Channels
#analog-design
Title
b
bluecmd
10/07/2020, 1:09 PM@User looking forward to hearing what you find out :)
k
Keno Fischer
10/07/2020, 1:12 PMI'll report back. I've talked to our simulations people, and they were all like "super easy barely an inconvenience" but they've yet to show me the code they promised. If they get it working, I'll do a blog post or something
The idea would be to take the standard spice model, but rather than running it on ngspice or xyce you just plumb it through a general optimizing model compiler, which should give you a whole bunch of programmability niceties for free. We'll see if it works out
b
bluecmd
10/07/2020, 1:14 PMIs it similar to Verilator / Icarus Verilog difference?
Or am I way off?
k
Keno Fischer
10/07/2020, 1:18 PMI don't know those tools, so I can't say. The pitch would be that the generic simulation world has built quite a bit of nice tooling for things like eventing that hadn't really been used much in the electronics world, because everyone is just stuck on SPICE
Also you might get niceties like GPU compute and distributed support for free
p
Pepijn de Vos
10/07/2020, 1:19 PM@Keno Fischer I assume the challenge would be implementing the built-in spice models. Like... the Skywater PDK does not contain like an actual model of the transistor in terms of differential equations, it just says, "yo spice, this is a nmos, model level 12, with this giant ball of parameters, you know what to do, right?"
b
bluecmd
10/07/2020, 1:19 PMMakes sense.
k
Keno Fischer
10/07/2020, 1:19 PMAlso xyce of course has the latter
@Pepijn de Vos Yes, that's the big piece, but our simulation folks said you can basically just black box it
b
bluecmd
10/07/2020, 1:20 PMIf it works and is very useful it seems reasonable that the fab folks could release a new model that's better for that particular simulation style
p
Pepijn de Vos
10/07/2020, 1:21 PMWhat does "black box" mean in this case? Call into spice for the equation?
k
Keno Fischer
10/07/2020, 1:21 PMRight, spice has an API that defines a device model
So may just be able to use that directly
p
Pepijn de Vos
10/07/2020, 1:22 PMSweet, that sounds very promising indeed.
k
Keno Fischer
10/07/2020, 1:23 PMPorting it would be better, since then the model compiler could see the structure of the model itself and use that for optimization
But just to get it working, I was told black boxing the whole model would be fine
p
Pepijn de Vos
10/07/2020, 1:23 PMSo what is this simulator that you are working on?
k
Keno Fischer
10/07/2020, 1:24 PMI admit to not really understanding the details though
The core computational engine is in https://github.com/SciML/ModelingToolkit.jl, but there is some work on top of it, that is still top sneaky and makes it look a lot more like a traditional sim tool
🌍 1
I'll write a blog post about it if we get it working. I was told to wait a week or two, while they add the black boxing support for me, so I'll try it once that feature is in. If it does work, I think it could be good, because we have a whole ecosystem around this stuff we should be able to leverage. We'll see.
The other thing that getting this running would be useful for is cosimulation stuff where you want to couple the device model to some other multi physics simulation, e.g. if you're building a regulator ic for some physical system.
p
Pepijn de Vos
10/07/2020, 1:44 PMomg, I'll add bondgraph support in that case haha
@Keno Fischer any updates on that? Sounded incredibly promising!
k
Keno Fischer
11/15/2020, 8:17 PMI'm told we have the prerequisites now, but I haven't had a chance to really work on it yet. I do still think it's a really neat idea, so I put in a funding proposal to actually let me take a team and do it properly. Will find out early next month if this got funded. If so, we should get some neat tooling. If not, I'll still do the blog post, but just haven't had the time so far.