Architecture: the lab at a glance

Five x86 hosts and an eleven-node ARM cluster: what runs where, with the full hardware below.

heterogeneous compute multi-factory fleet virtualization storage tiers

Click the map to enlarge.

Host specifications

HostCPURAMGPUStorage
MothershipRyzen 7 5800X · 8c/16t32 GBRadeon 7900 XTX 24 GB~4 TB SSD
BigHonkerCore i5-9600K · 6c16 GBRTX 3070 8 GB2 TB HDD + 1 TB SSD
SoftserveRyzen 7 5700G · 8c/16t48 GBRTX 3070 Ti 8 GB46 TB HDD + 1.5 TB SSD (shared)
RigbyCore i7-3770 · 4c/8t24 GBGeForce GT 730 · 2 GB1 TB SSD + 1 TB HDD
MordecaiCore i5-3470 · 4c8 GBnone500 GB SSD

Fellowship node classes

ClassProcessorRAMNodes
A8-core ARM64 · 2.8 GHz3 GBgandalf, aragorn
B8-core ARM64 · 1.8 GHz2 GBfrodo, sam, legolas, gimli, boromir, faramir
C4-core ARM64 · 1.8 GHz2 GBmerry, pippin
D4-core ARMv7 · 1.4 GHz (32-bit)1.4 GBsméagol

No two of these machines are alike: a 24 GB Radeon training box, two 8 GB chat cards, a pair of decade-old Ivy Bridge desktops, and an ARM cluster with 2 GB a node. The fleet grew one box at a time, and it works as one pool because each job is placed on the host whose capabilities fit it, not because the hardware matches. Softserve, the main server, anchors it: Plex, the network-shared array, and the Hyper-V layer that carries the Home AI orchestrator and the media stack as VMs. Heavier model work spills onto the other boxes because no single one holds it all. The other four maps break the fleet down from here: what runs on it and how work is placed are the AI and compute and scheduling maps, how the AI changes its own code under human gates is the self-engineering map, and the zero-trust front door is the network and security map. The current lab page has the host-by-host detail.