▍ The rig · Tartu

Four Blackwells on a welded frame

No datacenter, no kit case. A custom chassis on the workshop floor, built up over two cards, then four. This is where every finding on this site comes from.

GPUs

4× RTX 6000 Pro Blackwell

VRAM

384 GB · 96 GB each

Platform

ASUS Pro WS WRX90E-SAGE SE · dual 10G

Interconnect

LINKUP PCIe 5.0 risers

Cooling

Noctua · ducted air

Chassis

Custom welded aluminium

LINKUP PCIe 5.0 riser cable boxes and two GPUs in anti-static bags with yellow security seals, on a wooden floor with a screwdriver. — 01 Risers arrived LINKUP PCIe 5.0 · the right cable

The aluminium-frame chassis with four NVIDIA Blackwell cards mounted in a row, Noctua-cooled, on a stand in a dark workshop. — 02 Coming together four cards, one frame

Front view of the rig: four cards with DisplayPort outputs, riser cables threaded down to the motherboard, ASUS Pro WS WRX90E-SAGE SE board with dual 10G ports below. — 03 Wiring it up risers threaded · dual 10G

The rig running — lit 10G ports, a Noctua fan, the cards, and pink model output streaming in a terminal behind. — 04 First inference model output, streaming

05 It hums four cards under load

▍ The fleet · 6 nodes

Not just the one box

The Blackwell rig above is the heavy iron, but it doesn't work alone. A handful of nodes carry inference, image gen, storage and the cockpit i drive everything from.

node-01 the heavy iron

4× RTX PRO 6000 Blackwell

Threadripper PRO 7965WX · 48 threads

384 GB VRAM251 GiB ECC

node-02 the cockpit

2× RTX 5080

Core Ultra 9 285K · 24 threads

32 GB VRAM64 GB RAM

node-03 satellite node #1

RTX 5090 + 5070 Ti

Core Ultra 7 265F · 20 threads

48 GB VRAM64 GB RAM

node-04 satellite node #2

2× RTX 5060 Ti

Core i9-13900K · 32 threads

32 GB VRAM32 GB RAM

+ two more nodes — GitLab and model storage · linked by a 10 Gbit direct fabric · ~443 GiB system RAM across the fleet

The story of how it got here → How this started