(Input Protection and Node Autonomy)
The stability of the external power supply is a critical factor for the survival of any DePIN node. At peak load, our system consumes between 1.8 and 2.4 kW of power. Public utility power grids—whether residential or semi-industrial—are prone to sudden voltage sags, high-frequency interference, and brief outages.
For a regular computer, this would only result in a reboot, but for a server node running on an AI network (such as io.net or Render Network), even a microsecond-long outage means an instant drop in SLA (uptime), penalties (slashing), and a complete loss of ranking in the distributed computing pool. The goal of this stage is to isolate the server from external network failures and create a reliable backup power circuit.
1. The Pitfall of Cheap UPSs: Why Computer UPSs Are a No-Go
The first critical mistake when building a high-performance server is trying to buy a standard “computer” UPS (Line-Interactive UPS). For our architecture, this is a surefire way to cause equipment failure:
Stepped Sine Wave (Modified/Approximated Sine Wave): When switching to battery power, budget UPS units output an approximated sine wave (meander). Industrial server power supplies (HP/Delta “skis”) with active power factor correction (APFC) are physically incapable of operating at such a current—they either immediately trigger a protection circuit or their input circuits burn out due to overheating.
Switchover time: Line-interactive UPSs take 10 to 20 milliseconds to switch from the mains to the battery. For a heavy-duty server running at 100% GPU load, this “downtime” is long enough for the power supplies to detect a voltage drop and trigger a system reboot.
2. Professional Solution: 3 kW Inverter Systems
To ensure maximum reliability, the node’s power supply system is built around high-quality inverters from industry leaders (Victron or Deye). As a matter of principle, we focus on the basic and most affordable 3 kW model. This is more than enough to handle our peak load of 2.4 kW, and it prevents the assembly budget from being unnecessarily inflated.
Critical engineering requirement for switching speed:
Residential Systems (Delayed-Action Mine): Standard UPS units and inexpensive inverters have a switchover time of 10 to 20 ms. For a heavily loaded server, this is too long—the power supplies will detect a power shortage, and the node will reboot.
Victron vs Deye
A server with four powerful GPUs is neither a refrigerator nor a home computer. Any power dip or switch to battery power poses a risk of the server crashing or data corruption. For our “DIY build” (power consumption ~2.6 kW), choosing the right power supply is critical.
1. Why Victron (3 kW) Is Our Top Choice
For a single “unit,” Victron is the ideal standard.
Switchover speed: This is the main advantage. Victron switches to battery power in less than 4 milliseconds (sometimes even faster). For a server power supply, this happens so quickly that it doesn’t even have time to “blink” and reboot.
Comparison with household UPSs: Standard household UPSs have a switchover delay of 10–20 milliseconds. For server hardware, this is enough to cause the system to crash. Residential UPS units are designed to provide a “safety net” for household appliances, not to maintain the stability of high-load chips.
Scalability: Victron’s 3 kW model series fits perfectly into one of our setups. You don’t overpay for excess power, while still getting professional-grade power management.
2. When should you switch to Deye?
Deye is an excellent solution, but on a different scale.
The difference: Deye primarily covers the 8–12 kW segment and above.
Economies of scale: Buying a 12-kilowatt Deye for a single machine is like “using a sledgehammer to crack a nut” and results in unnecessary costs. However, as soon as your farm grows to four “units” (total consumption approaching 10–12 kW), the situation changes.
Growth Strategy:
Start (1 vehicle): Install 1x Victron 3 kW. This is reliable, fast, and budget-friendly.
Expansion (4 vehicles): If you decide to build a cluster right away, a single powerful 12 kW Deye will be a more economical and convenient solution than a set of four Victrons.
Conclusion for your technical setup:
Don’t try to skimp on the “brains” of your power system. Switching delays are your worst enemy.
Looking for a single unit? Go with the Victron 3 kW—it’s an investment in 4 milliseconds of safety that will protect your computing systems from sudden failures.
Planning a cluster of 4 or more machines? Consider the Deye 12 kW to consolidate power for your entire infrastructure into a single, reliable unit.
Technical note: Never use cheap offline UPS units. They produce an “approximated sine wave,” which can cause server power supplies to emit an unpleasant squealing noise or simply fail under load. Only a pure sine wave from a high-quality Victron or Deye inverter.
Professional Standard (Our Choice): In Victron-class equipment, the switchover time to battery power is less than 4–5 milliseconds. This is a critically important threshold. The relay operates so quickly that server power supplies with active power factor correction (APFC) are physically unable to detect the voltage dip and continue to operate stably in normal computing mode.
3. Buffer Battery Array: Choice and Cost-Effectiveness
The primary task of the battery backup in our scenario is to keep the server running for 2 hours (enough time to resolve a local failure or safely complete current computing tasks). Since such outages occur not several times a day, but on average once a month, the system integrator has every right to choose between the two technologies based on their budget.
Option A: Gel Batteries (GEL / AGM) — An Affordable Option
- Pros: They are significantly cheaper than their lithium counterparts, widely available, and widely used in off-grid power systems.
- Rationale: Since the battery operates exclusively in a low-discharge-cycle mode (infrequent discharge cycles), the lithium’s enormous cycle life simply will not be exhausted in this scenario. Gel batteries will perform their task perfectly, saving a significant portion of the initial investment.
- Cons: They are heavier and larger, and take longer to fully charge after being deeply discharged.
Option B: Lithium iron phosphate (LiFePO4) battery packs — Uncompromising premium quality
- Pros: They have a virtually unlimited service life (up to 4,000–6,000 discharge cycles), can handle extremely high currents (charge instantly), are compact, and feature a built-in battery management system (BMS) for cell balancing.
- Cons: High initial cost. Recommended for purchase if you have an unlimited budget for the setup and plan to expand the system further until it is fully self-sufficient (for example, by connecting solar panels).
Example of an assembly using 12V 200Ah batteries:
- One such battery contains 2.4 kWh of energy (12 V × 200 Ah).
- A string of 4 batteries: this gives us 48 V and 200 Ah (total capacity of 9.6 kWh).
Result: This capacity is more than enough to keep the server running for 2.5–3 Practical example: Calculating the buffer for 2 hours of autonomy (3 kW)
To ensure 2 hours of operation for our 3-kilowatt server, we need to store 6 kWh of usable energy (3 kW × 2 hours). Taking into account the inverter’s efficiency and the depth of discharge, we should aim for a total battery capacity of about 7–8 kWh.
48V Base Configuration:
All 12V batteries (GEL or LiFePO4) are connected in series in groups of four to produce 48V. These strings are then connected in parallel to power the clock (with a margin to account for battery degradation and high efficiency).
Integration of the Victron buffer power supply system with Linux for automatic protection of node ratings
Why automate the communication between the inverter and the server?
An off-grid power system based on Victron Energy inverters (MultiPlus-II / Quattro), combined with a lithium iron phosphate (LiFePO4) buffer, provides the server with up to 2 hours of uninterrupted autonomous operation during a complete power outage.
However, simply “keeping the battery running” is only half the engineering challenge. If a line failure drags on, the server must respond intelligently. In distributed computing networks (Render Network, io.net, Akash Network), severe penalties are imposed for a node suddenly going offline (Hard Offline):
- The internal reliability rating (Uptime/SLA Score) is decreasing.
- Noda is excluded from the priority distribution of computationally intensive tasks (rendering or AI training tasks).
- In the worst-case scenario, the accumulated balance is frozen or reduced.
Purpose of integration: To establish a direct connection between the Victron “head” (the Cerbo GX control unit or the built-in GX module) and the Ubuntu/Linux server operating system. If the external network goes down, the server will detect this, start a safety timer, after 15–20 minutes, it will correctly notify DePIN networks of the transition to maintenance mode, redirect tasks, and, if necessary, shut down normally while maintaining 100% of its rating.
Instructions for configuring the connection between Victron and the server can be found in our STEP 9 section on the website.
4. Input Filtering and Through-Ground Grounding
A protective panel must be installed upstream of the inverter at the power line input; this panel absorbs the impact of any faults in the external power grid:
- Voltage stabilizer: Smooths out sharp drops and spikes in the external power supply, ensuring the inverter operates within a safe range.
- SPU (Surge Protection Unit): It effectively cuts off microsecond-long high-voltage surges caused by thunderstorms or the switching of heavy equipment on the line.
- Grounding Circuit: The chassis of our 4U server, the metal parts of the racks, and the power supply unit circuit boards must be physically connected to a high-quality grounding circuit. In DePIN nodes, where four powerful graphics cards are in operation, electromagnetic interference and static charge on an ungrounded chassis can cause micro-glitches on the PCIe bus and sudden GPU crashes in the operating system.