Juan Fumero

Java achieving 14 TFLOP/s on an NVIDIA A10? 🚀

Is that possible? Our new article shows a PoC of how Projects Babylon and HAT use Code Reflection to achieve near-native GPU performance for matrix-multiplication.

Deep-dive here:
openjdk.org/projects/bab...

#Java #OpenJDK #GPU #HPC #AI

I can't quite believe that we're now announcing the eighth Programming Language Implementation Summer School (PLISS), but we are --- in May next year, if you want to learn more about programming language implementations, this is the place to come!

The fact that games such as Cyberpunk 2077 on a Pi 5 (along with the 5090) are playable is just impressive.

Heading to Redwood Shores for #JavaOne 2026 this March! 🚀

I’ll be speaking about what we are doing in project Babylon for programming, compiling and accelerating Java workloads on GPUs.

Details: reg.rf.oracle.com/flow/oracle/...

#java #babylon #gpus #ai

How to put Tailscale on a jailbroken Kindle (and why you should) Jailbreaking a Kindle makes it able to do much more. Adding Tailscale takes the Kindle even further.

“It’s a rite of passage to run Tailscale on weird devices": How to get easier access to a jailbroken Kindle, and drop files to it, with Tailscale. tailscale.com/blog/tailsca...

Symbolic Modeling and Transformation of Java Code #JVMLS YouTube video by Java

How code-model and code-reflection works in project Babylon to model Java programs for Machine Learning - presentation by Paul Sandoz
www.youtube.com/watch?v=jbXo...

Writing GPU-Ready AI Models in Pure Java with Babylon Project Babylon introduces the experimental Code Reflection technology that lets you define machine learning logic in plain Java code, without needing Python or external model files. It then uses Foreign Function and Memory (FFM) API to connect your code to native runtimes like ONNX Runtime for fast inference, including GPU acceleration. Furthermore, the Heterogeneous Accelerator Toolkit (HAT) provides a developer-facing programming model for writing and composing compute kernels, which can be more broadly applied-allowing Java libraries to seamlessly harness GPU power for high-performance computing tasks. Presented at Devoxx Belgium 2025 by Ana-Maria Mihalceanu and Lize Raes - Download as a PDF or view online for free

A big thank you to all of you who attended our @devoxx.com talk on writing #GPU ready #AI models in pure #Java!!!

If you wish to revisit the concepts and run the demos, here are the slides:

www.slideshare.net/slideshow/wr...

Give Babylon a try and share feedback with @lizeraes.bsky.social and I!

The Future of AI Development with Paul Sandoz - Live Q&A from Devoxx BE 2026 YouTube video by Java

The Future of AI Development with Paul Sandoz talking about the Babylon project and GPU Compute
www.youtube.com/watch?v=-I1F...

Write your own tiny programming system(s)! - YouTube The goal of this course is to teach how fundamental programming language techniques, algorithms and systems work by writing their miniature versions. The cou...

I'm teaching 𝗪𝗿𝗶𝘁𝗲 𝘆𝗼𝘂𝗿 𝗼𝘄𝗻 𝘁𝗶𝗻𝘆 𝗽𝗿𝗼𝗴𝗿𝗮𝗺𝗺𝗶𝗻𝗴 𝘀𝘆𝘀𝘁𝗲𝗺(𝘀)! again. I'll be posting the videos & tasks on YouTube too.

In the first lecture, I explain what's a tiny system, why write one and show plenty of demos!

🎞️ Playlist: www.youtube.com/playlist?lis...
👉 More info: d3s.mff.cuni.cz/teaching/npr...

Hey #Java devs!

At @devoxx.com, the Java Team at #Oracle is hosting an "Ask The Architects" panel.

Post your questions here by Wed 23:00 UTC! Then check out the recording to hear the experts’ answers (as many as we can).

📝 Please read the questions others have suggested to prevent double-posting.

Java - Visual Studio Marketplace Extension for Visual Studio Code - Java Platform Extension for Visual Studio Code

📢Announcement📢

#Java peers...the #Oracle Java VSCode extension 24.1.0 is now available with support for #JDK25 language features. Download now from the @microsoft.com #VSCode marketplace: marketplace.visualstudio.com/items?itemNa...

New blog post: How to Install NVIDIA Drivers and CUDA Toolkit on Oracle Linux 10 - Desktops

jjfumero.github.io/posts/2025/0...

#ol10 #linux #nvidia #cuda

Today's my last day at @officialuom.bsky.social ! After almost 8 incredible years, it feels like a second home. Huge thanks to everyone for the amazing memories, especially developing #TornadoVM into a world-class open-source platform. So proud of what we achieved! Excited for what's next!

Updating open WebUI to run local models with CUDA

gist: gist.github.com/jjfumero/a30...

link: jjfumero.github.io/posts/2025/0...

New blogpost: How to enable NVIDIA Nsight Compute CLI in Fedora/Linux.
🔗https://jjfumero.github.io/posts/2025/07/04/nvidia-ncu-enable-fedora

#profiling #cuda #configuration

Thanks to @mukel.bsky.social for the llama2.java port and Roger Allen (github.com/rogerallen/l...) for the CUDA port, we were able to understand and shape TornadoVM for running LLMs and transformers architectures.

This is pretty cool: switching between NVIDIA Driver Module Flavors at any time.
docs.nvidia.com/datacenter/t...

Java's 30th Birthday YouTube video by Java

🎉 Java is turning 30! ☕️

On 22 May, join the celebration of three decades of innovation, code, and community!

📺 www.youtube.com/live/LHHPbI7...

🗓️ Checkout the schedule dev.java/community/ja...

Don't miss it! See what's brewing next and be part of Java's joy!

#JavaTurns30 #HappyBirthday

airhacks.fm podcast podcast with adam bien

Really fun and interesting podcast with @adambien.blog. In this one we talk about the status of TornadoVM and all new changes, API design and optimisations we had to do to run LLMs more efficiently on integrated and discrete GPUs.

🎧https://airhacks.fm/#episode_346

Google's hybrid approach to research | Communications of the ACM By closely connecting research and development Google is able to conduct experiments on an unprecedented scale, often resulting in new capabilities for the company.

Really interesting article about Google's hybrid approach to research by combining engineering from day one. Engineering at Google has always intrigued me. Google makes stronger research by proving it at scale. I wish Universities have similar resources.
dl.acm.org/doi/10.1145/...

Guidelines to install and configure @unsloth.ai , a framework for optimising fine tuning of large language models.
This guide details the configuration for Linux systems to run with NVIDIA GPUs.

jjfumero.github.io/posts/2025/0...

#llm #ai #linux

🚀 Learn how to accelerate performance on RISC-V CPUs using TornadoVM & vector instructions with OCK in my latest article. #Java #RISCV #Performance #OpenSource

🔗 jjfumero.github.io/posts/2025/0...

We just released a new version of #TornadoVM: v1.1.0
Highlights:
- New API calls to share and optimise device buffers
- Multi-task graphs fine-grained executions
- Support for FP16 and mixed precision computations
- Memory management improvements
- Extended Math library

and much more!

Preprint is now available:

🔗 research.manchester.ac.uk/en/publicati...

#riscv #java #acceleration #vectorization #tornadovm

Nice. If you want to use a pre-compiled library with an existing JVM, I used it from,
chriswhocodes.com/hsdis/ and builds.shipilev.net/hsdis/

Hoy voy a exponer mi trabajo como postdoc en la Universidad de Manchester y mi trayectoria como investigador en la universidad en la que estudié, La Laguna, Tenerife! Siempre un placer!

NVIDIA DGX Spark A Grace Blackwell AI supercomputer on your desk.

NVIDIA DGX Spark: raspberrypi-ish size with 1000 TOPS AI performance (in FP4)

www.nvidia.com/en-us/produc...

Our last paper is accepted at the #RISCV Summit Europe 2025! We evaluated Java parallel workloads expressed with #TornadoVM on real RISC-V CPUs using vector instructions generated by the oneAPI Construction Kit. How cool is this?

Pre-print coming soon!

Last week @kotselidis.net and myself presented the work we are doing to run and optimise Java/ #TornadoVM applications for RISC-V hardware accelerators in the #RISCV J-Group.

See the slides: github.com/jjfumero/jjf...

If you are interested in the topic, stay tune!