---
pretty_name: "The Hidden Canopy | Three GPU Optimization Receipts"
license: other
language:
- en
tags:
- gpu
- machine-learning
- performance
- training
- cuda
- opencl
- nvidia
- amd
- benchmark
---

# Three GPU optimization receipts

Three **curated, hardware-specific research benchmark summaries** from The Hidden Canopy. Published for reproducible discussion of measured or explicitly projected performance and its limits. The JSON documents are the three receipts; this README is only the dataset card.

| Receipt | Device | Headline | Status |
| --- | --- | --- | --- |
| AMD OpenCL hybrid | Radeon 610M / gfx1036 | 13.12 → 107.31 tokens/s; about 8.18× on Swift-shaped canary | Internally reported completed test, CPU/GPU hybrid |
| Omni native training | RTX 4050 Laptop | ~82 → ~1,900 tokens/s; ~170× projected cumulative wall-time improvement | Published source JSON; optimized full-run time is **estimated** |
| 1F1B pipeline | 2× RTX 5090 | 42,213 → 53,968 tokens/s; +27.9% on matched small ablation | Internally reported 20-microstep experiment |

## Read before comparing

These are **three different workload shapes**, not three entries in a common model-throughput ranking. No claim is made that they outperform vendor-native NVFP4, AMD Instinct, or any unrelated production benchmark.

The RTX 4050 ~170× number is **a projection**, not a completed full-run stopwatch result. The AMD receipt includes work performed on CPU as well as GPU. The dual-5090 result is a short, matched pipeline ablation rather than a completed convergence test.

No private host addresses, user directory paths, infrastructure credentials, proprietary weights, or unrelated source files are part of this release. Original private-repository reports remain private. No independent third-party audit is implied.

The receipt IDs and source references are provided so the owner can produce the underlying run material to a qualified verifier.