Engineering Notes on Server Power
Short, numbers-first write-ups for hardware engineers and sourcing teams: what titanium efficiency is actually worth, how CRPS differs from M-CRPS, and why AI racks broke the old power playbook.
Titanium vs Gold: the Efficiency Math Buyers Skip
92% to 96% at 50% load sounds incremental. Across a 10kW node it is roughly 430W of heat every hour — enough to change your cooling bill and your rack density ceiling.
Read article →Titanium Efficiency as a TCO Argument
How to build the payback model a finance team will accept: efficiency points into watts, heat costed at your PUE, thermal headroom valued, and a sensitivity table around it.
Read article →CRPS vs M-CRPS Explained
Same 73.5×40×185 mm starting point, two different eras. Where classic CRPS tops out near 2400W, the modular M-CRPS segment scales to 5.5kW and 54V — here is how to choose.
Read article →The CRPS Mechanical Envelope, Dimension by Dimension
The CRPS-185 envelope explained: what 73.5×40×185 mm constrains, sled and cage tolerances, the golden-finger pinout, airflow direction and the format's practical power ceiling.
Read article →M-CRPS for ORv3: What Actually Changed
Classic CRPS versus M-CRPS item by item, the 3000W–5500W power classes, the 54V output section, and the mechanical and thermal differences a shelf forces on your design.
Read article →54 V vs 12 V Rack Distribution
Busbar sizing worked out in copper mass, where the conversion stages move to, what ORv3 standardized, and an honest itemisation of what a bus-voltage migration costs.
Read article →GPU Rack Power Trends: from 5kW to 120kW
H100 draws 700W, B200 draws 1000W, and racks have gone from 5–10kW to 40–120kW in a hardware generation. What 54V busbars and 33kW power shelves mean for PSU sourcing.
Read article →PMBus on Server PSUs: a Working Guide
Which registers are worth polling, how often to poll them, why warning thresholds matter more than fault thresholds, and how to wire PSU telemetry into rack management.
Read article →N+N and N+1 Sizing Without Guesswork
Units per shelf, post-failure load per module, derating at temperature and altitude, and two worked failure cases — one that holds the node and one that cascades.
Read article →Burn-In and ATE: How to Read a Test Record
What ATE proves versus what burn-in screens, the chamber profile variables to freeze, what a per-serial record must contain, and how to audit a supplier's line.
Read article →Written for the Buying Table
These are the conversations our engineers have most often with OEM and ODM sourcing teams: whether the titanium tier pays back (it usually does, and the payback is calculable), which form factor a platform should standardize on as power levels climb past 2.5kW, how the 54V rack architecture changes PSU selection for AI nodes, and what the production evidence behind a power supply is actually worth. Every article states its assumptions and links the claims back to product pages — 1600W–2400W CRPS, 3kW–5.5kW M-CRPS, GPU power solutions — so you can check the numbers against what we actually ship. Wattage stages are labeled honestly throughout: 550W–2400W in production, 3kW–5.5kW at development-sample stage, 8kW+ on the roadmap. New articles are added as the roadmap moves — bookmark this page or subscribe by asking sales to add you to the engineering-notes mailing.
Reading About It or Building It?
TitanWatt titanium CRPS runs 550W–2400W in production, with 5.5kW M-CRPS samples available.
Request a Quote