Computing · Server Efficiency

Server Work per Watt

More work from the same watt: 21.3× since 2009, in two long climbs and two stalls.

• 2,098 → 44,648 overall ssj_ops/watt
• 21.3× across 17 elapsed years
• Stalled 2014–2018; flat again since 2024
• Vendor-submitted record line; 2026 partial

SPECpower_ssj2008 runs a Java workload against measured wall power and reports operations per watt: the efficiency counterpart to Moore’s Law, in Koomey’s lineage, counting computing per unit of energy rather than transistors per chip. The best published one-node Server result rose from 2,098 overall ssj_ops/watt in 2009 to 44,648 so far in 2026, a 21.3× gain across 17 elapsed years. It did not get there evenly. Growth ran at 38% a year through 2014, and then the frontier stalled for four years, 10,654 to 13,478, about 6% a year. The stall broke when AMD’s EPYC took the lead from Intel in 2019, and from 2018 the frontier climbed 21% a year for six years, to 42,312. It has flattened again since: 2025 and 2026 have added 5.5% between them.

Results are voluntary, vendor-submitted and tuned for the benchmark, so this is a record line: the efficiency frontier, not the efficiency of the servers running the world. Points are dated by SPEC publication year, which runs ahead of deployment; the 2024 winner’s hardware was not available until February 2025. The 2026 point is the best published so far this year and may still rise. SPEC® and SPECpower_ssj® are registered trademarks of the Standard Performance Evaluation Corporation.

Source: SPECpower_ssj®2008 results

Download the data: specpower_efficiency.csv · specpower_efficiency.json

All curves · The chart is drawn in your browser and needs JavaScript. Every number it plots is in the table above.