Blackthorn

The physical hall where Guildhall runs. Named after the Blackthorn — Dalinar Kholin’s warrior identity in the Stormlight Archive. A processing beast that sleeps until called.

Hardware

• CPU: 2× Intel Xeon Gold 6148 — 40 cores / 80 threads total
• RAM: 768GB DDR4 2133 MHz (12×64GB)
• GPU: 2× NVIDIA GeForce RTX 2080 (8GB GDDR6 each)
• Storage: 1TB M.2 NVMe (boot) + 24TB HDD (4×6TB)

Operating profile

Blackthorn spends most of its time in suspend-to-RAM. Woken via Wake-on-LAN from Hoid when work is queued. Completes the job, suspends.

This is ideal for Quorum’s phased workflow:

1. Wake Blackthorn
2. Phase 1: fast sync kickoff (all 13 seats, small models, parallel, ~1-2 minutes)
3. Phase 2: async deep dives (heavy models on CPU, let it cook)
4. Phase 3: final decision meeting (Tenth Man + Facilitator)
5. Suspend

Inference strategy

The 2080s (8GB VRAM each) can run two 7B models on GPU simultaneously. Everything else runs CPU inference. With 80 threads and 768GB RAM, this is viable for multiple 32B+ models concurrently.

GPU assignments (latency-sensitive seats):

• GPU 1: Seat 3 (Gut) — Qwen3.5-9B, fast instinctive response
• GPU 2: Seat 6 (Evangelist) — Mistral 7B Instruct, punchy persuasion

CPU inference (throughput-oriented):

• QwQ 32B, Cogito 14B, DeepSeek R1 14B/32B, Qwen3-30B-A3B, Llama 70B
• All fit comfortably in RAM with ~200-250GB total footprint
• 500GB+ RAM remains free for parallel model loading

Parallel batching

With 80 threads, 3-4 models can run concurrently on CPU without thrashing:

• Batch 1: Seats 1-4 (Facts & Emotion departments)
• Batch 2: Seats 5-8 (Optimism & Creativity departments)
• Batch 3: Seats 9-11 (Risk & Exploitation department)
• Sequential: Seat 12 (Tenth Man) then Seat 13 (Facilitator)

Network

• Hoid — always-on low-power device that handles WoL and job queuing
• Blackthorn is on the local network, accessible via Tailscale for remote access

Ollama

All models served via Ollama. The model registry tracks which models are pulled, their quantization levels, and their resource requirements.