NT
inference
Thermaltake W200 · dual loop
Core
CPU
AMD Ryzen Threadripper PRO 9985WXmeasured
64 cores / 128 threads
Motherboard
ASUS Pro WS WRX90E-SAGE SE
Memory
8×96 GB DDR5-6400 — 768 GB total
Graphics
GPUs
4×NVIDIA RTX PRO 6000 Blackwell Workstation Editionmeasured
96 GB GDDR7 each — 384 GB of VRAM in one chassis
Storage
Boot
Micron 7450 Pro 480 GB NVMe
Model store
2×Kingston DC3000ME 7.68 TB U.2 NVMemeasured
9.3 GB/s sequential read, O_DIRECT
Array
4×Micron 7450 Pro 3.84 TB U.3 NVMe
Cooling — dual loop
Distribution
Singularity Computers Axiom Dual Loop Distribution Plate
Two independent loops from one plate
Pump
Singularity Computers
Radiators
Alphacool
360 × 86 mm · 560 × 86 mm · 2× 560 × 60 mm
Blocks
Optimus — CPU and GPU
Network & power
NICs
NVIDIA ConnectX-6
2× 100 GbE direct-attached to EPYC, MTU 9000 — no switch between hostsmeasured
EPYC
support & voice
air cooled
Core
CPU
AMD EPYC 7763measured
64 cores / 128 threads
Motherboard
Gigabyte MZ32-AR0
Memory
16×64 GB DDR4-3200 — 1 TB total
Graphics
GPU 0
NVIDIA RTX PRO 4500 Blackwell Server Editionmeasured
32 GB — the fallback language model, so the assistant survives a mode switch on NT
GPU 1
NVIDIA GeForce RTX 3090measured
24 GB — speech recognition and synthesis for the satellites
Storage, network & power
Array
4×Micron 7450 Pro 3.84 TB U.3 NVMe
NICs
NVIDIA ConnectX-6 — 2× 100 GbE to NTmeasured
PSU
Seasonic Prime TX-1600 Noctua Edition
Fans
Noctua 120 mm (current generation)