Efficiency of GPU servers: which figure can you ask for in a specification?

In the specification for the servers of your expansion you wrote an energy efficiency threshold, and the supplier accepted it without objection. It is worth asking why, because the answer lies in the measurement method rather than in the price. The figure they will hand you is the one the European regulation obliges them to publish, and that figure is obtained after the accelerator cards have been pulled out of the server — precisely the components you are buying those machines for. The threshold you wrote does not touch the machine that will arrive in the room.

The figure the manufacturer is obliged to publish

Regulation (EU) 2019/424 of 15 March 2019 lays down ecodesign requirements for servers and online data storage products. It applies directly in every Member State; the consolidated text available is the one amended by Regulation (EU) 2021/341, in force from 1 March 2021. Outside the scope are, among others, servers with more than four processor sockets, server appliances and large servers.

From 1 March 2020, Annex II, point 3.1, obliges the manufacturer to publish — in the instruction manuals and on free-access websites, for at least eight years after the last unit of that model is placed on the market — PSU efficiency at 10, 20, 50 and 100 per cent of rated output power, power factor, idle state power, maximum power, declared operating condition class and, under letter (m), “the active state efficiency and the performance in active state of the server”. If the model belongs to a family, items (e) to (m) must be given twice: low-end and high-end performance configuration. That is where the figure quoted in offers comes from, and it is the only efficiency value the regulation requires a manufacturer to declare.

The line in Annex III that empties it

Annex III, point 2, sets out how the test is run: “For servers with expansion APA, the unit under test shall be tested with the expansion APA removed”, when measuring idle state power, active state efficiency and server performance in active state. APA stands for auxiliary processing accelerator: GPUs, FPGAs and the cards that carry them. If the accelerator relies on a separate PCIe switch for communication with the CPU, those cards and their risers come out too, for active state and idle state testing of all configurations.

The thresholds themselves sit in Annex II, point 2, and apply only to servers with one or two sockets: idle state power must not exceed a calculated value, and active state efficiency must not fall below 9.0 at one socket, 9.5 at two, 8.0 for blade and multi-node. Both apply “with the exception of resilient servers, HPC servers and servers with integrated APA”.

Now put the two together. A two-socket server with eight accelerators in expansion slots is subject to the thresholds — it has no integrated APA — but it meets them with the accelerators unplugged. One with an accelerator integrated into the motherboard is outside the thresholds and inside the disclosure obligation. In neither case does the published figure describe the machine that will go into service.

The same symbol, two different formulae

The regulation defines active state efficiency as Eff_server = exp[W_CPU × ln(Eff_CPU) + W_Memory × ln(Eff_Memory) + W_Storage × ln(Eff_Storage)], with weights 0.65, 0.30 and 0.05 and worklets called Compress, LU, SOR, Crypto, Sort, SHA256, Hybrid SSJ, Flood3, Capacity3, Sequential and Random: general-purpose server loads.

Recommendation ITU-T L.1311 (07/2025), Energy efficiency measurement methodology and metrics for heterogeneous servers, edition 1.0, approved on 29 July 2025 and freely downloadable, uses the same symbol and a formula of the same shape, but with only two terms: training and inference, weighted 0.5 and 0.5, over six artificial intelligence worklets: image classification, image segmentation and natural language processing.

One detail visible only if you open the document: clause 7.2.2 still describes those weights as “Wtraining is the weighting assigned to the CPU worklets = 0.5.”, a residue of the text written for general servers. The operative values remain 0.5 and 0.5, but the sentence shows how recent the transplant is.

And the cards? Table 1 prescribes the opposite of the regulation: “The EUT shall be tested with extra expansion APAs or add-in cards removed, except for the tested APAs” — the spare accelerator cards come out, the ones under test stay in.

What you can write into a specification today

The declared category: CPU+GPU, CPU+FPGA or CPU+ASIC, in blade, rack-mounted or pedestal form. “Products of categories not listed shall not be compared using the EE metric of this Recommendation.”
The configuration: as shipped, BIOS included, at least two hard disk or solid state drives, all memory channels populated alike; 25 ± 5 °C, relative humidity between 15 and 80 per cent, temperature sensor accurate to ±0.5 °C no more than 50 mm upwind of the air inlet.
The measurement point: between the AC power source and the unit under test, with no uninterruptible power supply in between; for blades, at the input of the half-populated chassis, while a multi-node system is measured with the chassis fully populated.
The instrument: an analyser with a valid calibration certificate, measurement accuracy above 99 per cent, one reading per second, a crest factor of at least 3, and a resolution of 0.01 W below 10 W.
The tolerances, written in advance: “It is accepted that the measurement accuracy of the metrics is ±10%.”, with re-test deviations accepted within 15 per cent of the passing level.
The report: twelve items of documentation, from the revision of the test suite to the extremes of inlet air temperature, including the calibration date of every instrument used.

29 July 2026, and why we do not write about its content

On 29 July 2026, exactly one year later, ITU-T approved five further Recommendations in the same series, among them L.1309 (07/26), Energy saving design for data centres considering harmonization of cloud and edge, and L.1312 (07/26), Guidelines on multi-dimensional environmental metrics and management for data centres. The catalogue lists them as “In force (prepublished)”: in force, with the text still restricted to TIES users. Whoever writes a specification today cannot quote their content: the citable document, and the one that concerns the machine you are buying, remains L.1311.

What we do not know

Of the five Recommendations approved on 29 July 2026 we report only number, title, edition, date and status from the catalogue: the text is not accessible without a TIES account, and we have not read it. We do not know which software tool will be used for the tests: L.1311 sets out its requirements and output files without naming one, and two reports are comparable only at the same suite revision. Article 8 of the regulation called for an evaluation by March 2022: the consolidated text available remains the one in force since 1 March 2021, and we do not know whether a revision will take up a method for heterogeneous servers.

The point

Declared efficiency is not a property of the server: it is the result of a method, and as long as the method remains the 2019 one, the figure describes a machine without accelerators. In a specification the gap is closed with three lines — which metric, by which method, on which configuration — repeated at acceptance, with the test report delivered together with the machines rather than promised afterwards. As with the environmental footprint of artificial intelligence systems, the measurement boundary is written first.

For a data centre we do it on both sides. The clause goes into the specification and the check is made at acceptance, with our own measurements and documentation delivered together with the supply. And with CSIDIA, the other company in the group, consumption per cabinet, test reports and in-service power curves become a single map of the room, on which an AI flags the gap between the figure declared at purchase and the real one under load — inside the client perimeter, on-premises or in a dedicated cloud with data centres in Italy. Our own teams, no subcontracting.

Do you have to write an efficiency threshold for servers with accelerators, or check one already declared by a supplier? Talk to an engineer: the site visit is free of charge, and the measurement method is written before the order, not after delivery.