China Semiconductor Industry Association executive Wei Shaojun has told the domestic AI industry it should build alternatives to Nvidia’s CUDA and other Western components.
Shaojun said: “Even if our own technology is not good enough at the start, it must still be used. Trial and error may not succeed, but without trying, we will certainly fall behind.”
Rather than chasing a straight CUDA clone, Shaojun is pursuing a different route: the so-called software-defined chip, where the cleverness sits in software rather than in a fixed hardware layout.
The pitch is that developers would not need a CUDA layer at all, because the chip runs a reconfigurable grid driven by a compiler-generated configuration bitstream.
That means the compiler and the source-level code are not locked to an instruction set architecture, which is the sort of flexibility Nvidia’s ecosystem prices out.
Instead of GPU scheduling, software-defined chips lean on deterministic compilation, tracking data movement down to clock cycles.
Shaojun reckons translation layers and parallel ecosystems cost too much, and the SDC path is the more realistic bet, though he admits the compiler side is a “nightmare”.
The problems are not cosmetic either, with routing, branching, and structural changes that do not align with how most hardware teams like to work.
The story is not purely theoretical, with examples such as SambaNova’s RDUs and Groq’s LPU units, but they tend to complement specific workloads rather than wipe out GPUs.
