AI Compute
Preview. This section starts to catalogue the AI accelerators, NPU IP, and AI SoCs in the RISC-V ecosystem, and to define the fields and scoring used for the upcoming “model / operator / software-version maturity” matrix.
Current data covers only publicly documented hardware specifications and toolchain information; items without a public benchmark are not given invented scores. A few ARM-based AI SoCs are kept for reference and are explicitly marked in the “architecture” field and in the filters. Measured results and corrections are welcome via issues.
| Product / vendor | Type / origin | Throughput | Precision | Frameworks | SDK / toolchain | Maturity | Representative silicon |
|---|
No records match
Try loosening the throughput or framework constraints.
Next: model / software-version matching maturity
Hardware throughput is only the starting point. What really sets the project schedule is whether a given model runs on your toolchain version, how much accuracy quantization costs, and whether operator coverage is complete. Each AI record will grow into the matrix below.
Throughput and precision
INT8 / INT4 / FP16 peak, memory bandwidth, power and energy efficiency, package and interfaces.
Toolchain versions
Version combinations of SDK, compiler, quantization tools, drivers, and runtime, with install and reproduction steps.
Models and operators
ONNX / TFLite / PyTorch export, operator coverage, dynamic shapes, and post-quantization accuracy regression.