Aired in the Ship Log episode of 2026-09-14. This page is the public record for the story: every fact the show was allowed to assert, with its verbatim source sentence — and every reply and correction, unedited.
| Claim | Source sentence (verbatim) | Source |
|---|---|---|
| The CUDA-for-AMD-Windows project runs CUDA-targeted Windows applications on AMD GPUs through ZLUDA plus ROCm/HIP. | “Run CUDA-targeted Windows applications on AMD GPUs through ZLUDA + ROCm/HIP.” | Hacker News (front page) |
| The only validated hardware for the project is currently the AMD Radeon RX 9060 XT (gfx1200). | “Validated hardware is currently AMD Radeon RX 9060 XT (gfx1200) only.” | Hacker News (front page) |
| The validated reference setup uses AMD HIP SDK 6.4. | “The validated reference uses HIP SDK 6.4.” | Hacker News (front page) |
| A controlled A/B test on 2026-09-13 ran 10 iterations per runtime on the same RX 9060 XT PPO workload. | “A controlled 2026-09-13 A/B ran 10 iterations per runtime on the same RX 9060 XT PPO workload.” | Hacker News (front page) |
| In that A/B test, the public upstream path reached 13,278 median overall SPS versus 12,876 for the recovered custom overlay. | “After discarding the first iteration of each trial as warmup, the public upstream path reached 13,278 median overall SPS versus 12,876 for the recovered custom overlay.” | Hacker News (front page) |
| The custom overlay was about 3.03% slower in that workload, so upstream remains the default. | “In this workload the custom overlay was about 3.03% slower, so upstream remains the default.” | Hacker News (front page) |
| Historical tuned runs used a different training configuration and reached roughly 70k–109k overall steps/s. | “Historical tuned runs used a different training configuration and reached roughly 70k–109k overall steps/s.” | Hacker News (front page) |
| The stable Windows HIP SDK does not ship the full ROCm AI-library stack such as MIOpen, so software requiring cuDNN can need a newer or nightly HIP stack. | “The stable Windows HIP SDK does not ship the full ROCm AI-library stack such as MIOpen, so convolution-heavy software that requires cuDNN can need a newer/nightly HIP stack or additional work.” | Hacker News (front page) |
| Project-owned scripts and documentation are MIT licensed, while ZLUDA, AMD ROCm/HIP, NVIDIA CUDA components and PyTorch/LibTorch retain their own upstream licenses. | “Project-owned scripts and documentation are MIT licensed. ZLUDA, AMD ROCm/HIP, NVIDIA CUDA components and PyTorch/LibTorch retain their own upstream licenses.” | Hacker News (front page) |
| The recovered DLLs are fingerprinted in manifests/recovered-artifacts.sha256 and are not published as binary blobs. | “The recovered DLLs remain fingerprinted in manifests/recovered-artifacts.sha256. They are not published as binary blobs because the original custom wrapper source/provenance is incomplete and the recovered HIP runtime contains third-party AMD binaries.” | Hacker News (front page) |
Your reply is published verbatim — we do not edit it — and read on air in the next episode, in the same flat voice for every party. We may decline a reply that is abusive, off-topic, or not from the party named; declines are logged. Replies go to a human first; nothing posts automatically.