Hello, so I'm Francesco Lanutti from NGspice team and these guys are the other members.

So a little bit of history. SPICE is born 42 years ago at Berkeley, mainly from Larry Nagel PhD.

And NGspice started in 1999 by Paolo Nancy with the main goal of fix the original SPICE-3 code

and add new models like B-Sim-S-O-I. But now NGspice is a free open source software,

G-set on SourceForge. So now the goal is to keep the pace with the commercial tools and maybe

adding a new features. So SPICE is de facto standard for circuit simulation in both academic

and industrial field. So SPICE was designed to handle simple circuits, but now we have system

packages, system on package, system on chip and so on. And we have complex model like deep sub-migron MOSFET. So SPICE and also NGspice can handle all of this after 42 years and this

is something great. The commercial SPICE are still based on some code from SPICE-3. For

example, we expect that SPICE and SPICE-like C-mu letters will be active for the next years. So we continue to support NGspice and we also have other projects like Quacks,

SPICE team is spread across the Europe and we have three new members, Justin Fisher, Marcel Hendricks and Francesc Villa and contribution are user driven pretty much and we

try to maintain a strong code without including everything. So we test the code first and then

we create a new branch to include the code in a separate place and then we merge with the master.

NGspice was registered on Sourceforge in 2001 and now it counts 300k downloads pretty much

which are spread across these five countries. So it's a quite big community, less than Quacks

but still big and NGspice has a huge ecosystem. For example, FEGO which is a commercial tool but use NGspice for some simulation. IMOS which is sort of cloud computing from Professor

Mansunchan and it use NGspice on the cloud. We have Oscar the team which is presented by the transistor symbol, this one. It's an Indian tool for developing PCB and it use NGspice

simulator tool and then other front-end or tools that use NGspice somehow. Okay so NGspice

is divided in this way. We have the main SPICE3 core here and then we have some extension

like NGspice and tool for device simulation, a sort of T-CAD which is 1D and 2D only.

Then we have some extension to the front-end world. For example, the old Nutmeg which is the original front-end of SPICE3. So it's based on Xe11 is very poor but for plotting for base

plotting is fine. Then we have T-CoreSpice. For example, here you can use T-CoreScript to

drive NGspice compiled as a shared library and we have an interface with some T-CAD tool. On the other side we have modeling extension so we use also ADMS. We have a couple of models

included. I will show you later and then the Xspice code model toolkit. Okay this is the

list of the model which is implemented in NGspice. The bold ones are Velogay models and they are converted in C using ADMS like Quax so it's the same concept but we don't have

V-Sim 6 right now. The italic ones are the code model from Xspice. We have some

improvements to the original SPICE3 because for example we can create components which are

expression-defined like this resistor and then we have a lot of different source generator

like PBL, PAH, PALS and so on and we have non-linear control at the sources like this V-GVS and V-GGS. Then we integrated some KSpice model for example transmission line and coupled

multi-conductor line so you have to just include the model and describe the parameters.

From the semiconductor point of view we have a lot of models starting with the old Diode, then BJT, JFET and so on. We have a lot of MOSFETs until V-Sim 4 version 8 which

is the latest one and V-Sim 3 which is old but still maintained and we have some compatibility with PDKs so we maintain all the version to be able to run PDKs. We have AKV

inside in version 2.6 which is open and is in Velogate. So as I said we use a DMS by

Laurent Lemaître which convert Velogate into C and more precisely this process is divided into step. So first of all you have to use a DMS XML which is the real tool that

read the Velogate and convert it in a sort of XML3. Then NGspice has this XML script that pass the XML3 and translate in the so-called SPICE3 kit which is a list of C file that is

your model ready to be compiled by NGspice. One problem in this process is that we don't

have the dynamic loading so you have to modify a couple of files to inject the model inside NGspice and recompile because everything is static so it's linked inside the executable.

This is an old style compiling stuff but we didn't modify it for a long time. AMS and this model has some problem for convergence and noise is not implemented

but we are working on this. So the roadmap for the model include a lot of things but mainly

the introduction of Visyn6 and Visyncomo Multigate and the integration of NXP SYMkit mainly for PSP models. On the front-end side we have a lot of improvement especially bug fixes

and fix of memory leak and we support parametric net, netrist and a command which is called dot

measure which is a post processing tool so you have to write what you want to be computed

and you can compute for example the propagation delay of standard cells if you do standard cell

characterization so you just pick the two points you want to measure and extract the delay and if I remember correctly also the derivative of the trans characteristic.

We have Monte Carlo analysis also and it's based on our random generator.

Okay TicoSpice said something on TicoSpice and I skipped this. T-CAD. Okay in this case we have two approaches. One is linked to an external T-CAD software by Cogenda which is

partially under GPL and the other ways use CIDR which is inside ngSpice and we support just these five models so it's 1D and 2D models. We have a new periodic steady-state analysis.

This way you can simulate RF systems and at this moment we just support our autonomous

system to extract period and DFT. Non-autonomous systems are ongoing but they are pretty easy. We have to just insert the node but the other part is on the way and it's not supported yet.

This is an example. The solution is not very good but you can see the period of this OC and the frequency plot. Okay some slide on Veriloguet. I'm building a new Veriloguet

compiler from scratch mainly for my PhD and mainly because I don't know anything about Flex Bison and so on so I wanted to learn. So pretty much I'm starting from scratch and you

want to use C or C++ as main language. Flex and Bison is useful tool for Flex and partial and T-GOL as user interface. I have also semantic analyzer. So this part will be the

front end and then you have to include the T-GOL script which will be provided with NGspice to extract the model in C. So it's not really different from ADMS but I don't

use XML and I want to support every construct of Veriloguet. So I don't want to simplify anything. So at the moment I create the preprocessor. The preprocessor is working.

Just go faster if you have some question I will ask them. And then I have the main parser which is ongoing. It reads the output of preprocessor step. This is an example. So

you have the console here. Mouse. I have the console here. So when you reach the console,

the parser has already parsed the model and created the structure. So you can query objects and attributes of the model. For example, this is very basic stuff. You query

the disciplines and for example, you want to retrieve all the name of the disciplines. In this other example, you want to extract, for example, the name of the modules. You

have just one modules. And then the name of the part of the module is in comma multigate. This follow, let's say, the synopsis way of things. So it should be similar to,

I don't know if similar to Hspice, but for sure we design compiler and IC compiler methodology. Just a quick overview on acceleration. We have multi-core, a massive parallelization

and KLU inside the NGspice. So openMP implementation achieve pretty much 2x,

but this depends on the circuit and then you just stay on 2x because there is a problem, I think, in the reduction code. So we should fix this, but I didn't work on this part. So this is the plot just to show the speedup. Then I started KLU inside the NGspice and we

benchmarking suite. So if you look at the biggest one is 13x. Then I created Kuspice,

which is NGspice for CUDA. It's just a concept. So it doesn't support everything,

it just supports the minimal models. But if you want to extend this pretty easy, it's just long. So it's long to extend to every model. So I just created this basic stuff

and I skip this. If you want details, we can discuss how it works. There is a new algorithm here to create the matrix and the right hand side so you don't have the reduction tree

to sum the contribution in the matrix, but you just apply matrix vector multiplication using a topology matrix. You have two topology matrix, one for the circuit max and one for the right hand side. And the speedup is 9x with the same benchmark suite. So if you use KLU

and this one, you gain almost 100x, more than 100x. So in conclusion for this part,

we have roughly 2x for RobinMP, 13x for KLU, and 9x for CUDA. We want to improve this, for example, we didn't approach the changing of the model code, so we just simulate every device in the order. But if you reorder the computation, you can avoid some thread divergence

inside CUDA. And then we want to extend to the matrix solver, the parallelization technique, but this is really tough and the speedup depends on circuit. So in conclusion, SPICE is here

since 42 years and is used for both academic and industrial simulation. And NGSPICE is on

SourceForge, you can freely download it. It supports Gita's versioning system, so you can create branch, fork, and so on to add the contribution. So please visit the

site. Yes, DMS, I understand correctly. You said AADMS tool has this licensing problem.

So the question is if all the script are free, right? Okay. So NGSPICE is under BSD license,

so the original SPICE3 license, so no GPL problems. You can download and integrate in commercial tool, for example. You have to just maintain the name of the author. So from NGSPICE point of view, there is no kind of licensing issue, just this small thing. So

for example, if you want to download the CUSPICE and integrate in a commercial tool, you have to maintain the initial copyright, which is from NVIDIA, because I developed this

at NVIDIA, so they release for free, but you have to maintain the statement. Okay. Oh, okay. So the question is why I choose CUDA and not OBE-NCL. Okay. So

basically because I learned CUDA, so I don't know OBE-NCL very well, but at NVIDIA they use

mainly their compilers, so their CUDA, and I just used that one because I was there. So it's just this one, not anything else. So the question is if there is some collaboration

between NGSPICE and CUAX, I have to say we are working on it because we discussed the way to enhance ADMS or maybe use my compiler for the first part and then ADMS for the second part,

we have to say see what happens because ADMS is not complete, so it supports just a subset of AMS. It doesn't support all the construct that maybe you want to have in the model,

because from my point of view, let's say the main thing is to avoid to rewrite the model. The model has to be written by the model developer, just to grab the model and do whatever you have to do. So we are working on this. Okay.