I suspect before the end of next year they will have enough data to bootstrap a production level self supervised learning environment.
I wonder if they do training verification with real pick and place circuits too.
Mathematics is generally done in the open. SOTA chip design is done by a handful of people all behind closed private doors.
So instead of transformer based learning, they can probably combine design space exploration with some sort of understanding/generalization.
The biggest question is about efficiency, simulation time, etc...
So far it doesn't exist.
So you have to figure out what training data is acceptable and also make deals to get access to proprietary data.
> SID (Synthesis of Integral Design) was a logic synthesis program used to generate logic gates for the VAX 9000. From high-level behavioral and register-transfer level sources, approximately 93% of the CPU scalar and vector units, over 700,000 gates, were synthesized.
> SID was an artificial intelligence rule-based system and expert system with over 1000 hand-written rules. In addition to logic gate creation, SID took the design to the wiring level, allocating loads to nets and providing parameters for place and route CAD tools. As the program ran, it generated and expanded its own rule-base to 384,000 low-level rules.
> Initially it was somewhat controversial but was accepted in order to reduce the overall VAX 9000 project budget. Some engineers refused to use it. Others compared their own gate-level designs to those created by SID, eventually accepting SID for the gate-level design job. Since SID rules were written by expert logic designers and with input from the best designers on the team, excellent results were achieved. As the project progressed and new rules were written, SID-generated results became equal to or better than manual results for both area and timing. For example, SID produced a 64-bit adder that was faster than the manually-designed one. Manually-designed areas averaged 1 bug per 200 gates, whereas SID-generated logic averaged 1 bug per 20,000 gates. After finding a bug, SID rules were corrected, resulting in 0 bugs on subsequent runs.
https://en.wikipedia.org/wiki/VAX_9000#SID_Scalar_and_Vector...
But that's not really relevant at all to asking Claude to fix your counter, or figure out the logic and layout for a certain goal, or whatever.
The big difference is in hardware it's much easier to measure cost/bug and days_of_schedule_slip/bug. Because we have those metrics, and people trust them, we can much more easily justify decisions that software teams never make. Like spending big on verification.
Why do you think we try to push everything into software. You guys are free, right? :)
It's really not though. Both AMD and Intel have learnt bitter lessons about trusting simulators too much. Chips and antenna design are hard because it's a chaotic system, small changes in placement can whack the whole interference pattern.
LLMs are very good at RTL (all I can consider is that the parallel posters arguing otherwise haven’t tried a modern one against significant RTL?). It’s easy to verify, the optimization constraints are relatively straightforward, and there’s adequate simulation to build a good gym for it.
LLMs are not good at layout and floor planning and because there is near zero public data available, I’m not sure how well they’ll iterate. For PCB design they’re improving because while it’s a similarly thorny problem, the physical constraints (trace properties on a given substrate) are highly public and can be simulated and turned into an RL gym. For ASIC, any public PDK is 10-20 years old so making an RL gym becomes outrageously difficult on top of the fundamental problem (highly multivariate/layered spatial optimization) being hard to solve.
Surely you can train an RL model to do this
anyone who thinks llms can't get good at this is engaging in wishful thinking
"mean normalized reward" hahaha
You have not had the joy of being introduced to the dark arts of Analog/RF chip design.