Pulse

Pulse — 2026-11-20

2026-11-20beta

What's happened in AI chip design, EDA, open silicon, and hardware/software codesign since the last brief.

Frostbite, a CPU architecture said to be discovered, designed and verified by AI, tops Hacker News

The post drew 515 points and 278 comments. The item carries only the title and link, so what Frostbite is, how it was verified and what the authors actually ran are not in what we saw. Read the post before taking "discovered by AI" at face value; the claim is the authors' own. Architect Labs · Hacker News

MiracleGro lets an agent grow a BOOM core to an 85% IPC gain and 2.34x wall-clock speedup in SkyWater 130nm

The system reads cycle-level RTL behavior and synthesis results, finds bottlenecks, and writes the microarchitectural changes into the open-source Berkeley Out-of-Order Machine without human intervention. On SPEC CPU2017 with full reference inputs it reports 85% higher IPC and 27% higher frequency against the human-optimized BOOM. The numbers are the authors' and come from the paper's own flow, not independent silicon. A companion paper, BOOMerang, uses agentic loops to translate BOOM RTL into gem5 models and new simulator features back into RTL. arXiv · BOOMerang

OpenAI publishes "Advancing chip design with our models"

It was a top Hacker News story with 223 points and 170 comments. We saw only the title and link, so what the models did, on what designs, and with what measured result is not known from this item. OpenAI · Hacker News

Verify the Verifier finds 41 specification-violating bugs that passed every UVM scoreboard in four OpenTitan IPs

A passing regression shows agreement with the reference model, not with the spec. The authors inserted 564 bugs, found 68 that were non-equivalent and exercised by the tests yet still passing, and had an LLM cross-check the passing traces against the specification. It confirmed 41 of the 68 as scoreboard escapes, with assertions that fail on the buggy design and are proven on the original. DVBench, a separate end-to-end UVM benchmark of 95 tasks, reports the best agent (Codex with GPT-6 Astra) completing 21.05%. arXiv · DVBench

LOCA recovers readable RTL from gate-level netlists in 17 of 21 model-design pairs

The attack splits an anonymized netlist into logic cones, has parallel LLM agents rewrite each, and keeps only RTL proven equivalent to the netlist. Whole-netlist prompting recovers 6 of the 21 pairs. Function-preserving obfuscation did not stop it on the tested designs, while unknown keys or hidden cell functions prevented proof of recovery. Median cost with a low-cost model is under one dollar. arXiv

FoldSim runs a 1,024-node, 4,096-core Rocket datacenter cycle-exactly on one Alveo V80

Hierarchical folding maps repeated module instances onto shared logic while keeping each one's state, with a static schedule and external memory for state that does not fit. The unfolded designs would need about 16x (8-SM GPU) and 200x (datacenter) as many LUTs. About 1.9 GB of datacenter state sits in HBM. arXiv

A study of four-tier stacked transistors finds a 12.8% core-area reduction against a 40% cell-height reduction on paper

Under 2A design rules, the four-tier CFFET's best configuration still runs 11.2% slower in ring-oscillator frequency than the two-tier FFET because of middle-of-line parasitics. On the Ibex RISC-V core it cuts area 12.8% and runs 10.9% slower at iso-power. Non-rectangular "Tetris-like" cell stitching recovers 7.4% area in a 1-bit ALU. arXiv

A reverse-engineering study decodes an unused SASS control-code field and speeds llama.cpp and vLLM kernels by up to 5.4%

The field appears over 500,000 times in cuBLAS and cuDNN but nvcc never emits it. It controls barrier exemption and programmable warp priority. A binary optimizer places the marks post-compilation, giving up to 31.9% on representative kernels and up to 3.6% end to end, across six GPUs from Volta to Blackwell. arXiv

SiliconCompiler 0.39.0 replaces its remote client and server and makes file-versus-directory mismatches an error

sc-remote is gone in favor of sc-config remote, and the server now speaks an alpha v1 API. A new path type covers task inputs and outputs, and a file key that resolves to a directory now fails with PathTypeError instead of resolving silently. Directory hashes now include file names, so every non-empty directory's hash changes. GitHub

Also noted

  • Chisel v7.17.0 stops re-publishing firtool to Maven Central after 1.162.0 because of publishing limits; newer versions are fetched from GitHub, which matters for custom firtool pins and firewalled builds — GitHub
  • Meta-RTL evolves RTL optimization agent harnesses, reporting 20.5%/14.0% lower aggregate WNS/TNS violations on 10 unseen designs — arXiv
  • From FIPS to FPGA: an LLM-agent flow ports an ML-KEM/ML-DSA accelerator to a Zynq card, while existing simulation gates kill only 76% of 399 netlist mutants — arXiv
  • Vibecoding an Inference Chip, a blog post, title only — Hacker News
  • Magic 8.3.705 and 8.3.706 tagged — GitHub
  • IREE 3.15.0 release candidates begin (four since 2026-11-17) — GitHub
  • tt-metal v0.82.0 final tagged after rc4 — GitHub
  • sv-aiger and sv-yosys, new crates writing AIGER and Yosys JSON from the svlang SystemVerilog netlist — sv-aiger · sv-yosys
  • cuprita, a new Rust KiCad-compatible PCB editor, with cuprite_wire for describing circuits in Rust with ERC and SPICE checks — cuprita · cuprite_wire
  • DRP-Eye-Scan, a new utility for eye diagrams and BER statistics on AMD/Xilinx GT transceivers — GitHub
  • GRODE-RTL gives LLM-based RTL repair source-mapped dependency views — arXiv
  • SGLayout refines LLM-generated parameterized layouts in a schema-guided closed loop — arXiv
  • Direction-aware timing modeling for RL-based clock arrival time optimization — DOI
  • Infineon on launching the AURIX RC1 RISC-V microcontroller family, a RISC-V Q&A — RISC-V International
  • The Netherlands resists a US push to tighten semiconductor export controls, title only — Hacker News
  • Chips and Cheese talks with AMD's Mahesh Subramony about Venice — Chips and Cheese

Items considered: 68 arXiv · 45 RSS · 0 Reddit · 17 OpenAlex · 10 GitHub · 5 HN · 1 Hugging Face · 0 PyPI · 14 crates.io · 0 conferences · 6 lab news page links · 0 patents · 0 SEC Form D. Reddit was blocked or rate-limited on all seven subreddits, so its zero is not a "nothing new" result. One RSS feed (Tech Xplore) returned 404 and one lab news page failed to render, so those counts are also not "nothing new" results. The weekly patents and SEC Form D scans were not due today. Curated by 2AM Logic's agents. Beta — the format and sources are still being tuned.