Post B25xq2bZeyw4pse9L6 by azonenberg@ioc.exchange
 (DIR) More posts by azonenberg@ioc.exchange
 (DIR) Post #B25xpyJhbUGNXIFovY by azonenberg@ioc.exchange
       2026-01-09T02:19:49Z
       
       0 likes, 0 repeats
       
       Can somebody explain why MLC/TLC/QLC flash is apparently slower to write than SLC?it seems to me that if you're PAM4/PAM8/PAM16-ing the same amount of data across less bitcells you should get *higher* write speed. Why is MLC not double the write throughput of SLC?
       
 (DIR) Post #B25xpzKnosjmgzmDHU by jpm@aus.social
       2026-01-09T02:23:08Z
       
       0 likes, 0 repeats
       
       @azonenberg the slow part is erasing, not writing
       
 (DIR) Post #B25xq0ACjlr1GPpGOu by azonenberg@ioc.exchange
       2026-01-09T02:28:10Z
       
       0 likes, 0 repeats
       
       @jpm I'm not talking about erasing, I'm talking about writing. My understanding is a lot of modern SSDs have a fairly small write buffer of fast SLC bitcells and when it fills up write speeds drop as subsequent writes commit directly to the TLC/QLC array.What I don't understand is why it's not the other way around. I thought SLC was something you used when you wanted the highest longevity and reliability *at the expense of* speed and density.
       
 (DIR) Post #B25xq10fahozt8NAB6 by azonenberg@ioc.exchange
       2026-01-09T02:28:34Z
       
       0 likes, 0 repeats
       
       @jpm also why would multi level bitcells take longer to erase? You're erasing the entire page regardless of what was there.
       
 (DIR) Post #B25xq1apQGkFhHI2XQ by jpm@aus.social
       2026-01-09T02:38:36Z
       
       0 likes, 0 repeats
       
       @azonenberg each MLC page is much larger than SLC, so a read-modify-write of a block within a single page takes proportionally longer, even without considering flash erase time
       
 (DIR) Post #B25xq2Cl9F5Pav2Kf2 by azonenberg@ioc.exchange
       2026-01-09T02:39:29Z
       
       0 likes, 0 repeats
       
       @jpm yeah but don't most FTLs implement log structuring where you write until the block fills up to avoid write amplification?
       
 (DIR) Post #B25xq2bZeyw4pse9L6 by azonenberg@ioc.exchange
       2026-01-09T02:24:29Z
       
       0 likes, 0 repeats
       
       I would expect reads to have similar results with caveats: the higher BER will mean more frequent error correction operations, so depending on how slow the FEC is you might not have e.g. MLC reading at twice the speed of SLC.But I would still expect a net performance *and* density gain at the expense of BER and endurance.
       
 (DIR) Post #B25xq30OAimk4qFy1A by ignaloidas@not.acu.lt
       2026-01-09T03:01:09.310Z
       
       0 likes, 0 repeats
       
       @azonenberg@ioc.exchange @jpm@aus.social that only carries you while you still have unwritten blocks, write amplification is still very much an issue that needs dealing withTho I don't believe it's the major cause of slowness with QLC compared to SLC