Architect Labs published a paper on arXiv stating that its Architect Labs Platform system designed the Redwood accelerator with virtually no human involvement: after the specification was set by two human architects, no one controlled the pipeline work at levels below the specification, and the entire cycle took less than two weeks. Redwood is aimed at single-threaded, low-power inference with ultra-low latency for “physical AI,” i.e., robots and drones. The viability of the approach is demonstrated on the Redwood Nano FPGA prototype, which runs models with billions of parameters. The authors position the result as the first production accelerator designed by AI from start to finish.

image

What happened

The paper was submitted to arXiv on August 26, 2026, and updated on August 28. The Architect Labs Platform (ALP) pipeline received a specification from humans as input, after which the system operated autonomously, without human intervention below the specification level: it generated a performance model, RTL code at the register transfer level, UVM environments for verification, formal proofs of correctness, firmware, and compute cores, including a FlashAttention implementation. All blocks closed 95% test coverage. Specification changes underwent re-verification and were deployed to hardware in less than 48 hours, and no errors were found on the first RTL reset to FPGA. The only measured result pertains to the Redwood Nano FPGA variant on an AMD Versal VPK180 board: the Qwen3-0.6B model delivers 13 peak and 12.1 average tokens per second.

Context

Traditional ASIC design takes years and requires large teams and budgets, and the share of projects that hit the target on the first silicon wafer is declining: according to a 2024 study by Wilson Research and Siemens, only 14% of IC/ASIC projects are successful, the worst figure in two decades. The claimed “specification → performance model → RTL → verification → firmware → cores” pipeline hits exactly this pain point: verification serves as a single goal for all stages, rather than a separate phase at the end. Early demonstrations of hardware generation by AI models were limited to educational examples like a RISC-V codec, whereas here a full stack of a production accelerator is claimed. The authors also report that the Qwen model running on Redwood helped find optimizations for designing the next-generation accelerator.

Why this matters for the industry

If the ALP pipeline is confirmed, hardware design will transform from a multi-year project into an iterative product process with a cycle from specification to working hardware measured in days, not years. With two architect-specifiers at the input, NRE and time-to-entry for custom silicon drop by an order of magnitude, so chip specialization for a specific workload becomes accessible not only to giants with multi-year budgets but also to startups. The human role shifts to formulating the specification, and the AI role — to generation, verification, and deployment. Strategically, the very possibility of such a result shifts the expectations of investors and incumbents and potentially affects the positioning of EDA vendors Cadence and Synopsys, as well as the wave of AI chip startups.

Why this matters for users

For reader-engineers, the material is useful primarily as a reference for AI hardware generation methodology: the paper is short, seven pages and fifteen figures, and the tile architecture is described in detail. There is no practical benefit at the moment: there is no access to Redwood and ALP, nor a product, prices, API, or toolchain. Verifying the claims by hand is already partially possible: the AMD Versal VPK180 board is commercially available, so if the RTL and cores are published, independent teams will be able to reproduce the FPGA results. For end users, this is a claim for the future: specialized accelerators with low latency and low power consumption in niche devices, i.e., robots, drones, and edge sensors, where hardware iteration is integrated into product sprints.

What is still unknown / limitations

All key figures are claimed by Architect Labs itself, with no independent confirmation: the discussion on Hacker News at the time of measurement consisted of one item and zero comments. Silicon characteristics — throughput 1.75 times higher and power consumption 1.9 times lower than Jetson Orin Nano — were obtained by projection onto the Samsung 8 nm process, not by measuring a real chip, and should be correctly read as a hypothesis to be tested, not a dethroning of NVIDIA. The formulation about the first production-worthy accelerator designed by AI from start to finish remains the authors' claim, not an independently verified fact. The thesis of recursive self-improvement relies on a self-report without a counterfactual baseline: no measurable improvement between design iterations is presented in the sources. There is no news yet about the availability of RTL, tape-out plans, or the timing of real silicon.

Sources

Author

Look at AI, editorial team