Built for one person’s actual workload.

Specified from the training job backwards: memory first, then GPU bandwidth, then everything else. Quiet under sustained load, clean enough to service in ten minutes.

2×

GPUs, full x16 lanes

128 GB

ECC memory

34 dBA

at 1 m, full load

24 h

burn-in before handover

Thermal validation · 60 min sustained load

34 dBA at 1 m

cpu · plateau 76° gpu · plateau 71°

Included with every build

Assembly and cabling plan, photographed
24-hour burn-in under a real training job
Thermal and acoustic report
Linux + CUDA + drivers configured
Documentation to service it yourself
Two-year build warranty, parts by manufacturer