3...2...1

I am taking a break from Twitter

See you on the other side, provided both I and this platform are still alive :-)

#HPC

Are you wondering how to make you #SC24 SC24 submission even better?

Check out the scientific benchmarking paper for parallel systems.

Twelve rules to improve the scientific quality of your #HPC paper.

htor.inf.ethz.ch/publications/i…

Congrats again, Bill!

'Computing & data, especially but not only AI & ML, are transforming science & society, & it is more important than ever that the AAAS speak up for both the opportunities & the risks in applying computing technologies.' :Bill Gropp, NCSA Director

If confirmed, that’s ~100x the cost of the largest supercomputers used for traditional #HPC (Simulation etc.)

Drop the 6 from #FP64 , FP4 is where all the💰is 😜

#AI #GenAI OpenAI Microsoft

Our NVIDIA Compute Sanitizer is a powerful tool that can save you time and effort while improving the reliability and performance of your #CUDA applications.

🐛 Learn how to efficiently debug code in the CUDA environment:

nvda.ws/3VuY1Hs

Do not use AI detectors in schools!

Accuracy of AI detectors, already low, dropped to 22% when researchers added simple changes, like purposeful spelling errors. Worse, the writing of non-native speakers are more likely to be falsely flagged as AI writing arxiv.org/ftp/arxiv/pape…

Through the use of very efficient sampling methods enabled by a new Scalable Ensemble Envelope Diffusion Sampler (SEEDS) technique, neural nets now rain supreme in medium-range weather forecasting, and in particular can more accurately predict extreme weather events at longer…

You might try this link as the article below is behind a paywall

Microsoft would likely be responsible for financing the project - they expect to launch Stargate by 2028

reuters.com/technology/mic…

#AI #Supercomputer via @EvanKir

Microsoft and OpenAI Plot $100 Billion Stargate AI Supercomputer theinformation.com/articles/micro… #HPC HPC Guru (Taking a break)

Full house for my lecture hosted by the Center of Mathematical Sciences and Applications at Harvard yesterday.

Video will be available from the CMSA YouTube channel.
Slides here: drive.google.com/file/d/1Ymx_LC…

#RayBanMeta

By assuming that technical talent is fungible and interchangeable, a previously-successful engineering company will be run into the ground...
...along with a few of its own products and everyone on board.

It's astonishing how certain types of managers will prefer…

Compute-memory balance: NuLink is a PHY with ~10X data rate of a MCM package and traces can be 2-3 cm between NuLink PHYs, which is a 20X to 30X increase over the 0.1 mm trace lengths supported by CoWoS and other 2.5D packaging

nextplatform.com/2024/03/28/how…

#HPC #AI via Tim Prickett Morgan

'Fortran is not immune to [memory errors], though it is less vulnerable than pointer-based languages.  Fortran does have a restricted form of pointers, but for most applications, allocatable arrays provide a flexible and memory safe alternative.' fortran.uk/isfortranmemor…

Nvidia said AWS will build a server cluster featuring 20,000 GB200 chips that can power up to a 27-trillion parameter model.

For context, this would be more than 15x larger than OpenAI’s GPT-4, which is ~1.7 trillion parameters.

$NVDA $AMZN $MSFT $META

.@Amazon (@awscloud) plans to spend almost $150 billion in the coming 15 years on data centers, giving the #cloud -computing giant the firepower to handle an expected explosion in demand for #AI applications and other digital services

bnnbloomberg.ca/amazon-bets-15…

#HPC via Dan Nystedt

Last week, I described four design patterns for AI agentic workflows that I believe will drive significant progress this year: Reflection, Tool use, Planning and Multi-agent collaboration. Instead of having an LLM generate its final output directly, an agentic workflow prompts…

👇Peer reviews are unwavering in their intricate use of #GPT ! 🤪

#GenAI via Selim Chris Samuel

A minisymposium at #SIAMPP24 focused on cutting-edge fast Fourier transform #algorithms and their parallel implementations. Samar Aseeri recaps the session, which provided a valuable overview of the state of the art, in SIAM News Online. Check it out! sinews.siam.org/Details-Page/n…

🚀🌟🚀Excited to announce Samba-CoE v0.2, which outperforms DBRX by Databricks Mosaic Research and Databricks, Mixtral-8x7B from Mistral AI, and Grok-1 by Grok at a breakneck speed of 330 tokens/s.
These breakthrough speeds were achieved without sacrificing precision and only on 8 sockets,…