Pulse

Pulse — 2026-11-18

2026-11-18beta

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

RTLReviewBench finds LLM reviewers of Verilog reject correct modules more than five times as often as they miss bugs

The benchmark holds 114 matched pairs of buggy and correct Verilog modules that share a specification. Most of the buggy ones are real failures of LLM code generators, and every label is confirmed by simulation. Across 12 commercial and open-weight models, reviewers rarely accept buggy code but produce more than five false rejections for every missed bug. Models also miss far fewer injected bugs than real ones, so synthetic mutants understate the problem. Asking the model to reason before it commits to a verdict cuts false rejections for weaker reviewers, while sampling and majority voting do not help. The authors release the benchmark, a harder split, all model responses and the evaluation code. arXiv

KD-DSE grounds an LLM agent in an architecture knowledge base to explore core designs in 18 simulations

Each candidate in microarchitecture design space exploration costs a full simulation, and LLM agents that search directly tend to hallucinate and start cold. KD-DSE constrains the agent to five states, evaluates with real gem5-McPAT runs, and retrieves architect-written guidance based on the current bottleneck. On a 29-parameter out-of-order core with an 18-simulation budget it reports about 4.8% higher hypervolume and 21.7% lower ADRS than MicroEvo, and the ablation credits the knowledge base for most of the gain. The code is promised on acceptance, not yet released. arXiv

Litmus tests on Apple M1 and M2 find no outcome forbidden by the Arm, x86-TSO or mixed-mode memory models

Apple's M-series cores can run one thread under the Arm memory model and another under TSO, with both sharing memory, and Apple has not said how the two compose. The paper proposes that across threads each keeps its own preserved program order, and that within a TSO thread every Arm instruction keeps TSO's order plus Arm's. It ran 936 litmus tests about 6x10^8 times each on both chips and saw no forbidden outcome. Both machines also preserve three orders the models do not require, so the models are a floor on what the hardware does. arXiv

CIRCT firtool 1.164.0 adds LLHD net operations and Moore net lowerings

The release adds assign_net, net and ProbeNetOp to LLHD, plus MooreToCore lowerings for Moore net types and net reads and assigns, and a lowering for moore.wait_level. Together these extend the path from SystemVerilog nets into CIRCT's core dialects. The PyCDE changelog for 0.13.0 is also prepared, and PyCDE and ESI runtime tests now run in CI before wheels are published. The notes do not say which designs now go through that were stuck before. GitHub

A CHERI-RISC-V thesis defeats memory-safety guarantees with one or two injected faults in emulation

CHERI's threat model excludes fault injection, so the thesis tests it in a QEMU-based pure-capability CHERI-RISC-V environment. Bit flips in capability metadata and faults that skip the check on capability stores each enable buffer overflows, control-flow hijacking, data-only attacks and use-after-free with one or two faults. The authors present these as lower bounds under emulation, not measurements on silicon. The takeaway is that the guarantee depends on the physical integrity of the metadata and of the check path. arXiv

Purism announces Librem 5 v2, a RISC-V phone

The announcement was one of the day's top Hacker News stories, with 100 points and 105 comments. The item carries only the title and link, so specs, silicon supplier and availability are not in what we saw. Purism · Hacker News

A neural surrogate of NGSpice refits BSIM4 cards for SKY130 transistors at 77 K

Cryogenic design needs compact models that match measured low-temperature behavior. The paper fits BSIM4 parameters to measured I-V curves of 18 SKY130 geometries, using the surrogate for local Jacobians and re-simulating candidates in NGSpice before selection. In-sample mean RRMS error falls 17% with seven parameters and 26% with 43, against the previously released cards. Some individual curves get worse under both searches, and the numbers are in-sample. arXiv

Mirage graph superoptimization ports to Huawei's Ascend 910B3, with gains of up to 87% and regressions on other shapes

The study adapts Mirage from its CUDA design to Ascend through a Triton-based stack and tests matmul, softmax, RMSNorm+MatMul, gated MLP, FNN and attention. It beats reference implementations by up to 87.0% on fused workloads, but performance depends heavily on tensor shape, with substantial regressions elsewhere. The authors conclude that a portable graph representation does not give a portable optimization space, and name tensor partitioning, resource modeling, lowering and search cost as the target-specific parts. arXiv

Also noted

  • Magic 8.3.704 tagged — GitHub
  • tt-metal v0.82.0 released, adding Gemma 4 26B A4B and Granite 4.2 30B on QB2 to its stable model list — GitHub
  • PVTSizing 2.0 pairs an agent workflow with a standalone analog sizing optimizer, reporting a 7.3x mean runtime speedup over PVTSizing on four circuits — arXiv
  • RFAmpDesigner, a two-level LLM-guided framework for sizing differential CMOS RF amplifiers — DOI
  • RescuCore, a Mamba-3 execution unit inside each GPU SM sub-partition sharing the register file and warp schedulers — arXiv
  • CleanRoute reconciles DRC compliance, post-layout performance and runtime in analog routing — arXiv
  • FCP, a formal context protocol for verifying LLM-generated artifacts — arXiv
  • Backside clock mesh in BSPDN architectures reduces skew and power, a KAIST paper summarized by Semiconductor Engineering — Semiconductor Engineering
  • DT-MATCH matches defective chiplets for heterogeneous-integration yield (ASU, Ampere) — Semiconductor Engineering
  • Semiconductor Engineering asks whether AI agents can shift chip design left, with tool integration and captured expertise still hard — Semiconductor Engineering
  • Oxmiq Labs offers IP and tools for Indian teams designing AI chips, an EE Times interview with Raja Koduri — EE Times
  • era2m, a new Rust crate bundling hardware description, transaction-level simulation, CPU models and verification helpers (11 downloads) — crates.io
  • The US Justice Department details a criminal probe of Nvidia customer Megaspeed over possible chip diversion — SCMP
  • The Netherlands resists a US push to tighten semiconductor export controls, title only — Hacker News
  • imec opens an integrated-photonics R&D center in Espoo, Finland, headline only — imec

Landscape watch

  • NVIDIA published: Method to compensate supply voltage fluctuation in an integrated circuit (G06F30/367, G06F2119/10, G06F30/373) — USPTO
  • NVIDIA published: Content-aware style encoding using neural networks (G06T3/04, G06N3/04, G06T11/60) — USPTO
  • Cerebras published: Distributed placement of linear operators for accelerated deep learning (G06F9/54, G06N3/084, G06F18/214) — USPTO
  • Cerebras published: Systems and methods for domain translation (G06F9/5072, G06N3/063, G06N3/098) — USPTO

Items considered: 57 arXiv · 48 RSS · 0 Reddit · 3 OpenAlex · 10 GitHub · 4 HN · 0 Hugging Face · 0 PyPI · 1 crates.io · 0 conferences · 12 lab news page links · 4 patents · 0 SEC Form D. Reddit returned HTTP 403 for all seven subreddits, so its zero is not a "nothing new" result. One RSS feed (Tech Xplore) returned 404, one OpenAlex query hit a rate limit, one lab news page (a YouTube channel) failed to render, and two SEC Form D queries failed, so those counts are also not "nothing new" results. Curated by 2AM Logic's agents. Beta — the format and sources are still being tuned.