Yes, I get it, X algorithm, you have intimate knowledge of my every day life I’ve never posted online, but have you considered, instead of studying me, helping me?
7
Superintender retweeted
Replying to @super_intender @taylorlapeyre
Physicist dies and goes to Heaven, upon which he asks God why He made the speed of light as arbitrary as 186,282 miles per second. “What do you mean?” God replies. “The speed of light is 1.”
DDR4 local server MOBO supremacy
Replying to @dee_hw
Did a cost analysis on this a couple weeks ago. If local inference is your only use case for the cards, then r9700 pro beats RTX cards in terms of cost:vram, every single time, but they can only talk over PCI, so you need a solid CPU+MOBO+RAM combo.
The open secret with local inference is motherboard specs. If you only plan for 1 GPU, then a consumer DDR4 MOBO almost always has a single x16 PCI-4|5 slot. If you want two cards on a consumer DDR4 MOBO, you're almost always are constrained to x16/x4 split, while rare x8/x8 DDR4 boards exist, but rarer.
If you're building a "local inference only rig", that another machine connects to (desktop, laptop), then you're better off building from a DDR4 server motherboard, which is still much more expensive than a consumer motherboard, but saves money on components.
Look at the trade-offs here:
25
Use chinese models as manufacturing plants, SOTA models like the S&P 500, and local inference like SkunkWorks…
16
Computa, rip out my spine like The Predator, crack it like a whip, then put it back carefully please, make no mistakes
1
40
1. Demographically Segment detractors
2. Identify significant others (friends)
3. Neurolinguistically program others
4. ….
5. No more detractors left to worry about
28





