Skip to content
LocoLLM LocoLLM

ADR-0005: WEIHO 8-GPU Enclosed Chassis Replaces B250 Open Frame for Colmena

Status

Section titled “Status”

Accepted

Date

Section titled “Date”

2026-03-07

Context

Section titled “Context”

ADR-0004 adopted a B250 Mining Expert motherboard in an open air frame as the multi-GPU platform (Colmena), fully retiring Cerebro. During hardware sourcing, the plan changed substantially:

  1. A WEIHO 8-GPU enclosed mining rig (72x42x18cm, steel chassis, integrated 2000-3300W PSU) became available. It provides 8 native PCIe slots without riser cables in an enclosed, cooled chassis — a better fit than an open frame with USB risers.
  2. Cerebro should not be fully retired. It continues to run CloudCore Networks, a separate project from LocoLLM and LocoBench. One RTX 2060 Super remains dedicated to CloudCore inference in Cerebro.
  3. LocoBench’s benchmarking philosophy crystallised around floor cards per VRAM tier rather than multiple identical cards. Three matched 2060 Supers are no longer needed — one 2060 Super represents the 8 GB Turing floor.
  4. The host system is deliberately constrained (i3-3220, 8 GB DDR3) because LocoBench benchmarks GPU capability on modest hardware, which is what most users actually have.

Decision

Section titled “Decision”

Replace the B250 Mining Expert open frame design with the WEIHO 8-GPU enclosed chassis for Colmena. Supersedes ADR-0004.

Hardware:

  • Chassis: WEIHO 8-GPU enclosed mining rig
  • Motherboard: Intel LGA1155 (likely B75/H61 chipset)
  • CPU: Intel i3-3220 (Ivy Bridge, dual core)
  • RAM: 8 GB DDR3 SODIMM (board maximum)
  • Storage: 128 GB mSATA (OS) + WD Scorpio Blue 750 GB SATA (model storage)
  • PSU: Integrated 2000-3300W
  • Cooling: 4x 120mm fans (to be replaced with Arctic P12 PWM)
  • 8 native PCIe slots, no risers

GPU lineup — floor cards per VRAM tier:

CardVRAMTier Role
GTX 1050 Ti4 GBFloor of 4 GB tier
GTX 1060 6 GB6 GBFloor of 6 GB tier (pending acquisition)
RTX 2060 Super8 GBFloor of 8 GB Turing tier
RTX 3060 AORUS Elite12 GBFloor of 12 GB tier
RTX 3090 (reserved)24 GBReference ceiling (work budget)
3 slots reservedTBDFuture expansion

Cerebro stays active:

  • One RTX 2060 Super remains in Cerebro for CloudCore Networks (separate project)
  • Second 2060 Super (Michael’s testing card) migrates to Colmena

Three distinct projects with distinct hardware:

  • LocoLLM (Burro + Colmena) — infrastructure, architecture research, adapter training
  • LocoBench (Colmena primary) — benchmarking platform, community results
  • CloudCore Networks + Cerebro (standalone) — student assessment simulation

Consequences

Section titled “Consequences”
  • LocoBench benchmarks use floor cards per tier, producing conservative baselines — if it runs here, it runs on your card
  • Enclosed chassis replaces open frame; card visibility is traded for better cooling and a smaller footprint
  • 8 native PCIe slots (no risers) simplify the build and avoid x1 electrical bandwidth concerns
  • Deliberately constrained host (i3-3220, 8 GB RAM) means sequential benchmarking rather than parallel, but results are identical
  • Cerebro is no longer retired — it remains active for CloudCore Networks with a dedicated 2060 Super
  • CloudCore Networks is cleanly separated from LocoLLM and LocoBench infrastructure
  • RTX 3090 reserved as reference ceiling validates whether floor-tier results scale predictably across the VRAM range
  • ADR-0004 is superseded; the B250 Mining Expert open frame design is not being built