Switchboard: Calling All Hardware Models
summary
Switchboard wires RTL, FPGA, C++ and Python models via simple 52-byte packet messages over mmap-backed shared-memory queues that present a ready/valid, single-producer single-consumer channel. Each link hits up to 1 GB/s / 200 ns latency in practice, exposing queues as named files so RTL on simulators or F1 cloud FPGAs and software all talk the same protocol. Built-in adapters include axi and Umi; Zero ASIC runs it to boot Linux and run ML models near realtime.
AI-written summary · about
chapters (7)
- 0:00Switchboard framework intro
- 2:09Switchboard architecture
- 7:02Python and RTL example
- 9:30High-level interfaces like AXI
- 12:03Mix-signal simulation
- 13:30Performance and demos
- 14:33Q&A on QEMU, FPGA, Wishbone
related talks
15mDRAM simulation with the simulator DRAMSys
24mFireSim: Cycle Accurate Rack Scale System Simulation Using FPGAs In The Public Cloud
22mIntroducing 6-pack, the First Open Hardware Modular Switch
16mRuntime Programmable Switches
20mVirtual Prototypes and Open Source Hardware Design in Research and Education
16mSoftware-driven Hardware in the Data Center: LibreBMC, Renode, RowHammer...
53mProgramming Languages & the Power Grid
24mComplex system simulation and HW/SW co-design with Renode open source simulation framework
