Christos Argyropoulos

@ChristosArgyrop@mast.hpc.social
60 Followers
73 Following
374 Posts

Interested in #HPC for clinical (electronic health records) and bioinformatic (sequencing) research.
In perpetual student mode (in 2024 this means lots of #OpenMP, #OpenMPI and #Fortran!)

Associate Professor and Chief of Nephrology Department of Internal Medicine,
Faculty Lead, Hub Research Capacity, Clinical and Translational Science Center, University of New Mexico Health Sciences Center, Albuquerque, NM,USA

Redditu/ReplacementSlight413
Google Scholarhttps://scholar.google.com/citations?user=Hj15fvUAAAAJ&hl=en
mstdn.science@ChristosArgyrop
Programming LanguagesR | Perl | C | Mathematica | SQL | Python | Fortran
Christos Argyropoulos16h ago

Truth about "thinking" LLMs (screenshot).

I am currently using for a pet project and it has been behaving like this. At least it saves me from the headaches of looking up stuff for the makefile

Source:

https://x.com/burkov/status/1934129008108568682?t=wGjoAmKD_vrlyLJdF9eFoQ&s=19

Christos Argyropoulos5d ago
Show threadBalthasar Reuter5d ago
@ChristosArgyrop @openmp_arb depending on the GPU hardware that you use, single precision execution may utilise tensor cores or produce different residual loop behaviour. With Nvidia compiler, options -Kieee and -gpu=uniform (and not using fastmath) may restore a more coherent behaviour.
Show threadChristos Argyropoulos5d ago
@reuterbal @openmp_arb It was just bizarre to see this (actually turning off optimization flags makes this error go away - which apparently has been seen in other contexts). It's most likely a gcc lm issue.
Christos Argyropoulos5d ago
The problem with AGI/ASI and LLMs (thinking or not) is that *linear algebra* (the matrix multiplications from high school),can never be sentient in the same manner as differential equations of the transport processes inside the kidney can produce homeostasis (or make urine).
Christos Argyropoulos6d ago
Assn for Computing Machinery6d ago

In an opinion piece in The Conversation, #ACMTuringLaureate Jack Dongarra takes a stand for traditional high-performance computing (#HPC), opining that AI's reliance on HPC necessitates greater focus from the U.S. government.

"US national strategy should include funding new machines and training for people to use them," he writes. Read more: https://theconversation.com/challenges-to-high-performance-computing-threaten-us-innovation-255188

Christos Argyropoulos6d ago
This is really bizarre an error with @openmp_arb (or gcc) when offloading to a #gpu. Double precision version of the code that uses trigonometric functions offload and executes flawlessly. Single precision version of the same code dies with an illegal memory access. I suspect it has to do with the lm library (or rather with sinf cosf). Is there an alternative math library for gpus to try?
Christos ArgyropoulosJun 8
Fighting the HOA surveillance drones one desert baby bunny a time
Christos ArgyropoulosJun 7

Learning #Perl

https://dev.to/lnation/learning-perl-introduction-3d1a

Learning Perl – Introduction

Perl has long been known as the “duct tape of the Internet,” or "the Swiss Army chainsaw of scripting...

DEV Community
Christos ArgyropoulosJun 6
Since we can't use the cloud to automate our #EHR analysis projects , we tried #nextflow (traditionally used in bioinformatics https://www.nextflow.io/), and it worked as a charm coordinating #duckdb #R #python and within node tasks.
Next in line is (R)?ex https://www.rexify.org/
A DSL for parallel and scalable computational pipelines | Nextflow

Nextflow enables scalable and reproducible scientific workflows using software containers. It allows the adaptation of pipelines written in the most common scripting languages.

Show threadChristos ArgyropoulosJun 6

@jannem Let's make this concrete (ThS: Throughput Per Second)

| Compiler Flag | Algo A | Algo B |
| -O0 | 50 | 10 |
| -O1 | 70 | 40 |
| -O2 | 80 | 70 |
| -O3 | 90 | 90 |

Which one is the better algorithm?