C'est avec une grande tristesse que j'apprends aujourd'hui la mort de Pierre Giraud (1981-2026). Acteur et producteur, il était notamment connu pour avoir créé et incarné le personage de Joueur du Grenier entre 2009 et 2013, avant d'être remplacé par Frédéric Molas
Erasing (not writing!) is the main thing that kills SSDs.
With the industry pushing towards HBF as a DRAM / HBM alternative, endurance is the name of the game.
There’s two knobs you can adjust for NAND flash that make a huge difference:
Peak erase voltage (~20V) and the length of the pulse (~3.5ms).
As you (might) expect, modern SSDs very much err on the side of caution; risk of losing user data is too high. But what if you’re willing to live on the edge a little? Research came out of Korea recently where ordinary Samsung 3D TLC NAND lifetime was pushed +43%!
See, erasing on an SSD is kind of a multi-step process. As the cells age, the block will need multiple pulse “rounds” to erase. It usually looks like this:
Normal Erase:
3.5ms pulse -> check, too many cells fail -> 3.5ms pulse -> check
The REO research is a bit more clever. Using a basic predictive model, they a shallow erase by using smaller increments in the final pulse.
REO (the paper) Erase:
3.5ms pulse -> check, too many cells fail (but we’re close!) -> *1.5ms* small pulse -> check
You can actually be quite aggressive with this technique! I think of it like erasing a whiteboard. Sure, you can go overboard cleaning it perfectly…but you can still use it even if you only did a quick wipedown. Might be a little smudgy, but it works.
What’s awesome about this is that with these shorter pulses, you increase cell lifetime AND improve performance! p99.99 tail read latency got ~22% better.