Introducing Frostbite: a new class of CPU built around interaction nets.
Given only interaction net theory, our agents co-designed the ISA and chip architecture from scratch, running large swarms to design, verify, and optimize the CPU down to the final GDS.
On our first set of target workloads, Frostbite is 4.6x more performant per watt than modern traditional CPUs at similar area.
@OpenAI@AMD I'm curious how these models handle something as sequential as routing, since the context length of any large design's node's will be soo high.
@suarezvictor it comes down to a difference in how TritonRoute and i decide to handle cost. I'm implementing the pathfinder approach sing negotiated A* but TR does something different.
Another reason's that my compiler is wayy lighter lol
I'll write a much more technical post in a few days :)
i spent today building an asic router - it finished detailed routing 16x faster than OpenROAD's TritonRoute on any design i give it.
It meets timing and reduces area and cell usage by 30% thanks to an equally good floorplanner.
@Atulya_Ayush Yosys for synthesis and equivalence checkers, and OpenROAD for routing, sta, extraction tool. I'm not doing simulation checks might look into it later.
llms and coding agents will be awesome at rtl design but what area or timing to choose is still a human decision to make, decided by the cost of actually making an asic.
my compiler now has extended synthesis tools! here is the benchmark against yosys+ABC
Built for TinyTapeout IHP
(ms = time taken to place by my logic vs by yosys)
What if you could learn something about someone's phone activity just by observing how quickly it responds to messages?
No access to their chats. No need to read message content.
I wanted to explore this possibility through WhatsApp.