arxiv: 2605.05170 · v1 · submitted 2026-05-06 · 💻 cs.AR · cs.AI

Recognition: unknown

Design Conductor 2.0: An agent builds a TurboQuant inference accelerator in 80 hours

The Verkor Team: Ravi Krishna , Suresh Krishna , David Chin

Authors on Pith no claims yet

Pith reviewed 2026-05-08 15:51 UTC · model grok-4.3

classification 💻 cs.AR cs.AI

keywords LLM agentshardware design automationinference acceleratorTurboQuantautonomous RTL generationmulti-agent systemsFPGA mappingquantized inference

0 comments

The pith

An updated multi-agent harness autonomously designs an LLM inference accelerator with built-in TurboQuant support in 80 hours.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper describes Design Conductor 2.0, a multi-agent system driven by frontier models released in April 2026 that scales to tasks 80 times larger than its 2025 predecessor while maintaining full autonomy. It demonstrates the capability by generating four hardware designs from specifications, most notably VerTQ, which embeds TurboQuant quantization directly into a 240-cycle pipeline containing 5129 FP16/32 units. The design maps to an FPGA at 125 MHz and projects to 5.7 mm² in TSMC 16FF for eight attention pipes. This matters to a general reader because it shows AI agents can now produce complex, performance-oriented hardware starting only from a research paper, without continuous human input during the process.

Core claim

Design Conductor 2.0 autonomously produces VerTQ, an LLM inference accelerator that hard-wires TurboQuant support inside a 240-cycle pipeline with 5129 FP16/32 units; the resulting RTL maps to an FPGA at 125 MHz and occupies 5.7 mm² in TSMC 16FF when configured for eight attention pipes.

What carries the argument

The multi-agent harness that coordinates frontier LLMs to read specifications, generate RTL, refine the design, and produce FPGA-mappable output without manual intervention.

If this is right

The harness can now manage 80 times larger hardware tasks at higher quality than the prior version.
Specialized accelerators can be created directly from algorithm papers with integrated support for techniques like TurboQuant.
Designs reach the stage of FPGA mapping and ASIC area estimates through fully autonomous agent workflows.
Empirical measurements of token consumption and system limits become available for scaling studies.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If the autonomy holds at larger scales, custom accelerator development cycles could shrink from months to days.
The same harness pattern might transfer to other engineering domains that currently require large expert teams.
Verification and formal checking of agent output remain the clearest next bottleneck not resolved in the reported work.

Load-bearing premise

The generated hardware designs are functionally correct and achieve their performance targets without hidden bugs or post-generation fixes.

What would settle it

A complete gate-level simulation or FPGA run of the VerTQ RTL that confirms correct TurboQuant inference behavior and matches the claimed cycle count and area estimates against a reference implementation.

read the original abstract

Driven by a rapid co-evolution of both harness and underlying models, LLM agents are improving at a dizzying pace. In our prior work (performed in Dec. 2025), we introduced "Design Conductor" (or just "Conductor"), a system capable of building a 5-stage Linux-capable RISC-V CPU in 12 hours. In this work, we introduce an updated multi-agent harness powered by frontier models released in April 2026, which is able to handle 80x larger tasks, at higher quality, fully autonomously. Following a brief introduction, we examine 4 designs that the system produced autonomously, including "VerTQ", an LLM inference accelerator which hard-wires support for TurboQuant in a 240-cycle pipeline, starting from the TurboQuant arXiv paper. VerTQ includes heavy compute processing, with 5129 FP16/32 units; the design was mapped to an FPGA at 125 MHz and consumes 5.7 mm^2 in TSMC 16FF (8 attention pipes). We review the key new characteristics that enabled these results. Finally, we analyze Design Conductor's token usage and other empirical characteristics, including its limitations.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper reports an updated agent harness scaling to autonomously produce a large TurboQuant inference accelerator with FPGA and ASIC metrics in 80 hours, but provides no verification evidence to support the functional claims.

read the letter

The core update is that Design Conductor 2.0, running on April 2026 frontier models, handles tasks roughly 80 times larger than the prior 5-stage CPU example and produces four autonomous designs, one of which is VerTQ: a 5129-unit FP16/32 inference accelerator with a hard-wired 240-cycle TurboQuant pipeline. The paper gives concrete post-synthesis numbers—125 MHz on FPGA and 5.7 mm² in TSMC 16FF for an 8-pipe version—and discusses token usage plus some new harness characteristics that enabled the jump in scale and quality. That empirical scaling result is the actual new content relative to the 2025 version they cite.

Referee Report

2 major / 2 minor

Summary. The paper introduces Design Conductor 2.0, an updated multi-agent harness powered by April 2026 frontier models that autonomously generates large hardware designs up to 80x larger than prior versions. It presents four such designs, with detailed focus on VerTQ: an LLM inference accelerator that hard-wires TurboQuant support in a 240-cycle pipeline with 5129 FP16/32 units, achieving 125 MHz on FPGA and 5.7 mm² in TSMC 16FF (8 attention pipes). The work reviews enabling harness characteristics and analyzes token usage, empirical performance, and limitations.

Significance. If the generated designs prove functionally correct and meet performance targets without hidden bugs or manual intervention, the work would represent a notable empirical advance in autonomous hardware design automation, demonstrating that LLM agents can now tackle complex, paper-to-RTL tasks at scale with claimed quality improvements over the 2025 version.

major comments (2)

[Abstract / VerTQ description] Abstract and VerTQ section: the headline claim that VerTQ was produced fully autonomously and is functionally correct rests on post-synthesis metrics (125 MHz FPGA mapping, 5.7 mm² TSMC 16FF) alone; no simulation waveforms, testbench coverage, formal verification results, bug reports, or comparison against a hand-written reference implementation are provided, leaving open the possibility that the reported numbers reflect an incomplete or incorrect design.
[Introduction and results sections] The manuscript states the system 'handles 80x larger tasks, at higher quality, fully autonomously' yet supplies no quantitative evidence (e.g., error rates, verification coverage, or side-by-side quality metrics versus hand-designed baselines) to substantiate the quality and autonomy assertions for any of the four designs.

minor comments (2)

[Abstract] The abstract refers to 'we examine 4 designs' but the provided text gives detailed numbers only for VerTQ; brief quantitative summaries for the other three would improve balance and allow readers to assess the breadth of the claimed capability.
[VerTQ description] Notation for pipeline depth ('240-cycle pipeline') and unit count ('5129 FP16/32 units') is clear but would benefit from an accompanying block diagram or table in the main text to clarify dataflow and resource breakdown.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback, which helps clarify the strength of evidence needed for claims of autonomy and correctness. We address each major comment below.

read point-by-point responses

Referee: [Abstract / VerTQ description] Abstract and VerTQ section: the headline claim that VerTQ was produced fully autonomously and is functionally correct rests on post-synthesis metrics (125 MHz FPGA mapping, 5.7 mm² TSMC 16FF) alone; no simulation waveforms, testbench coverage, formal verification results, bug reports, or comparison against a hand-written reference implementation are provided, leaving open the possibility that the reported numbers reflect an incomplete or incorrect design.

Authors: We acknowledge that the manuscript emphasizes post-synthesis and FPGA mapping results as primary evidence of a working design. The agent workflow does include automated testbench generation and simulation steps prior to synthesis, but these were summarized rather than detailed. In revision we will add a verification subsection reporting testbench coverage, key simulation outcomes, and any bugs identified and resolved during the 80-hour run. Formal verification and hand-written reference comparisons were not performed, as they fall outside the agent's current toolset; we will explicitly note this scope limitation. revision: yes
Referee: [Introduction and results sections] The manuscript states the system 'handles 80x larger tasks, at higher quality, fully autonomously' yet supplies no quantitative evidence (e.g., error rates, verification coverage, or side-by-side quality metrics versus hand-designed baselines) to substantiate the quality and autonomy assertions for any of the four designs.

Authors: The paper supplies empirical data on token consumption, runtime, and final design metrics (FPGA frequency, area) across the four designs as evidence of successful autonomous completion at scale. Direct quantitative baselines against hand-designed equivalents and explicit error-rate tables are absent. We will expand the results section with additional internal quality metrics captured by the harness (e.g., self-reported verification pass rates) and add an explicit limitations paragraph discussing the lack of external baseline comparisons. We maintain that the 80x scale increase is evidenced by the concrete designs produced, but agree the presentation can be strengthened. revision: partial

Circularity Check

0 steps flagged

Empirical report of agent-generated hardware with no derivation chain

full rationale

The paper reports empirical outcomes from executing an updated multi-agent harness on hardware design tasks, such as autonomously producing the VerTQ accelerator starting from the TurboQuant arXiv paper. No mathematical derivations, equations, fitted parameters, or first-principles predictions are presented that could reduce to inputs by construction. The reference to prior Design Conductor work is purely contextual background and does not serve as a load-bearing justification for any claimed result. The findings are externally falsifiable via independent reproduction of the agent runs and are not self-referential or tautological.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests on the empirical success of the multi-agent harness in producing working hardware from research papers. No mathematical free parameters or invented physical entities are introduced. Domain assumptions about LLM capabilities for hardware understanding are implicit but not enumerated.

pith-pipeline@v0.9.0 · 5518 in / 1206 out tokens · 42085 ms · 2026-05-08T15:51:59.658359+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

12 extracted references · 2 canonical work pages

[1]

Aes 128 ip.https://www.altera.com/asap/offering/po-1902/aes-128-ip

Altera. Aes 128 ip.https://www.altera.com/asap/offering/po-1902/aes-128-ip. Altera ASAP partner offering by DesignGateway Co., Ltd. Accessed: 2026-05-03

1902
[2]

Performance and resource utilization for floating-point v7.1.https://download.amd.com/ docnav/documents/ip_attachments/floating-point.html#virtexuplus, 2025

AMD. Performance and resource utilization for floating-point v7.1.https://download.amd.com/ docnav/documents/ip_attachments/floating-point.html#virtexuplus, 2025. Vivado Design Suite Release 2025.1. Accessed: 2026-05-03

2025
[3]

Ethernet switch.https://github.com/corundum/ethernet-switch

Corundum. Ethernet switch.https://github.com/corundum/ethernet-switch. GitHub repository. Accessed: 2026-05-03

2026
[4]

Flashattention: Fast and memory- efficient exact attention with io-awareness.Advances in neural information processing systems, 35:16344– 16359, 2022

Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory- efficient exact attention with io-awareness.Advances in neural information processing systems, 35:16344– 16359, 2022

2022
[5]

Scalable hierarchical aggregation protocol (sharp): A hardware architecture for efficient data reduction

Richard L Graham, Devendar Bureddy, Pak Lui, Hal Rosenstock, Gilad Shainer, Gil Bloch, Dror Gold- energ, Mike Dubman, Sasha Kotchubievsky, Vladimir Koushnir, et al. Scalable hierarchical aggregation protocol (sharp): A hardware architecture for efficient data reduction. In2016 First International Workshop on Communication Optimizations in HPC (COMHPC), pa...

2016
[6]

Arjun Kharpal. A google ai breakthrough is pressuring memory chip stocks from samsung to micron.https://www.cnbc.com/2026/03/26/ google-ai-turboquant-memory-chip-stocks-samsung-micron.html, March 2026. CNBC. Ac- cessed: 2026-05-03

2026
[7]

Dworkin, Elaine Barker, James Nechvatal, James Foti, Lawrence E

National Institute of Standards, Technology (NIST), Morris J. Dworkin, Elaine Barker, James Nechvatal, James Foti, Lawrence E. Bassham, E. Roback, and James Dray Jr. Advanced encryption standard (aes), 2001-11-26 00:11:00 2001

2001
[8]

Spike, a risc-v isa simulator.https://github.com/riscv-software-src/ riscv-isa-sim

RISC-V International. Spike, a risc-v isa simulator.https://github.com/riscv-software-src/ riscv-isa-sim. Accessed: 2026-02-03

2026
[9]

Synopsys ultra high-performance aes-xts/ecb ip.https://www.synopsys

Synopsys. Synopsys ultra high-performance aes-xts/ecb ip.https://www.synopsys. com/designware-ip/security-ip/cryptography-ip/symmetric-cryptographic-engines/ ultra-high-perf-aes-xts-ecb.html, 2026. Accessed: 2026-05-03

2026
[10]

Design conductor: An agent autonomously builds a 1.5 GHz Linux-capable RISC-VCPU,

The Verkor Team, Ravi Krishna, Suresh Krishna, and David Chin. Design conductor: An agent au- tonomously builds a 1.5 ghz linux-capable risc-v cpu.arXiv preprint arXiv:2603.08716, 2026

work page arXiv 2026
[11]

Turboquant: Online vector quantization with near-optimal distortion rate,

Amir Zandieh, Majid Daliri, Majid Hadian, and Vahab Mirrokni. Turboquant: Online vector quantiza- tion with near-optimal distortion rate.arXiv preprint arXiv:2504.19874, 2025

work page arXiv 2025
[12]

Turboquant: Redefining ai ef- ficiency with extreme compression.https://research.google/blog/ turboquant-redefining-ai-efficiency-with-extreme-compression/, March 2026

Amir Zandieh and Vahab Mirrokni. Turboquant: Redefining ai ef- ficiency with extreme compression.https://research.google/blog/ turboquant-redefining-ai-efficiency-with-extreme-compression/, March 2026. Google Research Blog. Accessed: 2026-05-03. 12

2026