Comment by walterbell | Hacker News Reader

Comment by walterbell | Hacker News Reader

Would companies be willing to contribute to the OpenSSD project?

OCP (Open Compute Project) has shown that customer-operators can cooperate on open hardware designs, successfully influencing enterprise hardware supply chains. Commercial DPUs and SmartNICs were preceded by a decade of open hardware and research by the NetFPGA project (https://netfpga.org). Why not DiskFPGA?

2017 OpenSSD overview, based on Xilinx: https://github.com/Cosmos-OpenSSD/Cosmos-plus-OpenSSD/blob/m...

2022 status, http://www.openssd-project.org/

> OpenSSD platforms are still being actively used in many academic institutions. As of June 2022, we have renewed the homepage hoping that this site will be a forum to share various simulators, tools, traces, etc. not only for the conventional SSDs but also for the upcoming storage devices such as KVSSD, ZNS SSD, and Computational Storage (CSX). This site is being maintained by Systems Software and Architecture Lab. at Seoul National University as a part of the SW STAR Lab. project.

An open source SSD is also a lot more feasible than an open source hard drive. Even if you managed to get an open source HDD controller, you still need the precision mechanical parts that are impossible for the average person to make. With SSDs, however, it’s just a PCB with ICs.

Edit: this obviously ignores any troubles one would have sourcing the ICs (such as possible NDAs)

Is there anything special about making SSDs that the average person would not be able to do or is it a "if you can outsource PCB printing and maybe solder you can make one" situation?

The limiting factor would be the memory chips themselves and any firmware required for them (if any). I also don't know how well they are spec'd and if full documentation is available without NDA's and lawyers.

Looking on mouser and digikey it doesn't seem like flash chips, even into very fairly high density on a single chip[eg. 1], are all that difficult to get and get info on, though they all have very high minimum volume orders. So if a person wanted to try to do this on their own they'd probably be best off finding like 50 friends to go in on the order with them.

[1] https://www.mouser.ca/datasheet/2/671/micron_technology_mict...

The flash you linked to the datasheet for is over a decade old. Any SSD built from it would fall short of adequate by an order of magnitude in every important metric. The per-die capacity of typical current-generation NAND flash is 16x larger, the interface speeds are 8x higher, erase blocks are 24x larger, program latency can be 5-10x higher. And most importantly, all mainstream SSDs now use flash that stores three or four bits per physical memory cell, rather than one.

So using that flash, you could build something that is recognizably an SSD. But it would be almost entirely useless: too expensive and too small and slow for production use, and too far removed from the current state of the art to serve as a research platform for the most important challenges the SSD industry has been dealing with for the past several generations (error correction strategies for TLC/QLC, and SLC caching).

A couple things:

- I just picked one at random. I'm sure the bleeding edge is harder to get and datasheets are harder to get, but I wasn't trying to find the newest or best.

- the specific subthread here is about the diy-ishness of ssds vs. spinning rust, where the difficulties are of a fundamentally different kind. I feel like it goes without saying that a home built ssd is not going to perform to the level of mass production devices, the question was just can you.

The new, fast, high density flash chips from the big name flash chip vendors are not generally even listed on the vendor websites. You have to talk to a sales person and convince them you're actually going to buy in volume to even get data on the latest generation of flash ICs. You will also likely need more than 50 friends to meet the order minimums, unless your 50 friends each want to buy about an ExaByte worth of flash chips.

Also, with these multiple level flash technologies (quad level is current tech, triple is still used in some SSD/NVMe) the read, write, and ECC algorithms are non-trivial to the point where last I checked even mainline Linux's raw flash driver support won't do anything beyond single level cell flashes (and very few new embedded designs are choosing raw parallel NAND flash, instead opting for things like eMMC or UFS which have built-in controllers to handle this).

Like I said in another branch of this subthread, the question wasn't "can you build a high perf flash drive yourself" but "can you build an SSD more easily than a magnetic drive."

The comparison here is that no matter how much you hunt on digikey you won't find a disk platter or drive head or any of the other precision machined parts that go into a hard drive (never mind putting them together and keeping dust out etc).

Taobao has many sellers of pcb/controller and flash chips. There are even sections dedicated to DIY SSDs at Chinese forums.

But they all rely on leaked manufacturer firmware production tools, not open source firmware.

An FPGA-based SSD will always be too expensive for people with truly large scale. The controller ASICs are a lot cheaper.

Yeah, if we want an open SSD the path would be for a hyperscaler to strong-arm a controller vendor like Microsemi or Marvell into opening up their SDK. This worked with various Broadcom ASICs so it's not impossible.

Reply on news.ycombinator.com