I've replaced three processors due to electromigration failure in the last six months. So, these chips have had a fairly hard life, but probably an easier life than if they belonged to a consumer who folds or cryptomines. The remaining third of their power-on time is idle and they are running 24/7/365. These CPUs are cooled well and perform 4C/8T software rendering at the stock 4GHz for about two thirds of their power-on time. Giving a very small sample size, but one that is still useful statistically, we added 24 nodes (3 stacks of 8 nodes) to our renderfarm when the 4790K came out. It's the old voltage-squared problem, and actually makes underclocking more enticing How low can you get the voltage without dropping clockspeed? The reductions in noise and heat are often quite staggering (especially if we're talking about AMD GPUs). I moved away from extreme overclocks, water loops and loud fans a long time ago I firmly believe that the best point to stop overclocking is the point at which you have to add more voltage. I've always assumed that electromigration is roughly proportional to the power consumption - so if a stock CPU uses 84-91W and an overclocked chip uses, say ~160W, then it's using 80% more power and will burn through its lifespan 80% quicker - so a 14nm finFET chip that we'd expect to last around 5 years in a 24/7 situation may last only 2-3 years when overclocked. I certainly won't be recommending used processors any more. I'm starting to see multiple degraded Ivy and Haswell chips (22nm) that spent their half decade at stock speeds, whilst degraded Sandy chips (32nm) seem to be very rare outside of extreme overclocking. It's definitely getting worse with these newer processes. Increased voltage will hasten the electromigration, but that'll happen in time anyway, even at stock speeds.
0 kommentar(er)
