iiinova — where your digital self keeps evolving

Reposted by Stefan Abi-Karam

YosysHQ @yosyshq.com · 29/09/2025

Our next #YUG will be with Matt Young, talking about triple modular redundancy. Join us at 18:00 CET / 22:30 IST / 09:00 PT on Thursday 9th October. Sign up to our mailing list to get a reminder before the event: blog.yosyshq.com/newsletter/

0122

Reposted by Stefan Abi-Karam

Satnam Singh @satnam6502.bsky.social · 29/09/2025

I will work on exploring applications of Aristotle to the formal verification of hardware. This job is a perfect intersection of hardware design and verification, functional programming, formal methods and machine learning, bringing together several threads of my career so far.

1111

Stefan Abi-Karam @stefanabikaram.com · 29/09/2025

"High-Level Synthesis Synthesis" stefanabikaram.com/writing/hls-... Some fun writing on hardware design pedantry

stefanabikaram.com

Stefan Abi-Karam

Personal webiste for Stefan Abi-Karam.

030

Reposted by Stefan Abi-Karam

Catherine @whitequark.org · 15/09/2025

i built a pages server for Codeberg that's way cheaper to operate than the existing one, please enjoy grebedoc.dev/

grebedoc.dev

Grbrmlnd

24510

Reposted by Stefan Abi-Karam

TheZoq2 @thezoq2.bsky.social · 20/09/2025

I got inspired at orconf, so now play.spade-lang.org supports a full in-browser flow for submitting to tinytapeout.com/ Still experimental, so if you try it and run into any issues, let me know :)

074

Reposted by Stefan Abi-Karam

Satnam Singh @satnam6502.bsky.social · 22/09/2025

A lovely Xilinx 7-series FPGA chip layout of a sorting network produced from a dependently typed DSL in Haskell. More stuff like this during my keynote talk at ICFP 2025 in October in Singapore. icfp25.sigplan.org/details/icfp...

2606

Reposted by Stefan Abi-Karam

Satnam Singh @satnam6502.bsky.social · 24/09/2025

I did some more hacking to produce a totally floorplanned 64-input sorter circuit on a Xilinx XC7A200T FPGA, on the LHS picture shown as a tight rectangular block, on the RHS a close-up that shows the butterfly wiring pattern.

381

Stefan Abi-Karam @stefanabikaram.com · 08/09/2025

Shout out to all the CGRA folks, all jokes are in good fun

030

Reposted by Stefan Abi-Karam

Gabe Gomes @gabegomes.bsky.social · 06/09/2025

new preprint, tl;dr: • LLMs match or exceed SOTA strategies on chemical reaction optimizations. • we built the Iron Mind platform and we hope that it can serve as a new benchmark for both reaction optimizers and foundation models. a thread [1/5]

182

Reposted by Stefan Abi-Karam

Rachit Nigam @notypes.bsky.social · 06/12/2024

Lindsey Kuper's* group has produced this fantastic zine on choreographic programming that folks should definitely check out: decomposition.al/blog/2024/12... (* can't seem to find Lindsey here but please tag if you know the handle)

decomposition.al

“Communicating Chorrectly with a Choreography” is out!

0249

Reposted by Stefan Abi-Karam

afd-icl.bsky.social @afd-icl.bsky.social · 02/04/2025

Today at ASPLOS Chengsong Tan will present our work with @wicko3.bsky.social on mechanical formalisation of CXL.cache - check out the paper! www.doc.ic.ac.uk/~afd/papers/... #CXL

153

Reposted by Stefan Abi-Karam

Sung Kim @sungkim.bsky.social · 04/09/2025

This study of deep learning optimizers found that previously claimed speedups of alternative optimizers over AdamW in language model pretraining were often inflated due to methodological shortcomings like unequal hyperparameter tuning and limited evaluation.

2171

Reposted by Stefan Abi-Karam

Sung Kim @sungkim.bsky.social · 04/09/2025

Benchmarking the variety of different proposed LLM optimizers: Muon, AdEMAMix, ... all in the same setting, tuned, with varying model size, batch size, and training duration! They find that AdEMAMix is SOTA, the importance of weight decay, or how Muon “prefers” WSD, and more...

1203

Stefan Abi-Karam @stefanabikaram.com · 04/09/2025

I have yet to read “Computer Architecture: A Quantitative Approach”

220

Reposted by Stefan Abi-Karam

Lenna 💛🩸 @lenna.gay · 02/09/2025

We were cooked the moment operating system UIs stopped shipping looking like this

The internal operating system of the Fujitsu FM TOWNS computer. A few windows are open; one shows the disk drives and one shows the contents of the physical CD-ROM inserted into it.

2387

Reposted by Stefan Abi-Karam

Flux @willflux.bsky.social · 01/09/2025

Over the weekend I released the next part of Isle #FPGA Computer covering bitmap graphics. projectf.io/isle/bitmap-... I've added more docs and tests, but I'm still finding the right approach to the Isle blog and working on diagrams.

Lakritz FPGA dev board driving widescreen monitor showing picture of purple and white crocuses. The monitor is running at 1366x768.

041

Reposted by Stefan Abi-Karam

TheZoq2 @thezoq2.bsky.social · 31/08/2025

This Friday I successfully defended my PhD 🎉 If you want to watch a recording of the defense there is a recording here vimeo.com/1114208379/89dcd4f302

vimeo.com

Improved Tooling for Digital Hardware Development - Frans Skarman

This is "Improved Tooling for Digital Hardware Development - Frans Skarman" by Linköping University on Vimeo, the home for high quality videos and…

0193

Reposted by Stefan Abi-Karam

Jan Gray @jangray.bsky.social · 09/04/2025

Then and now, Prof. Cong and students trailblaze new design automation research for mainstream high level synthesis tools and myriad applications. Chuck Thacker made beautiful computers with FPGAs, so this award seems particularly apt. web.mit.edu/6.173/www/cu... www.cl.cam.ac.uk/~swm11/examp...

041

Reposted by Stefan Abi-Karam

jrenau.bsky.social @jrenau.bsky.social · 06/12/2024

arxiv 📄 Evaluating Large Language Models for Automatic Register Transfer Logic Generation via High-Level Synthesis arxiv.org/abs/2408.02793v1 The ever-growing popularity of large language models (LLMs) has resulted in their increasing adoption for hardware design and verification. Prio...

012

Reposted by Stefan Abi-Karam

arXiv cs.AR Hardware Architecture @csar-bot.bsky.social · 08/07/2025

Zedong Peng, Zeju Li, Mingzhe Gao, Qiang Xu, Chen Zhang, Jieru Zhao: ForgeHLS: A Large-Scale, Open-Source Dataset for High-Level Synthesis arxiv.org/abs/2507.03255 arxiv.org/pdf/2507.03255 arxiv.org/html/2507.03255

012

Reposted by Stefan Abi-Karam

John Wickerson @wicko3.bsky.social · 19/04/2024

Huge congrats to Yann Herklotz (@ymherklotz) for successfully defending his PhD thesis today, all about his proven-in-Coq high-level synthesis tool. And enormous thanks to George Constantinides (@gconstantinides) and Xavier Leroy for their thoughtful and thorough examining.

032

Reposted by Stefan Abi-Karam

John Wickerson @wicko3.bsky.social · 07/05/2021

Here is a short blog post about an #FCCM2021 paper that will be presented next week by @ymherklotz about our work (with Zewei Du and @nadeshr5) on finding bugs in high-level synthesis tools using #fuzzing. johnwickerson.wordpress.com/2021/05…

johnwickerson.wordpress.com

Fuzzing High-Level Synthesis Tools

High-level synthesis – the automatic compilation of a sof...

011

Reposted by Stefan Abi-Karam

arXiv cs.AR Hardware Architecture @csar-bot.bsky.social · 01/04/2025

Martin Langhammer, George A. Constantinides: Banked Memories for Soft SIMT Processors arxiv.org/abs/2503.24132 arxiv.org/pdf/2503.24132 arxiv.org/html/2503.24132

111

Reposted by Stefan Abi-Karam

Keenan Crane @keenancrane.bsky.social · 29/08/2025

I can't* fathom why the top picture, and not the bottom picture, is the standard diagram for an autoencoder. The whole idea of an autoencoder is that you complete a round trip and seek cycle consistency—why lay out the network linearly?

1115925

Reposted by Stefan Abi-Karam

Sung Kim @sungkim.bsky.social · 30/08/2025

"On the Theoretical Limitations of Embedding-based Retrieval" Paper: arxiv.org/abs/2508.21038 Repo: github.com/google-deepm...

github.com

GitHub - google-deepmind/limit: On the Theoretical Limitations of Embedding-Based Retrieval

On the Theoretical Limitations of Embedding-Based Retrieval - google-deepmind/limit

0255

Reposted by Stefan Abi-Karam

arXiv cs.AR Hardware Architecture @csar-bot.bsky.social · 28/08/2025

Rishov Sarkar, Cong Hao: OmniSim: Simulating Hardware with C Speed and RTL Accuracy for High-Level Synthesis Designs arxiv.org/abs/2508.19299 arxiv.org/pdf/2508.19299 arxiv.org/html/2508.19299

012

Reposted by Stefan Abi-Karam

George A. Constantinides @gconstantinides.bsky.social · 28/08/2025

NeuraLUT-Assemble open-sourced! My PhD @martaandronic.bsky.social has been working hard to open-source our SOTA NeuraLUT-Assemble work presented this year at FCCM2025. Extreme low latency inference for FPGA-based machine learning. Try: github.com/MartaAndroni... Read: arxiv.org/abs/2504.00592

github.com

GitHub - MartaAndronic/NeuraLUT: NeuraLUT-Assemble

NeuraLUT-Assemble. Contribute to MartaAndronic/NeuraLUT development by creating an account on GitHub.

043

Reposted by Stefan Abi-Karam

George A. Constantinides @gconstantinides.bsky.social · 28/08/2025

Want to see NeuraLUT-Assemble working and learn all about the latest in LUT-based neural networks and how to classify in nanoseconds? On Monday 1st @martaandronic.bsky.social and Olly Cassidy from @imperialeee.bsky.social will be delivering a tutorial at @ethz.ch ! indico.cern.ch/event/149667...

031

Reposted by Stefan Abi-Karam

Jan Gray @jangray.bsky.social · 27/08/2025

Congratulations @clairexen.bsky.social: #HotChips / IEEE TCMM 2025 Open Source Hardware Contribution Award

Slide from Hot Chips 2025:
"2025 TCMM Open Source Hardware Contribution Award:
Claire Wolf: In recognition of outstanding contributions to RISC-V —
including BitManip, RVFI, and PicoRV32—and to open-source tools like Yosys and IceStorm. [IEEE Computer Society TCMM / Technical Community on Microprocessors and Microcomputers]"

23111

Reposted by Stefan Abi-Karam

Guilherme @noghartt.dev · 12/09/2024

📌 New bookmark! What’s in an e-graph? | Max Bernstein This post follows from several conversations with CF Bolz-Tereick, Philip Zucker, Chris Fallin, and Max Willsey. bernsteinbear.com/blog/whats-in-an-…

bernsteinbear.com

What’s in an e-graph? | Max Bernstein

This post follows from several conversations with CF Bolz-Tereick, Philip Zucker, Chris Fallin, and Max Willsey.

011

Reposted by Stefan Abi-Karam

Mark Riedl @markriedl.bsky.social · 25/08/2025

Our work on Detecting Suspense in Stories will appear at COLM'25 arxiv.org/abs/2508.15794 Do LLMs know when stories are suspenseful? We ran LLMs through a bunch of classical psychology suspense studies The answer: kinda, sorta? Honestly, better than I expected. But the results are nuanced... 1/

38219

Reposted by Stefan Abi-Karam

Mark Riedl @markriedl.bsky.social · 25/08/2025

Suspense is also really common in stories: romcom, thriller, horror, sci-fi, etc. If we ever want really good AI writing assistants, we need to understand their limitations. Also, we know that LLMs are particularly bad at GENERATING suspenseful stories arxiv.org/abs/2402.17119 without help. 7/

arxiv.org

Creating Suspenseful Stories: Iterative Planning with Large Language Models

Automated story generation has been one of the long-standing challenges in NLP. Among all dimensions of stories, suspense is very common in human-written stories but relatively under-explored in AI-ge...

1182

Reposted by Stefan Abi-Karam

Underfox @underfox3.bsky.social · 26/08/2025

In this paper is presented RIROS, a parallel RTL fault simulation framework which integrates structural-level and fault-level parallelism approaches to minimize bubbles that arise from parallel core load imbalances. arxiv.org/pdf/2508.16376

121

Reposted by Stefan Abi-Karam

angela zhou @angelamczhou.bsky.social · 25/08/2025

one thing that I don't get about the new push for RL envs for AI agents is, I feel like there was a lot of genuine joy around RL envs for atari, games, starcraft, overcooked* etc I'm not sure who out there is hankering to create RL envs for accounting, excel, IB monkey (said affectionately)

271

Stefan Abi-Karam @stefanabikaram.com · 25/08/2025

I love how all high-level synthesis tools are not only cursed, but also cursed in their own quirky and unique way

020

Reposted by Stefan Abi-Karam

Jørgen Lund @jaalu.bsky.social · 19/08/2025

Fun breakdown of the process of implementing a toy TPU in Verilog, starting "from scratch" www.tinytpu.com

tinytpu.com

Tiny TPU

An attempt to understand and build a TPU—by complete novices.

055

Reposted by Stefan Abi-Karam

아리다 남편 @parkbot.bsky.social · 24/08/2025

I love this post by Todd Austin equating running a kitchen to a modern out of order microprocessor, especially with the inclusion of precise interrupts, prefetch, and power gating.

Todd Austin • • 1st
S. Jack Hu Collegiate Professor of CSE at UofM...

I am a COMPUTER ARCHITECT, which means I design computers. I was trained at a great school (UW-Madison), and I worked in industry (Intel) before I became a professor (University of Michigan). Despite these rarefied experiences, I first came to know computer architecture in the KITCHEN, working for 4 years first as a COOK. Let me explain.
As a cook, you are the control logic of a complex multi-issue machine, which operates with many of the optimizations found in today's CPUs:
IN-ORDER FETCH: The meal orders come in-order from the wait staff, and this order must be preserved.
OUT-OF-ORDER ISSUE: With a big menu, the latency to cook items will vary greatly, from fast salads to medium-latency burgers to long-latency steaks. Cooks will scan the upcoming orders and "issue" long latency items first, medium later, and short last, to make sure all items finish at roughly the same time.

IN-ORDER RETIREMENT: Importantly, the meal orders have to finish IN-ORDER. Wait staff will get very annoyed if any out-of-order completion occurs—no one wants to see the table that sat down after them get their food first! BTW, Tapas restaurants break this rule, which greatly improves the efficiency of their kitchens. What is the computer architecture dual to the tapas restaurant?
PREFETCH: Occasionally, someone from the wait staff will run into the kitchen and yell "BUS!" (typically with the same vigor as someone running into a beach bar and yelling "TSUNAMI!"), which means 50+ additional orders are coming to your line.
When this happens, you turn on all the stoves and steak burners and start frying staples, etc.

CACHING and SPECULATION: Many of the "specials" for the night are pre-cooked and kept in an oven (cache). This is optimizing the common case, since these popular meals take seconds to make. Often, if there are specials left over at the end of the night (mispeculation), they are donated to local food shelters.
NON-DETERMINISTIC LATENCY: Efficient architectures try to avoid all non-deterministic latencies, since they stymie schedulers and gum up the works. Kitchens operate the same way: everything is prepped to give deterministic cook times.
Once, I put a steak on a grill shared with the short-order cook window. Later, when I checked on my steak, it was gone... I confronted the short-order cook, and he said, "Do you want to fight me for it?" A non-deterministic latency made my line a mess that night!

PRECISE INTERRUPTS: When the shift changes, the incoming cook will negotiate which order to transition on. The working cook will finish all their orders, after which the new cook takes over. This is called a precise interrupt, and it is similar to how out-of-order processors effectuate a process switch.
POWER GATING: Kitchens are HOT. If stoves are not needed, they are powered down!
Besides architecture training, working as a cook is a great way to start a career. All later jobs you will have will be MUCH EASIER and LESS DANGEROUS!
#computerarchitecture

042

Reposted by Stefan Abi-Karam

SMBC Comics @smbccomics.bsky.social · 23/08/2025

The Stewart's Calculus movie was 4,000 hours long but it's okay because nobody ever gets to the final third anyway. COMIC ◆ www.smbc-comics.com/comic/flux PATREON ◆ www.patreon.com/c/ZachWeiner... STORE ◆ smbc-store.myshopify.com

29612

Reposted by Stefan Abi-Karam

Sung Kim @sungkim.bsky.social · 23/08/2025

Hong Kong University updated AI-Researcher: Agentic Scientific Discovery with a new technical report. GitHub: github.com/HKUDS/AI-Res... Paper: arxiv.org/abs/2505.18705

071

Reposted by Stefan Abi-Karam

Egor Kotov 🌐🏃‍♂️🚊🚋🚙 @ekotov.pro · 23/08/2025

Many of you already heard about this new algo that's faster than Dijkstra ( arxiv.org/abs/2504.17033 ), here I collected 3 software implementations: 1. #python github.com/sidharthpuna... 2. #rust github.com/alphastrata/... 3. #rust + some explanations of the paper medium.com/%40adnanmaso...

medium.com

Breaking the Sorting Barrier: Why a New Shortest‑Path Algorithm Matters (Even If You Still Love…

A Deterministic O(m·log^(2/3) n) Breakthrough in Directed Single-Source Shortest Paths — Escaping the Sorting Barrier with Bounded…

195

Reposted by Stefan Abi-Karam

arxiv cs.CL @arxiv-cs-cl.bsky.social · 19/08/2025

David Heineman, Valentin Hofmann, Ian Magnusson, Yuling Gu, Noah A. Smith, Hannaneh Hajishirzi, Kyle Lo, Jesse Dodge Signal and Noise: A Framework for Reducing Uncertainty in Language Model Evaluation arxiv.org/abs/2508.13144

041

Reposted by Stefan Abi-Karam

Mark Riedl @markriedl.bsky.social · 16/08/2025

Approximately 30% of answers in the so-called “Humanity’s Last Exam” could be wrong www.futurehouse.org/research-ann...

futurehouse.org

About 30% of Humanity’s Last Exam chemistry/biology answers are likely wrong | FutureHouse

2306