Mark Pupilli

[Matt Cohen @mattybcohen]

How do you scale 500 to 9,000 H100s in six hours!!

That's what it took to keep @DreamLabLA Dream Machine standing when it went viral. Half a million videos in 12 hours.

@keeganmccallum3 is now building @urunml for the interactive AI wave.

New Tank Talks [links below]

[Nicholas Fry @deepfry_n]

What if we had converged to using different hardware architectures for visual computing ?

Check out how we render 3D Gaussians on a Coarse-Grained Reconfigurable Array (CGRA): https://t.co/bbSuQi0f9F
Presented at EGSR 2026.

Work with @alzugarayign, @MarkPupilli, @paulhjkelly and @AjdDavison

[Jeremy Bernstein @jxbz]

Today I gave a talk @graphcoreai called "Neurons, Norms and Number Systems", about precision and scale in deep learning. I'm sharing my slides below...

My talk gave a very visual "neural circuits" perspective on Modula. We formally define two ways of building modules:

(1/6)

[PyTorch @PyTorch]

Welcome to the PyTorch Foundation, Graphcore! 🎉

Graphcore has contributed to the PyTorch ecosystem by developing integrations to run on their IPU hardware. We look forward to their involvement in our community as the first general member!

Details: https://t.co/ia8d3bmaTO

[Mark Pupilli @MarkPupilli]

@alignment_lab @MattNiessner What they up to? Rediscovering Kalman filtering and typesetting it badly?

[Mark Pupilli @MarkPupilli]

@BlackHC @MattNiessner Oh dear:

[Graphcore @graphcoreai]

Graphium has arrived. A powerful new way to build and train Graph Neural Networks for molecular science. @graphcoreai @valence_ai @Mila_Quebec @datamol_io
https://t.co/ym00swKB7j

[Tom 7 @tom7]

One time 20 years ago a computer science professor said something wrong (a technicality) in class, so I spent quite a lot of time proving them wrong in this video, and then just kept going, as if I don't have any self-control over my approach to projects:
https://t.co/IQmHLlW30L

[Graphcore @graphcoreai]

Do you need #LLMs like #GPT-4 and #GPT-3 for every language task? No... fine-tuning open source #GPT-J may be a more cost-effective alternative.
https://t.co/TQqpcqEJgW

[Mark Pupilli @MarkPupilli]

@AjdDavison @fdellaert @markvanderwilk Congratulations @joeaortiz !

[Mark Pupilli @MarkPupilli]

Path-tracing on IPU version 2: https://t.co/w3qd5i11e5

If you want to learn how to implement this kind of thing on Graphcore hardware apply for our in-person workshop in London next month: https://t.co/6JYSH0XhRZ

[Mark Pupilli @MarkPupilli]

Some more exploratory work I have been doing with IPUs:
https://t.co/esOZYOaljS

[Graphcore @graphcoreai]

We're thrilled to be partnering with @HelloPaperspace to offer developers free and instant IPU access! In just a few seconds, you can get state-of-the-art models up and running and experience the advantages of Graphcore's made-for-ML technology. Details: https://t.co/8XMfsA7uIh

[Graphcore @graphcoreai]

Good news, developers: the latest @wandb client update includes IPU profiling support! Metrics including utilisation, clock speed, temperature and power usage are now automatically logged and synced to the System section of the dashboard.

[Graphcore @graphcoreai]

It's never been faster or easier to extract insights from text data with the new cloud solution from Pienso and Graphcore. With no coding or AI skills required, Pienso + Graphcore puts control in the hands of business leaders and subject-matter experts.
https://t.co/in2okJdYAg

[Michael Bronstein @mmbronstein]

Are GPUs the best hardware choice for GNNs? In a new post authored with @emaros96 @Daniels_Data we explore the new @graphcoreai IPU architecture for Temporal GNNs showing it produces 10x gains in performance

https://t.co/qwHblByY1l

[Graphcore @graphcoreai]

Say hello to Bow: @graphcoreai's new range of AI compute systems powered by a brand new IPU - the world's first Wafer-on-Wafer processor. https://t.co/2Q3yX1Pdpc

[Graphcore @graphcoreai]

The next step in the evolution of AI: @graphcoreai's plan for an Ultra Intelligence AI Supercomputer.
https://t.co/hDhOnV4I6c

[Mark Pupilli @MarkPupilli]

@BumblingBee8 Might port Doom if I get some spare time though 🤣

[Mark Pupilli @MarkPupilli]

Exploiting Tile-Level Parallelism on Graphcore IPUs https://t.co/qibhWaTalL

[Mark Pupilli @MarkPupilli]

@BumblingBee8 Possible in principle but there is no dedicated rasterization HW or "video out" so would have to operate as a cloud gaming device like Google's Stadia. Not what it was designed for really!