Mark Pupilli retweeted
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]
Mark Pupilli retweeted
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): nmjfry.github.io/ipu-3dgs/
Presented at EGSR 2026.
Work with @alzugarayign, @MarkPupilli, @paulhjkelly and @AjdDavison
Mark Pupilli retweeted
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)
Mark Pupilli retweeted
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: hubs.la/Q020Yj8x0
Mark Pupilli retweeted
Graphium has arrived. A powerful new way to build and train Graph Neural Networks for molecular science. @graphcoreai @valence_ai @Mila_Quebec @datamol_io
hubs.la/Q01WKBkn0
Mark Pupilli retweeted
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:
youtu.be/Ae9EKCyI1xU
Mark Pupilli retweeted
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.
hubs.la/Q01J4g3H0
Path-tracing on IPU version 2: github.com/markp-gc/ipu_ray_…
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: graphcore.ai/accelerated-com…
Mark Pupilli retweeted
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: hubs.la/Q01jJjgN0
Mark Pupilli retweeted
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.
💥 wandb 0.13.1 released 💥
2️⃣ new contributors!
💫 Improved multiprocessing support‼️
💫 Save KerasCallback model checkpoints as artifacts
💫 Profiling support for @graphcoreai IPUs
🪲 Prevent run.log() from mutating arguments
👉`pip install wandb --upgrade`
Mark Pupilli retweeted
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.
hubs.la/Q01jyWK70
Mark Pupilli retweeted
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
michael-bronstein.medium.com…
Mark Pupilli retweeted
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. hubs.la/Q015fWMC0
Mark Pupilli retweeted
The next step in the evolution of AI: @graphcoreai's plan for an Ultra Intelligence AI Supercomputer.
hubs.la/Q015fXtr0
Mark Pupilli retweeted
Graphcore demonstrates outstanding performance at increasing scale in second #MLPerf submission, with IPU-POD16 outperforming Nvidia DGX A100 on ResNet-50.
hubs.la/Q0104Wy70
Mark Pupilli retweeted
Graphcore raises $222 million in Series E Funding Round. hubs.ly/H0D5d8K0
Mark Pupilli retweeted
Graphcore's IPU delivers outstanding performance on state-of-the-art models like EfficientNet. Next generation AI demands a new architecture. Read more: hubs.ly/H0Chh6m0
