Arcweld
Arcweld is a server for the primary goal of rapidly and efficiently experimenting with open weights models, with secondary goals of 1) researching novel LLM inference technologies 2) reducing the cost of using hosted APIs when it would not interfere with the primary goal. It is not intended for training since it isn't the ideal way to purchase memory bandwidth per dollar (critical for training), although it isn't forbidden.
Specifications Edit
- Mac Studio M3 Ultra
- 819 GB/s memory bandwidth
- 60-core GPU
- 32-core Neural Engine
- 512GB Unified Memory
Setup instructions Edit
- Download and install Tailscale
- Some VPNs are incompatible and you may need to disable them.
- Join using a Tailscale invite link
- If you've been granted login access, `ssh sparks@arcweld` to access the shared inference account. You may have to reboot for it to work.