The LedgerHow It RunsAbout UsHardwareThe LabRun Logs YouTube @localhost_live ↗ Request a Free Build →
Hardware matrix · recorded, not scored✦Printed exactly as observed — no vibes

the iron

Runs on hardware we actually own.

One DGX Spark, one laptop, a second Spark or two. Every build's badge prints its machine exactly as observed.

MachineSoC / GPUMemoryCPURole on the ledger
★
gx10-a6b9Asus GX10 — “DGX Spark”
NVIDIA GB10
121 GB unified LPDDR5x
20-core aarch64 · CUDA 13.0
THE THRONE
2×
dual GX10 cluster2× Asus GX10
2× GB10
256 GB combined
ran the GLM 5.3 3-bit one-shot
≈
Mac Studio M3 Ultra96 GB unified
M3 Ultra
96 GB unified
ran the Qwen 3.8 27B overnight build

Observed live on the gx10-a6b9 throne: NVIDIA GB10, driver 580.173.02, CUDA 13.0, 20 CPU threads, 121 GB usable unified memory.

interactive tooling · cost to build

Will it fit? VRAM & cost-to-build calculator.

Move the sliders like a stream operator planning a slot: parameter count, quantisation bits, and KV-cache overhead.

Model: weights GB ≈ B × 10⁹ × bits ÷ 8. Machine baselines: GB10 121 GB @ US$1,999 · dual GX10 256 GB · Mac Studio 96 GB+. Estimates, not gospel — the operators' clocks are the law.

topology · the arena

How the slots are wired.