This article of course focuses primarily on the consoles with an interesting gathering of some of the lower level details behind "fast SSD". But while it compares it to servers what it really brought to mind to me was mainframes, which often had not particularly impressive CPU specs at first glance but were kings of IO and offloading. It made me wonder if over the next decade we'll see yet another example of the cyclical nature of the tech industry in PCs. I haven't been following storage super closely for a bit, but it hasn't been that long yet the numbers the author tosses out in terms of "by the end of the year everyone will be doing 6-7+ GB/s" felt pretty staggering. Combined with the sudden leap forward in system bandwidth thanks to AMD, and after a fairly long period of stagnation around HDDs and the limits of SATA it feels like a huge amount of progress in storage is happening at an enormous clip. And that is leading to the discovery of bottlenecks elsewhere, like with ZFS Issue #8381 [1] where the ARC meant to improve performance actually becomes troublesome. AMD has really pushed available PCIe bandwidth up dramatically with EPYC 2 doing up to 128 GB/s, yet with drive numbers like those arrays could actually be made to get towards it already!

As the article says that kind of IO becomes a real strain on valuable general purpose CPU time. So I wonder if we'll actually start to see regular PCs begin offloading again, reversing the generalize everything a bit with more support specialized silicon so that they can actually feed the beast. I could see that having real implications for a lot of future software and OS development too, particularly for OSS. If things go that way I hope some good open standards can get out ahead of it.

----

1: https://github.com/openzfs/zfs/issues/8381