Back to projects

Active project

TensionForge

A verifier-first OpenCL training runtime for legacy AMD hardware, built around exact parity receipts before performance claims.

Claim discipline: Claims on this page should be treated as hypotheses unless linked to a receipt, benchmark, or reproducible artifact.

What problem it addresses

Legacy commodity GPUs are poorly served by current ML stacks, while custom kernels can easily appear correct without proving numerical parity.

How it works

Implement persistent device tensors, fused linear/tension operations, backward passes, AdamW, and recurrent cells, then compare each stage against a reference implementation.

What exists now

Correct enough to investigate honestly; too slow to present as practical acceleration.

Evidence / receipts

The runtime demonstrates OpenCL SAXPY, tiled matmul, full linear training, and Ten-SON forward/backward/multi-step optimizer parity.

Limits

It is currently much slower than PyTorch CPU on the tested workloads. It is not a performance win yet.

Next milestone

Reach within 2× of the CPU reference or explicitly classify the project as a verification/research backend rather than an accelerator.

How to run / inspect

Install / open

git clone https://github.com/BoggersTheFish/TensionForge && cd TensionForge

Inspect

Run the parity harnesses, then profile launch count, dispatch overhead, small-matrix occupancy, serial recurrence, temporary buffers, and device readbacks.

Links

Focus tags

Numerical parityKernel fusionPerformance diagnosis

Related receipts