Threadripper For Disruptive AI Research

Hardware Build a PC
1

I’m assembling a Threadripper build for the purpose of doing private AI research. Of course, other than learning a little basic on a TI99 when dinosaurs still ruled the earth, having studied a little electronics, and having assembled two pretty modest gaming rigs, I know absolutely nothing about the subject, but that’s never stopped me before. :smile:

The AI research I want to do involves using the same tiny fpga arithmetic accelerators currently being added to almost every graphics card and cpu being manufactured. These can be as small as 120 transistors, but produce compelling, if rudimentary, animations of any motion and behavior imaginable. Mathematicians have been busy assembling algorithms for leaves swirling in the breeze, lightning, etc. As rudimentary as these animations are, they can be equally complex with a single algorithm already being discovered that can animate a bot to follow sidewalks and crosswalks, jump, crouch, run, walk, roll, avoid obstacles, step over a stick thrown down in its path, and get and continue on if knocked down. My hope is to create a distributed computing platform that can run 64 of these animations simultaneously within the players immediate proximity and field of view using a simple adaptive AI program that animates everything according to what you are looking at or happen to be near at the time.

Many have eagerly anticipated the current dramatic changes in PC architecture that make it possible to play a serious game of Three Card Monty between the system ram, processor, and graphics card in this manner, with fpga circuits commonly being embedded in gram. The drawback is that these circuits are relatively slow right now, commonly rendering perhaps 8-16 fps, but I intend to get around that by using a zero latency MindMaze headset. The VR headset is manufactured by a major medical research group and can literally know what you are going to do before you do, because it reads your brain waves to provide the computer with 32-70ms of advanced warning before you can so much as move a muscle.

That kind of advanced warning is simply unheard of, would completely eliminate traditional lag problems such as draw distances, and means the computer can buffer frames in advance for any animation. Theoretically, with future progress, a next generation laptop could run such a program using entirely open source circuitry and programs in the public domain. By merely giving the system the ability to model a chicken flock pecking order it can be made significantly self-correcting and RPG programmers in particular might feel right at home adding more complex AI to the system.

That would make it a self-assembling AI program and video gaming platform constructed entirely within the public domain, extremely compatible with Vulcan and the new Linux 3D fonts which appear likely to be adopted as the new industry standard. My personal interest is in translating the fuzzy logic of the Tao Te Ching into such an interactive VR platform. It requires studying at least six to ten English language translations of the text for at least fifteen years to be considered the least bit competent with its outrageously convoluted fuzzy logic.

Like I said, I know next to nothing about the technical side of all this, just enough electronics and computer programming to get myself into serious trouble. However, designing an analog computer platform is as much an art as it is a science, and that’s where I come in. Currently, I’m in the seventh year of writing a book on the subject, titled “Truly Ignorant Wisdom”, and the design is intended to go in the last chapter of the book as empirical evidence to support the fuzzy logic.

Any suggestions on hardware are extremely welcome. So far I’ve assembled a phanteks enthoo luxe tempered glass class with three thick alphacool radiators for water cooling the entire motherboard, including the VRMs and south bridge, and gotten an Enermax Platimax 1200w power supply and am waiting for the new x499 chipset to be released. The biggest question I have is which graphics cards to use, and I am tempted to wait for Navi to be released if necessary, because I’m on a fixed income and can’t afford to do this twice.

2

I don’t know why this post caught my eye and maybe its too late, but some brutal honesty that may or may not help a fellow tinkerer with a basement full of humming computes:

  1. If you just want an excuse to build a killer rig, then you already have it. It’s Friday in March, why not?

  2. Water cooling is going to add significant cost/complexity with very little if any benefit to a Threadripper setup in particular. It does about as well with Noctua’s top-of-the-line heat-sink as its going to do. Particularly if you don’t plan on overclocking it to the bleeding edge.

  3. I find one science experiment at a time is best. I like rock-solid, can’t fail, could not be simpler hardware when I’m trying to break ground elsewhere (limit variables, focus your energy on the solution, not the tool).

  4. If you are the least bit cost conscious, you should only optimize for all-out throughput once you’ve made your way through a proof-of-concept with lesser hardware. You can prove your thesis on A and extrapolate to B based on observed scaling from 1 to N nodes (understanding the fudge factor and asymptotic limits inevitably finding Amdahl’s law somewhere in the system)

  5. If you are just starting with openCL/cuda and the like now, see all of the above again. You are standing at the foot of a steep learning curve. Spend your time now, spend your money later as prices will almost always (the last 18 months being a notable exception) make throughput-per-dollar better in the future. Prove our your idea, then ebay an older model quadro card to unlock features if needed (say fp64/int64 throughput).

I generally make very good use of my tinkering hardware acquisitions and I more than understand that having capability at your fingertips can be inspirational, but I’d be lying (a lot) if I said I never bought something before I needed and from that I learned to pick my battles.

I say all of that having ebay’d some 40GbE NICS and a router in the past few weeks to up my NFS/SAN game to support big-data computes in my noisy basement.

Good luck and happy disruption.