AMD Developer Hackathon · Unicorn Track

Any CUDA. Any Hardware. Verified.

Drop in a CUDA kernel. Our AI comprehends it, our Oracle mathematically proves it, and our compiler lowers it to any architecture — RISC-V GPUs, ARM, x86, and beyond.

5
Hardware Targets
12+
CUDA Patterns
100%
Oracle Verified
0
Silent Failures

Why POLYFORGE?

01

LLM Comprehension

Fireworks AI Kimi-2.6 reads raw CUDA and understands thread indexing, synchronization, and memory patterns. No manual annotation required.

02

Zero-Trust Oracle

Independent Clang AST simulation verifies all threads before touching silicon. Catches LLM hallucinations and data races that hardware can't reliably detect.

03

Hardware Lowering

Write CUDA once, run anywhere. Verified kernels lowered to RISC-V GPUs (Vortex), ARM Cortex-A72, ARM Cortex-M0, x86-64, and generic RISC-V.

04

Race Condition Detection

The Oracle detects Read-After-Write (RAW) and Write-After-Read (WAR) hazards that make parallel execution unsafe — before deployment.

05

RTL Capability Analysis

Scans Verilog/SystemVerilog RTL for thread_id, parallel_spawn, barrier_sync, and addressable_memory primitives. Know your hardware before compiling.

06

Full Execution Transparency

Every subprocess streams stdout/stderr in real-time. 100% visibility into what's happening at every stage of the pipeline.

Ready to forge?

Explore the full source code, pipeline, and documentation on GitHub.

Star the Repo →