Running IBM COBOL lab files locally

z/OS mainframe to Fedora Silverblue I have finally been able to get the IBM COBOL compiler installed and running on the ThinkPad correctly. I've now started to experiment with the IBM lab COBOL files and get them running locally. The first one is CBL0001. It deals with a data set of presidential names with account information. I downloaded the data set using zowe. Claude and I had to make six changes to the source code in order to…

https://felixquinihildebet.wordpress.com/2026/03/02/running-ibm-cobol-lab-files-locally/

YubiKey GPG Recovery Adventure

Recovery from a full Admin PIN lockout in OpenGPG What Went Wrong? I am not sure. For some reason my three PIN tries permitted by YubiKey were exhausted plus my Admin PIN was locked (after 3 tries). Not even my management key or reset key was of much help. What caused this is unclear. I thought it was hacking but Co-Pilot did not think so given the condition of the YubiKey particulars. All I had was a YubiKey that was completely locked on…

https://felixquinihildebet.wordpress.com/2026/02/28/yubikey-gpg-recovery-adventure/

The numerical dissipation introduced by a cheap semi-Lagrangian solver is a deterministic function of the current field state.

https://felixquinihildebet.wordpress.com/2026/02/16/turbulence-spectral-lab-notes-on-neural-correction-of-fast-fluid-solvers/

Scientific computing with large datasets (over 70GB) can be conducted on an 8GB RTX4060: WeatherBench2 ERA5 dataset is training (untuned of course) on our fortran cuDNN training engine with U-Net CNN. On github soon (after some further training runs).

https://felixquinihildebet.wordpress.com/2025/11/24/training-the-72gb-weatherbench2-era5-dataset-on-an-8gb-rtx4060/

currently refining the unified memory feature, does anyone have ideas for large dataset scientific projects that hobble PyTorch? CuPy and PyTorch crashing because of GPU memory is a constant annoyance.

https://felixquinihildebet.wordpress.com/2025/11/20/can-someone-suggest-a-scientific-project-involving-a-large-dataset-that-is-difficult-for-pytorch/

We now have a full workflow from data loading to training to inference that has access to all the python ml libraries but is able to train 4x faster than PyTorch. We can review the samples, check out the confusion matrix and do all the other tests as normal.

https://felixquinihildebet.wordpress.com/2025/11/19/cnn-fortran-training-engine-workflow-is-now-complete/

Oxford flowers 102: cutting training from 7 minutes to 1 second

a two stage methodology offers advantages A rather large change has taken place after the 5 month CIFAR-10 ordeal. In one day I was able to get my old Oxford flowers 102 training workflow up and running but what surprised me was that I was able to get the training time down from 7 minutes for 100 epochs to 1 second. With a two stage methodology I am also able to offload a lot of the work to existing…

https://felixquinihildebet.wordpress.com/2025/11/14/oxford-flowers-102-cutting-training-from-7-minutes-to-1-second/

CIFAR-10 cuDNN reaches PyTorch parity. We solved the column major/row major puzzle for a cuDNN intrinsic along the way.

https://felixquinihildebet.wordpress.com/2025/11/13/cifar-10-cudnn-reaches-pytorch-parity/

How to restore YubiKey in Fedora Silverblue 43

https://felixquinihildebet.wordpress.com/2025/10/30/how-to-restore-yubikey-functions-in-fedora-silverblue-43/

I noticed some things about Claude Sonnet when debugging a +10,000 line code base over the course of a month.

https://felixquinihildebet.wordpress.com/2025/08/21/observations-after-a-month-of-debugging-with-claude-sonnet/