Treat Elixir functions as circuit specifications and let the compiler decide clocks and registers—Elixir HLS 0.1.0 has taken the first step down that road. It currently covers only integer functions and fixed loops, its first target board is the Sipeed Tang Primer 25K running AtomVM, and the example is an FM radio whose entire demodulator runs in the FPGA. For people doing RTL, this is not a replacement but an entry point worth watching.
The core convention of Elixir HLS is that functions defined with defhls run as usual on the CPU while also being compiled into SystemVerilog circuits. mix ehls.compile turns each defhls function in a file into a circuit, then simulates and compares it against the corresponding function one by one, stopping if the results differ. In other words, the Elixir function itself is the specification for the circuit, and the acceptance criterion is behavioral equivalence rather than waveform comparison. This flow depends on hlstool and mlir-opt from CIRCT 1.160.0, uses Icarus Verilog for simulation, and requires Elixir 1.18 or later. For teams used to hand-writing RTL, this means the starting point of verification moves forward from the testbench to the software function.
The first version can synthesize integer functions and supports state, tables, memories, fixed loops, and connecting instances on FPGA or CPU through streams, hierarchy, and events; definstance is responsible for placing functions as circuits, and when no board is available they can also run as CPU processes. The official documentation explicitly states that division, floating point, recursion, and the parts of Elixir streams that require a full collection are not yet supported. The first target is the Sipeed Tang Primer 25K running AtomVM, and in the working example atomvm-fm-radio the entire demodulator runs in the FPGA. Types are bit widths: annotations such as u32 directly determine wiring and logic cost, and narrower bit widths buy smaller, faster circuits.
The example of a 32-bit core in the material is worth a closer look: the low 16 bits and high 16 bits are compared as signed numbers, and then the multiply-accumulate path is chosen. On the CPU the values are truncated to the same bit width, while in the circuit every bit costs resources, and only then do the two sides match. This detail shows that Elixir HLS is not pouring software code directly into hardware, but first imposing bit-width constraints at the type level. The real question is how this equivalence is maintained once division, floating point, and recursion are involved, and the official documentation has not yet given an answer. For high-speed interfaces and board design, the current capabilities do not yet reach scenarios such as SerDes or DDR controllers, but as rapid prototyping for control planes or preprocessing logic, the path is open.
In the short term, engineers in China will not put this into mass-production board flows: the lack of support for floating point and division means most stages of a signal-processing chain still have to fall back on hand-written RTL or mature HLS. But it provides a runnable reference—using software functions as specifications and simulation comparison as acceptance—and this approach has value for teams building their own HLS tools or internal DSLs. The toolchain consists entirely of open-source components, and there are no obstacles to obtaining CIRCT and iverilog in China; the risk is that Elixir itself has almost no installed base among hardware teams. Whether to follow up depends on whether you need a faster prototyping path than Verilog for control logic.
Daily Briefing · Weekly Digest · Monthly Deep Dive