Slashdot

Flash memory is no good because it has a limited number of write cycles (typically about 10,000 - after which it becomes 'random'. If a swap file was on flash memory, it'd soon die..)

Very low-end flash memory has that kind of write cycles. And it's typically limited to NOR flash, which is used only for code memory and limited data store due it its large cell size (largest NOR flash chips are around 256MB). Even so, Intel's StrataFlash had write lifetimes of at least 100,000 erase-write cycles, and most flash chips are underrated by an order of magnitude.

Modern bulk-dsta storage flash is NAND flash, which due to its smaller cell size (partly due to its design, and partly due to operation), means 16GB (byte, not bits) per chip is starting to become practical. NAND flash is faster erasing and writing than NOR flash, but much slower (order of magnitude) slower at reading. Plus it's I/O based - you can't "boot" from NAND flash like you can from NOR. (Write/Erase/Reads are on the order of microseconds for NAND - typically 100-500uS for write/erase, and 10uS for reads. For NOR, writes are typically 300-1000milliseconds, erases 1000ms, but reads on the order of 100ns or less).

Because of the operational characteristics of NAND flash, it typically has a 100,000 write-erase cycle limit at the minimum, with most offering at least 1,000,000 cycles (and typically lasts an order of magnitude more).

Wear-levelling algorithms and bad-block handling increase the time between writes and erases to the point where it almost isn't a consideration anymore - when the drive dies eventually, it'll really be timeto change it. And at least when an SSD dies, it dies on erases and writes, and very rarely on read. So if you get write errors, you still have a great probability of recovering all the data (except the data which was just written).

It's write-erase cycles, because erasing turns "0" bits into "1" bits. Writing turns "1" bits into "0" bits. Within certain restrictions, you can do multiple writes to a block (turning "1" bits into "0" bits, but you can't turn a "0" bit into a "1" bit without erasing), but those don't count towards write-erase cycles. (This behavior is often exploited when marking blocks as dirty and such). And they only fail on writes or erases due to internal timeouts (each cell takes progressively longer and longer to erase and write). Reads can be considered as never failing.