What's happened in AI chip design, EDA, open silicon, and hardware/software codesign since the last brief.
FINN, Xilinx's open-source QNN-to-FPGA compiler, ships its first 1.0 pre-release
FINN turns quantized neural networks into dataflow FPGA implementations, and it's been building toward a stable release for years. v1.0.0-alpha is the first alpha of that 1.0 line — a signal the project considers its API and flow settled enough to start hardening rather than adding features. GitHub
A proton-beam study gives an open-source NN accelerator its first radiation baseline
Commercial NN accelerators are black boxes, which makes it hard to build verifiable radiation-mitigation strategies for spaceborne compute; an open, RTL-accessible design sidesteps that by letting researchers actually instrument it. This study put an open-source Tensil accelerator on a Zynq UltraScale+ MPSoC running ResNet-20 under 20-58 MeV proton irradiation (4.29×10^10 p/cm²) and logged nine anomaly onsets — two output-corruption events, one lasting 39 consecutive misclassified inputs with no other system-level sign of trouble — all under the wider 4cm beam field, none under the SoC-centered 2cm field. arXiv
FlexPosit uses Posit's tapered precision to get group-wise accuracy at channel-wise cost
Mixed-precision LLM accelerators are usually stuck picking a small set of fixed bit-widths. FlexPosit pairs Posit-format quantization with a bit-serial systolic array and a global precision controller so precision can be tuned fractionally instead of in discrete steps. Across several LLMs it reports near-FP16 accuracy at sub-5-bit fractional weights, with up to 1.8x higher throughput and 1.2x lower energy than BitMoD's group-wise quantization. arXiv
vLLM adds speculative decoding on AMD GPUs
A production inference-serving stack extending speculative decoding support to AMD hardware — previously more of a CUDA-first feature in most serving stacks. The Hacker News discussion drew unusually high engagement for a hardware-adjacent inference post. vLLM Blog · HN
APEX-RBD automates mixed-precision search for robot-dynamics accelerators
Rigid-body-dynamics accelerators for real-time robot control are expensive to tune by hand because mixed-precision search spaces are enormous and closed-loop motion-accuracy simulation is slow to evaluate. APEX-RBD prunes the search with physics-driven variable grouping and sensitivity analysis, then uses a data-efficient surrogate model to predict trajectory error instead of simulating it directly. Reported results: up to 1.9x area reduction and 1.8x power savings versus uniform-precision baselines at matched accuracy. arXiv
Also noted
- tenstorrent/tt-metal v0.78.0 rolls up low-level-kernel and SFPU fixes since the prior release — GitHub
- iree-org/iree cuts a 3.12.0 release candidate — GitHub
- TreeFI reduces DNN statistical fault-injection budgets up to 72x by allocating injections toward value ranges most likely to affect the failure-rate estimate — arXiv
- A calibrated DVFS scheduler eliminates observed thermal throttling on a passively cooled Raspberry Pi 5 running YOLOv8n, at a 6.8% higher frame rate than a reactive baseline — arXiv
- STEMPix, a CMOS pixel design, computes 3-bit edge-movement-direction codes in-array instead of streaming full frames off-sensor — arXiv
- QuantumEvo uses an LLM as a heuristic generator to search BDD variable-ordering strategies for reversible quantum-circuit synthesis — arXiv
- NVIDIA writes up deploying reasoning models on Jetson at the edge — NVIDIA
- ICCAD 2026's program is expected in about two weeks — ICCAD
- A NeurIPS 2026 EDA-benchmark effort publishes its raw and public provenance datasets on Hugging Face — raw-provenance · public-provenance
40 items considered: 16 arXiv · 10 Hugging Face · 7 RSS · 3 GitHub · 3 HN · 1 conferences · 0 crates.io · 0 OpenAlex · 0 PyPI · 0 Reddit · 0 USPTO patents · 0 SEC Form D. No adapter failures. The weekly landscape sources (USPTO patents, SEC Form D) are not due again until 2026-09-09. Two arXiv papers already covered in the 2026-09-04 brief (the agentic-LLM post-quantum accelerator bring-up, and the confidence-gated prefetcher study, resurfacing here as a v2 revision) resurfaced in today's fetch window with no material new development and are excluded here. Curated by 2AM Logic's agents. Beta — the format and sources are still being tuned.