We propose a method to accelerate the Softmax function, a key bottleneck in Transformer models, by integrating a custom exponential instruction into Snitch. See "VEXP: A Low-Cost RISC-V ISA Extension for Accelerated Softmax Computation in Transformers" https://t.co/yFVRAqSZ0U

Reporting from the Session Radiation and Flight Results at #RISCV in Space in Gothenburg, here is Frank with his keynote "Open RISC-V Hardware in Space – One Small Step for Space, One Giant Leap for PULP"! Thank your for the pics @MikeRogenmoser.

The slides & poster from Andrea's work on a technique to accelerate the softmax kernel in transformers "SoftEx: a Low Power and Flexible Softmax Accelerator with Fast Approximate Exponentiation" presented at #DATE2025 are online: https://t.co/jvvABSM4Y1 & https://t.co/9MfATg4R6l

We are delighted to see a mention of our Basilisk, which uses the IHP 130nm Open PDK, in Giovanni De Micheli's #DATE2025 keynote "A Vision of Systems and Technology in a Connected Europe ". More on Basilisk: https://t.co/WFq7dTGM0j.

#DATE2025 starts today in Lyon. The PULP team will present their work at the times in the chart below. Looking forward to meeting you, all!

#Carfield is here! The first open-source, silicon-proven #RISCV platform for secure, reliable, time-predictable hardware-accelerated mixed-criticality applications" based on the Cheshire architecture https://t.co/7BbAGwQuZF. Explore also on GitHub https://t.co/Jo71XUbAJR @Ang__93

Pulp Cluster v3.0.0 is here! Native support for accelerators Neureka, RedMulE, SoftEx. Enhanced reliability: Error correction support for SRAMs, Fault-tolerant interconnect, Integration of HMR unit. Improved parametrization for greater flexibility. See https://t.co/C5AxfZn6Mr

We present Maestro, a SoC bringing Wearable Ultrasound processing to the edge without offloading computation to remote PCs. See "Maestro: A 302 GFLOPS/W and 19.8GFLOPS RISC-V Vector-Tensor Architecture for Wearable Ultrasound Edge Computing" https://t.co/PDMv5mpEU6 @mattia_siniga

Here is our work about Maestro, an application-specific RISC-V processor for wearable ultrasound edge computing. The unified Vector-Tensor Unit (VTU) and the memory-coupled multi-precision FP16/FP32 FFT accelerator boost the WUS application by 5x. Check it out for more in-depth!

This picture was from 2019! I had an opportunity to spend a day listening to the works of graduate students focused on AI at SJTU in Shanghai, one of the top engineering schools in the world. Upon return to US, I wrote an internal report that said if there was no geo-political

Next for #DATE2025 in Lyon, @chaoqun_L will present a versatile Transaction Monitoring Unit for AXI4-based SoCs. Find the paper on arXiv under "Towards Reliable Systems: A Scalable Approach to AXI4 Transaction Monitoring" https://t.co/7ZaxpkN2ec @ThommyThomaso @aottaviano96

I'll get straight to the point. We trained 2 new models. Like BERT, but modern. ModernBERT. Not some hypey GenAI thing, but a proper workhorse model, for retrieval, classification, etc. Real practical stuff. It's much faster, more accurate, longer context, and more useful. 🧵

Today OpenAI announced o3, its next-gen reasoning model. We've worked with OpenAI to test it on ARC-AGI, and we believe it represents a significant breakthrough in getting AI to adapt to novel tasks. It scores 75.7% on the semi-private eval in low-compute mode (for $20 per task

In our recent paper we present a partitioning scheme tailored for deploying Transformer models on edge devices. Check out "Distributed Inference with Minimal Off-Chip Traffic for Transformers on Low-Power MCUs" https://t.co/FF7spiMS1D

For all of you interested in Edge AI here is a recording of @Vic_JB_Jung's talk "Energy-Efficient Gen AI on Open-Source RISC-V Heterogenous SoCs" by EDGE AI Foundation. Deeploy, our bottom-up DNN compiler for heterogeneous MCUs & Siracusa inside:

We present "A Flexible Template for Edge Generative AI with High-Accuracy Accelerated Softmax & GELU" based on a 8-core RISC-V cluster with a 24 × 8-PEs tensor processing unit and a novel accelerator for BFloat16 softmax. See https://t.co/yyEQRP7RUE @YvanTortorella @Ang__93

The (true) story of development and inspiration behind the "attention" operator, the one in "Attention is All you Need" that introduced the Transformer. From personal email correspondence with the author @DBahdanau ~2 years ago, published here and now (with permission) following

Ben 166 GW di solare ed eolico installati in Germania, nelle prime 2 settimane di Novembre, con poco vento e pochissimo sole, hanno generato appena il 18,5% dei 18,4 TWh della domanda. Il resto lignite, gas, carbone e legna, a più non posso, e per fortuna anche un po' di

Wondering what is up in the Italian PULP section? PULP Transformer workshop event! Great pleasure to meet everyone planning our group's next goals relating to transformer acceleration. And of course, a nice break with some typical Bolognese food🥩😀. @arunax

We have recently taped out ASTRAL, a flexible, high-performance, secure & reliable heterogeneous SoC based on PULP, targeting space applications🚀. Learn more: https://t.co/EEpxQLRQdu #RISCV @YvanTortorella @CianiMaicol @MikeRogenmoser @chaoqun_L @Ang__93 @arunax