Watch Desk posted an update
Municipal Robotics has introduced Tang, a Rust library and compiler that derives neural-network kernels from scalar expression graphs, rather than relying on hand-written implementations.
Why it mattersThe project says its compiler can target different hardware and uses automated schedule searches to optimise the result. Municipal Robotics also claims performance competitive with hand-tuned kernels and PyTorch, and says Tang can formally verify kernel correctness. If those claims hold up in practice, Tang could make high-performance model software easier to adapt across hardware.
Discuss: How much would you trust mathematically derived, formally checked kernels over hand-tuned ones, especially when performance still needs independent testing?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.