OpenLANE was mentioned in that episode, but that is run by eFabless
Traditionally, the semi industry had been very closed. Small companies struggled to get a custom chip made, and it was still $500K to $1M to start.
When efabless started, they wanted to design in the browser
Was the open PDK the gating moment?
The PDK is the IP of the fab. eFabless uses the 130 nm PDK from Global Foundries, 80 nm from Sky Fab (which are not open).
Does it have to be on the browser? No, but that made it easier to deal with closed source IP.
Lots of people are working on it around the world, even at highGer latencies
Black box design / only knowing the input and output
eFabless offers both open and closed IP
Mohammed worked in smartphone chip development, starting in 2000. He designed chips for smartphones at TI. He saw that there can be as few as two people designing chips at a company, but the rest of the company is designing infrastructure that makes it possible. What if this infrastructure was outside of any one company?
He asked if it can look like an app store, since there would be small players who could access the resources and develop small ideas.
Was this a validated idea? Do engineers want this sort of thing?
Looked at Topcoder as an example from the software world. There were also data points from the open source world.
What is the volume for making a custom chip?
The first chip off the line costs the entire NRE. Each additional chip amortizes that up front cost. Need to sell enough to cover the NRE cost with the margin in the chip.
eFabless want to reduce the volume requirement so it’s less of a hassle when someone is asking “Do I need an ASIC?”
That knowledge is residing on the IC side, so a system dev wouldn’t consider doing it
Reasons for using an ASIC (after getting the costs down)
OpenROAD can do 1M+ gates, previously was only 100K
GDS is “graphic design system”. It is like gerbers all in one file, but also has thickness information.
Xfab – 350 nm for high voltage, 180 nm for normal
Global Foundries 130G, can do front end with it
What does it take to get a new fab onto their platform?
GDS is readable in any tool, but it’s not as easy as it might be with large scale EDA
Start with design rules, which are in a PDF manual
Validating the designs and design rules is done against known designs. The number of layouts will go up a lot with the open source PDK, which is why getting more designs is important! This will allow people to push the rules
Analog process always trails digital
Hopefully this is all the beginning, with Skywater as a beachhead for convincing other fabs to open up.
Statistically speaking, more designs means more potential hits in the marketplace (for Skywater)
It’s like a currency: convert IP to a process technology is valuable, but doesn’t translate well to other places.
Parallel processing might be possible now that there are more people testing