Objective: To examine the interaction between CD4 cell count, viral load suppression and duration of ART on mortality. Methods: Cohort analysis of HIV-infected patients initiating ART between April 2004 and June 2011 at a large public-sector clinic in Johannesburg, South Africa. A log-linear model with Poisson distribution was used to estimate risk of death as a function of the interaction between current CD4 count, current viral load suppression and duration on ART in 12-month intervals. We calculated predicted mortality using estimated coefficients within combinations of predictors. Results: Among 14,932 ART patients, 1,985 (13.3%) died. Current CD4 was the strongest predictor of death (<50 vs. ≥550 cells/mm3 – RR: 46.3; 95%CI: 26.8–80), while unsuppressed current viral load vs. suppressed (RR: 1.8; 95%CI: 1.5–2.1) and short duration of ART (0–11.9 vs. 66–71.9 months RR: 1.7; 95%CI: 1.2–2.3) also predicted death. Our interaction model showed that mortality was highest in the first 12-months on treatment across all CD4 and viral load strata. As current CD4 and duration on ART increased and viral load suppression occurred, mortality dropped. CD4 count was the strongest predictor of death. The relative effect of current CD4 count varied strongly by viral load and duration of ART (from 1.3 to 55). Lack of suppression increased the risk of mortality upwards of 6-fold depending on time on ART and current CD4. Conclusions: Our findings show that while CD4 count is the strongest predictor of death, the effect is modified by viral load and the duration of ART. Assessment of risk should take into account all three factors.