Hardware — SolarAI LLM Server V2

Full stack

Component	Part	Role
Motherboard	Asus Z9PE-D16	Dual LGA2011, 40 PCIe lanes — the whole reason four P100s fit and stay fed
CPU	2x Intel Xeon E5-2620 v2	12 cores / 24 threads total; handles tokenization, sampling, OS, and the serving layer
GPU	4x NVIDIA Tesla P100 16GB PCIe	64GB HBM2 VRAM combined; Pascal, 16-bit compute, the inference engines
RAM	128GB DDR3	Model staging, OS cache, headroom for KV spill and multi-request serving
Storage (OS)	120GB SSD	Ubuntu 24.04.5 server, CUDA, the serving stack
Storage (models)	4x WD Velociraptor 500GB	10K RPM HDDs holding Strata QWEN3.8-120b and its quant variants

Why these parts

Z9PE-D16: a server board sold into a niche that died. Cheap now, and one of the few platforms where four double-width PCIe cards coexist without bifurcation hacks.
Tesla P100: 16GB of HBM2 each at ~700 GB/s memory bandwidth. Four of them is 64GB of VRAM — enough to host a 120B model in 4-bit quant with room for context. The secondary market priced these as e-waste.
DDR3 at 128GB: the rig streams weights from disk through RAM into VRAM. Bandwidth of the previous generation, doing the job of the current one.
Velociraptors: 10K RPM, 500GB each. Model files are sequential reads — HDDs are honest about that. The models load once and live in VRAM; the disks just have to survive the boot.
Xeon E5-2620 v2: not fast by modern standards. Inference doesn't care. The GPUs do the math; the CPUs move bytes and answer the network.

Power and thermals

Four P100s at ~250W each plus dual Xeons puts the wall draw in the 1.2–1.4kW range under full inference load — and every watt of it is solar. Air-cooled blowers, a case with more airflow than furniture, and 85kWh of battery behind the outlet: the rig runs full-load for roughly 60 hours with zero sun. The board takes the RAM the server world stopped making and the GPUs the datacenters stopped wanting — the entire bill landed near a grand.

64GB VRAM / ~$1,000

$16 per gigabyte of VRAM. That is the entire argument for this machine.