Microprocessor based loads for computer hardware need high fidelity voltage for continuous fault-free operation. For more than 20 years that power quality profile has been mandated by the CBEMA curve (Computer Business Equipment Manufacturers Association) that was embodied in IEEE466 and the Emerald edition of the IEEE Colour Series Books. Within that North American centric standard the “ride-through” time for zero-volts as well as the time-based immunity for over/under voltages showed the required level of power quality although it excluded the sensitivity to frequency and rate-of-change of frequency. Although only ever issued for 120V-60Hz the pre-1997 version indicated that zero-volts could only be tolerated for half a cycle and post-1997 this was extended to one full cycle. In 50Hz Europe (20ms per cycle) this was always interpreted as 10ms and, post-1997, 20ms.
This level of immunity is always exceeded by the transients experienced on a utility network. Whilst actual “blackouts” are still relatively seldom occurrences in mature grids (for example in the UK, where urban locations may experience 2-3 seconds of power failure every 2-3 years) transients, caused by consumer connections, weather, infrastructure works, faults and network switching events, result in a “sub-20ms event” every 250h or so. In other words, if you connect your ICT load directly to the utility you should expect a high risk of mis-operation every 10 days or so. Of course in areas fed by an immature or weak infrastructure this MTBF (Mean Time Between Failure) can be as little as 2-3h.
Since the earliest days of the mainframe (after the demise of 441Hz power supplies) UPS has become the only way to protect the critical load from the vagaries of the utility power system. It is worthy of note that the specification on the UPS output power quality was, and to a large extent still is, based upon the mainframe requirements rather than the CBEMA ICT PQ curve – for example the typical tolerance on the output voltage is ±1% even though the CBEMA specification was ±5%. On the other hand the frequency stability of the typical mature utility is very high, usually less than ±0.3% – compared to the traditional view that the limit for ICT hardware should be ±1% with a slew-rate of less than 1Hz/s albeit not specified by the CBEMA curve.
However many pressures are building up on the application of UPS in the data-centre. These pressures include both operational and environmental issues:
- Efficiency – both on power cost grounds and carbon footprint
- Partial load efficiency – as most facilities run UPS at fractions of their rated load, especially so when high levels of redundancy
- In locations where PUE (Power Usage Effectiveness) is used as an improvement metric for carbon taxation the contribution to the improvement in PUE is increasing as the cooling systems become less energy intensive (the UPS is becoming the ‘new’ low-hanging-fruit)
The reaction from the UPS industry has, particularly in Europe, been quite dramatic. Full-load efficiencies have risen from a 1990s industry average of 88% to over 96% and, crucially, partial-load efficiencies have improved even more – with 25% load operating efficiency improving from c60% to over 90%. This has been largely achieved by substituting IGBTs (Insulated Gate Bipolar Transistors) for Thyristors in both rectifiers and inverters, avoidance of input and output filters and removing power & isolating transformers in the input, bypass and output. The result has also been smaller footprint, lower weight, cleaner input power demand and higher quality output voltage.
However the full-load efficiency is hardly relevant to the vast majority of data-centres. A newly constructed data-centre starts life with <10% load and, history has shown us, rarely exceeds 75% even after several years of operation. Although scalable (not necessarily ‘modular’) UPS topology is used to combat the problem of partial load, due to modular redundancy (N+1) or system redundancy (2N) an individual UPS module can spend the majority of its 10-12 year service life at well below 20%. In extreme cases (i.e. where 2(N+1) systems are applied in large Financial sector facilities) the UPS normal running load can c5%. The problem is that no large UPS system can be optimised for energy efficiency below 20% load without extreme scalability leading to high-component count and compromised reliability.
There are three basic topologies for UPS:
- Off-line, where the load is supplied by “raw” utility and a fast static-switch is used to transfer the load to a battery-fed inverter in the event of a detected utility excursion. The efficiency is very high, typically 98-99%. This topology is popular for very small UPS (c1kW) where cost is paramount and risks are acceptable. If the input power remains sub-standard the UPS will drain its battery and then shut the load off. The system cannot correct frequency and feeds the load with the input frequency. Operation on emergency genset can be problematical unless the frequency of the supply is very stable. Transient Voltage Surge Suppression (TVSS) must be fitted to protect the load from short-duration transients such as utility switching events. Almost no data-centres utilise off-line UPS since there are no large capacity modules built.
- Line-interactive, where the inverter/output stage interacts with the utility and corrects the voltage level (buck-boost). A line choke (inductor) is often used and, just as in off-line, a static switch transfers the load to the inverter only if the utility deviates. Efficiency is high with many designs reaching 97-98% and very high partial load efficiency. All diesel-rotary UPS (DRUPS) and most hybrid-rotary UPS are line-interactive and thousands of large facilities are fed from line-interactive topology UPS. Line-interactive UPS are just that, line-interactive, not on-line. The definition of ‘on-line’ should include the requirement of the inverter handling the entire load, all of the time and not switching between ‘interacting’ (5-10% load handling) and on-line. Frequency variations on the input cause the module to switch to battery operation.
- On-line (also referred to double-conversion) decouples the input from the output and, continuously, will control both voltage and frequency on the output – working from very poor quality utility input power. The only downside is that this high level of decoupling comes at a loss of energy and the full load efficiency ‘ceiling’ is c96% – although this itself is a huge technical step forward. The ability to run with high-fidelity output and low quality input has made this topology the mainstream choice for data-centres worldwide.
So, how to break through the glass ceiling of on-line UPS? An answer (if not the answer) lies in an “eco-mode” feature.
The eco-mode product principle is simple:
- The power quality of mature grids has an MTBF of c250h and an MDT (Mean Down Time) of c3s. The “down time” in this case is not a “power failure” per se but a deviation from the limits of the CBEMA curve. This means that for 99.99967% of the time the utility is of sufficient quality so as not to need UPS protection.
- The static switch that is in the output of all static UPS is capable of transferring the load between the inverter and the utility and the inverter (and vice versa) in 4ms – must faster than the 20ms time limit required by the CBEMA curve.
- The usual IGBT based inverter can accept a 100% step load with sub-5ms transient response.
- Therefore, with the addition of a power quality monitoring function on the input, the UPS can run in ‘eco-mode’ for the majority of the time, effectively off-line, and switch to the inverter mode when required.
In most western European countries it is anticipated that eco-mode will be active for more than 96% of the running period, allowing for monthly diesel generator testing.
Eco-mode can be applied to both line-interactive products (with an efficiency improvement of 1% plus any gains from cooling the UPS losses, c1.25% overall) or on-line products (with a 3% gain plus any cooling savings, at least 3.5% overall). Clearly the application of eco-mode to on-line products gives the greatest security of power supply and flexibility when running on emergency diesel generators etc. One product on the general market can run in all three modes depending upon the utility quality over the prior 30 minutes or so.
In eco-mode the UPS controls and switching components are all enabled and active, just not conducting (apart from battery re-charging) and so there is a load-independent standing-loss virtually equivalent to the eco-mode full-load losses. In this respect the larger the UPS module size the lower will be the standing losses and so the scalability of the power solution should be carefully chosen.
In early generations of eco-mode enabled products the transfer time was c4ms and paralleling of modules was problematical but the latest developments transfer in 2ms and can run in parallel for power or redundancy at will. All systems can be inhibited in switching into eco-mode and most are adjustable for the period of acceptable quality before the eco-mode automatically enables.
With eco-mode there is always the recommendation that all-mode TVSS is fitted in the eco-mode power-path – as it should be on the bypass of all UPS systems.
Further power savings (e.g. up to 99.5% efficiency) can be envisaged where the load is configured as dual-cord. In this architecture (i.e. Tier III) there is an active-path and a passive-path which can be configured to be supplied by UPS on the active path and directly by the utility on the passive path. In this way the load on the UPS is halved and the overall losses potentially reduced – but only if the partial-load efficiency does not fall away too fast as the load reduces.
Of course, the data-center industry is very conservative and resistant to change so where power is low cost and the data-center PUE is >1.5, the impact of eco-mode would be low and rarely applied, but, as the cost of power increases (for example in the UK currently GB£0.08/kWh and forecast to rise 15-20% per year for the next 5 years) the pressure to adopt eco-mode operation will grow and application increase rapidly. PUE driven “cloud” services (where power costs represent >50% in the 10 year TCO and the client relies on SLAs rather than facility approval) will be in the application vanguard of eco-mode.
Ian Bitterlin is CTO for Ark Continuity – a developer of high integrity, low carbon, data-centre’s based in Corsham, Wiltshire. With a strong real-estate portfolio, well over 100MVA of power and planning consent for >100,000m² of critical space in multiple UK locations, Ark are at the forefront of the low-energy data-center market.