Ali Asif

The AI log, science · agents and coding

Frostbite AI-designed CPU, DAQIRI, Willow contextuality

Architect Labs reports a CPU designed by AI agents, NVIDIA ships DAQIRI for GPU data capture, Google tests Willow, and Dynamo gets Rust tool parsers.

By , published

What shipped

  1. Architect Labs

    Frostbite: A new CPU architecture discovered, designed and verified by AI

    Agent swarms iterated from ISA to final GDS on a 7 nm design, with 800k+ lines of Lean and a reported 4.59x performance per watt against an undisclosed baseline.

  2. NVIDIA

    Enabling Real-Time AI for Scientific Instruments with GPU-Native Data Acquisition

    DAQIRI v2026.11.00 adds NIC packet reordering, an ibverbs engine and direct polling, used for a 500 Gbps electron microscope camera stream.

  3. Google Research

    A game you cannot win unless you have a quantum computer

    Willow beats classical bounds on contextuality games, 98.30% against 88.89% on the magic square game, with a 71-player parity game.

  4. NVIDIA

    Build Tool-Calling Parsers in Rust with AI Agents

    Reusable Rust parsers behind Dynamo tool calling, a shared conformance corpus, and a method for having an agent write a day-0 parser for a new model.

Four items today, all from pages I opened and read. I have not run any of it, so every number below is what the source reports. OpenAI also posted about GPT-6.2 Astra and Sol, but its page still returns a block for me, so I am leaving it out until I can read it directly.

Frostbite, a CPU designed by agents

Architect Labs announced Frostbite on November 18, a general-purpose CPU with a new instruction set that it says was discovered, designed and verified end to end by AI. Agent swarms iterated on the compiler, ISA, architecture, microarchitecture, RTL, verification and final GDS. They also built a specification language called ASL and a system model called Arcmodel, which was recalibrated against RTL until the two agreed within 5%.

The design rests on interaction nets, where computation happens by rewriting local pairs of nodes. Programs are written in Vine and compiled to an IR called Ivy. The ASL contract is formally backed and translates to Lean, with more than 800k lines of Lean, and it checks RTL, UVM, C and SVA outputs.

The headline numbers are against an undisclosed dual-issue, in-order core. Performance per watt is 4.59x as a geometric mean over 25 workloads, 6.02x on the 10 used in the search and 3.83x on the 15 held out. The process is 7 nm, with 800 MHz at synthesis and 500 MHz in the final GDS. The page says the comparison CPU cannot be named for legal reasons, and no repository is linked yet. RTL and ISA release is planned, without a date.

DAQIRI for scientific instruments

NVIDIA released DAQIRI v2026.11.00, a library that moves high-bandwidth sensor data straight into GPU memory. It adds hardware packet reordering on ConnectX-7 or newer NICs, an ibverbs engine for raw Ethernet, and a direct polling mode. On IGX Thor, NVIDIA reports about twice the per-core throughput of the previous DPDK backend for packets up to 1024 bytes above 200 Gbps, and loopback send-to-receive latency of about 2 to 4 microseconds.

The case study is Lawrence Berkeley National Laboratory's Molecular Foundry, which built a high-dynamic-range electron energy-loss spectroscopy camera whose stream exceeds 500 Gbps. GPU code assembles frames, applies corrections and reduces spectra while the experiment runs. The repository is github.com/NVIDIA/daqiri.

Willow and contextuality

Google Quantum AI published a Science paper on tasks where measurement contextuality separates quantum from classical strategies. On the magic square game the best classical strategy wins at most 88.89%, and Willow won 98.30% over 100,000 plays of each version. A contextuality inequality measured 5.590 against a classical limit of 4. A parity game ran with up to 71 players, one qubit each, with entanglement verified across 45 qubits.

The authors also propose these tasks as a benchmark, since the ceiling is set by physics and they surfaced coherent errors that random-circuit benchmarks tend to miss.

Rust parsers for tool calls

NVIDIA describes frontend-crates, Rust components that parse model output into ordered reasoning, visible text and tool-call events for Dynamo. A shared conformance corpus of YAML fixtures covers complete, malformed, truncated and interleaved calls, and the post walks through having an agent draft a parser for a new model while a person reviews the expected outputs. It gives no speed or accuracy figures and tells readers to measure their own.

What I'd watch next

Whether Architect Labs releases the RTL and names its baseline, and whether other groups reproduce the Lean-backed flow. I also want to see DAQIRI numbers from outside NVIDIA.

Sources

Topics: NVIDIA, Google Research, Dynamo