no idea how they work, and so what hardware is actually required, this is the only machine they could have realistically built… which is a pity.
The ideal machine for neuromorphic computing would be a single shared massive memory core with the million cores on the same bus. The model would exist within the memory core and its state could be read or written too directly by the cores, they can obviously simulate this architecture but the MPI and Ethernet will slow processing down immensely.
The good news is they could improve this system for neuromorphic processing just by changing the network topology. I can’t find specifications on their current wiring or protocols but I bet because they are using off the shelf parts they are using standard switches/ hubs and TCP/ UDP protocols, two NIC cards in each node bank, for I/O. They need clusters of nodes with the same IP and two I/O streams/ cards running half duplex, cables actually split so the nodes Ethernets are wired in parallel… omni-directional protocols, etc… more like a ‘brain’ less like a computer network.
Who currently feels constrained by the hardware that they have access to?
Now that I have a working theory/ model I’ve recently started scaling up my neural core model, as the core expands and ages, more neurons/ synapse/ etc are added and the simulation obviously begins to slow down, this is not a major problem at the moment but will be in the near future. I have to subdivide the model periodically for efficient parallel processing on my cluster, but each node is reaching its limits for real time simulations.
What do you think you could do if you had the best hardware (or software) that money could buy?
From my perspective having the best machine in the world wouldn’t help ATM, throwing more processing power at a bad model is pointless. Processing power is only a very small part of the AGI problem. Standard computer architectures are not optimal for this problem, which I believe can only be solved by simulating biochemical/ electrical structures, the hardware needs to be optimized for biological simulations, not generic logic processing.
The problem has been ‘how’, not ‘how fast’.
What do you think you could accomplish if you had a great software development team working for you?
Probably less than I do now, lol. I’ve worked with teams before and they can be very counter productive, the structure of the team is very important.
It’s the network topology problem rearing is ugly head again; bi-directional communication takes time, explaining your ideas, trying to get everyone on the same page, writing specifications for coders to follow, other members are only human… and so want to express/ explain their ideas.
The interpersonal communication required, and as a consequence extra mental load created, often stifles the team’s creative abilities/ members… and don’t get me started on daily/ weekly/ monthly meetings or even worse… team building exercises WTF.
Pamela, a 20 year old from PR thinks you are mentally distant and not participating in the team efficiently... Pamela... Fu*k Off.
Fancy a pint and a game of pool... or a steak... I'm there... or give me a paint ball gun so I can shoot Pamela all at point blank... I'm in. (I'm old and grumpy lol)
My main problem is time; I have to run my business to put bread on the table and keep my immediate family happy... ish. My ideal scenario would be to have an income and the time to work without distraction, I only get a few hours a day to think about/ program/ solve the AGI problem space. Up until last year I had it sorted, contracts coming in, very little work needed, £200 an hour average for the hours I actually worked... but the government installed school trusts and the money dried up, I've got to start again now.
Luckily I'm nearly finished with my creation... the world is about to change... this should bring a few bob in, hehe.
So yeh!...ideally... Just I and my machines… if you want a progress report… just ask the machine lol.