Where Correction Stops Paying: Error-Correction Energy for Memory at Fleet Scale

Error-correcting code (ECC) memory costs energy three ways: extra stored bits, patrol scrubbing, and correction logic on every access. This paper prices those costs at fleet scale from published inputs (DRAM soft-error rates from the large-fleet studies of Schroeder et al. 2009, Sridharan et al. 2013/2015 and Meza et al. 2015, which span orders of magnitude; ECC overhead fractions for SECDED (64+8) and a stronger device-correcting code (128+32); DRAM access energy from Horowitz 2014) and prices what correction buys under a plainly stated energy valuation of a prevented crash (1.2e6 J: a 500 W server restarting for 600 s plus re-executing half of a 1 h checkpoint interval, with one uncorrected error in 1e4 taken to cause a crash; silent corruption is assigned no energy value). For a reference fleet of 100,000 servers with 512 GiB each (54.98 PB), SECDED draws a continuous 707759.3 W (6.204 GWh per year), 97.09% of it the background power of the extra stored bits; scrubbing draws 114.53 W and logic 450.0 W. At a representative 25000 FIT per Mbit the fleet corrects 4.2290e-3 errors per joule of ECC energy (236.46 J per correction, against 2.8836e-9 J for the correction act itself) and the avoided crash energy is 0.5075 times the energy spent. SECDED breaks even in energy at 49263.0 FIT per Mbit, at the top of the published range; the stronger code breaks even against SECDED only at 2539132.7 FIT per Mbit, far above any published fleet rate. Conclusion: ECC is not an energy-saving measure at any published error rate; it is a continuous capacity cost paid for correctness. Model calculations are labelled SIMULATED, NOT MEASURED. The deposit includes the paper PDF and markdown, the computation script (compute.py) that produces every number in the paper, and its outputs (results.json, results.txt).