Forge F1

F1 is our first forge, in a data hall on the Reykjanes peninsula, Iceland. It was listed on 2026-07-28 after its burn in. Region id is-south, forge id f1.

#Each node

#Between the nodes

Each GPU has its own 400 Gb/s InfiniBand NDR port, 3.2 Tb/s per node, on one non blocking switch. A 16 GPU job gets GPUDirect RDMA between the two nodes. The same fabric reaches the shared storage cluster.

#Cooling and power

GPUs and CPUs are cooled by liquid on cold plates. The loop rejects heat through dry coolers to outside air, which averages 5 °C over a year; on the warmest days a small chiller helps. Each node draws up to 14.3 kW, on A and B feeds from separate breakers. Measured PUE over the first 60 days was 1.15.

Under sustained load our B200s sit between 52 °C and 72 °C. You can watch every one of them on the status page.

#Burn in

Before a node is listed it goes through the same steps, in order:

1. Vendor acceptance and serial check.
2. Firmware and driver pinned to the versions every node runs.
3. 96 hours at 95 % load or more on every GPU, with NVLink, InfiniBand and storage tests each hour.
4. Heat agent reporting to the public feed.
5. Listed in GET /forges.

Any fault in step 3 restarts the clock after the fix. F1 passed at 97.4 % mean load.

#Software

Ubuntu 24.04 LTS, NVIDIA driver 580, CUDA 13.0, NVIDIA container runtime. Host updates happen in maintenance windows announced 7 days ahead in the incident log.

#Forge F2

F2 is the same build at Beauharnois, Québec, next to a hydro dam on the St. Lawrence. It is listed when it passes burn in. The ledger reports its progress each month.