2x RTX Pro 6000 build maxes PSU
A Reddit user documented an extended stress test of a dual-RTX PRO 6000 Blackwell workstation built for local LLM serving. The rig draws roughly 1650W at the wall, forcing the GPUs to stay around 535W each, while the air-cooled Threadripper CPU sits near 95C without throttling.
This is a power-delivery benchmark disguised as a thermal test: the machine’s ceiling is the PSU and circuit, not the cooling stack. That’s the real signal for anyone building serious on-prem inference rigs.
- –The setup pairs two 96GB RTX PRO 6000 Blackwell cards with a Threadripper Pro 7965WX, 256GB ECC RAM, and a 1600W Titanium PSU.
- –NVIDIA rates the cards at up to 600W, but the builder capped them lower to stay under the PSU’s hard limit during full CPU plus GPU load.
- –The air-cooled Noctua TR5 cooler appears to be doing its job: the CPU holds around 95C under combined heat soak without thermal throttling.
- –For local AI workloads, this says more about systems engineering than raw GPU throughput: airflow, board behavior, and power headroom now matter as much as the silicon.
- –It’s a useful real-world datapoint for inference teams sizing racks, circuits, and cooling before they ever think about model latency.
DISCOVERED
45d ago
2026-04-26
PUBLISHED
45d ago
2026-04-25
RELEVANCE
AUTHOR
Signal_Ad657