What 320 Lanes Taught Us About 4
When you can't fit what you want, remove the thing that seems like it's helping.
warp-core¶
The Snow That Wasn't Bandwidth
When the obvious diagnosis is wrong, the real bug is always simpler.
120x in an Afternoon
Three multiplicative improvements that each looked trivial on their own.
The Spec That Ate Its Registers
How a 62-register diagnostic braindump fed itself to an architecture and came out as 9 peripherals.
Debugging by Flashlight
When your diagnostic output shares a data path with the thing that's broken, you're debugging in the dark.
Zero New Opcodes
The entire warp scheduler fits in CSR space — no new decode logic, no new pipeline paths, no new encoding pressure.
Every Constraint Was a Door
The hardware limits that killed the original roadmap turned out to unlock a better one.
A Fourth Point in the SIMT Divergence Design Space
A divergence model the industry skipped, from type theory to silicon.