betrusted-io / betrusted-wiki

Looking for docs on Precursor/Betrusted? Start here.
https://github.com/betrusted-io/betrusted-wiki/wiki
Apache License 2.0
173 stars 6 forks source link

Chips for free #9

Open dumblob opened 3 years ago

dumblob commented 3 years ago

I just read https://fossi-foundation.org/2020/06/30/skywater-pdk and thought it might (or might not) be suitable for (parts of) betrusted stuff.

If you already heard of this, please wrap up your thoughts about that as it might be interesting to know what pros do think about it.

bunnie commented 3 years ago

It is interesting for a lot of reasons but there are many problems as well.

Two "for example" problems are (1) even if custom chips were taped out in 130nm, we have no way to confirm they are actually fabricated as specified. So before we go into a node, we need to build a post-fab verification methodology that can be performed in a non-destructive fashion, and ideally is "practical" enough that it can be done at all by someone without a million dollars worth of lab equipment. (2) 130nm node SRAM density is a bit low, this is touched on at https://betrusted.io/betrusted-architecture/. Basically at 130nm 4MiB of RAM alone is an area that's 8mm on a side, and I suspect most applications would want 2-4x that amount. While these aren't "ludicrous" geometries, they are on the big side for good yield and per-unit pricing.

On the other hand, it's hard to argue with free tape-outs. Since this project is 100% open source, nothing really stops anyone from trying to take our final verilog netlist and porting it to the Skywater PDK and just trying it out to see what happens.

dumblob commented 3 years ago

I know there are definitely more "problems", but I thought that (1) is being worked behind the scenes on and (2) was primarily motivated by price and not "general practicability" (I didn't take a look at the Skywater PDK technology, but if they provided some way of fast chip interconnection, they could make several 8x8 chips next to each other interconnected or they could somehow "stack them up" though I think this would be too advanced and thus is a nonsense; not talking about the need to analyze their interconnection to protect it against attackers with physical access).

dumblob commented 1 year ago

Heads up, they now provide 90nm technology: https://opensource.googleblog.com/2022/07/SkyWater-and-Google-expand-open-source-program-to-new-90nm-technology.html . That sounds already almost reasonable...

dumblob commented 1 year ago

Related: https://github.com/gsmecher/minimax/issues/2