Ankit Gupta used coding agents to run a language model on an FPGA, a chip whose circuits can be reconfigured. He had never written Verilog, the hardware-description language involved. Over several days, the agents optimized against two concrete targets: faster output and exact agreement with a pinned CPU reference.

  • The reported gain: 6.43 to 213.13 tokens per second on eight chats with three turns each. The final implementation matched the reference on all 24 turns.
  • The comparison that matters: the 33-fold gain is against the first working FPGA implementation. The fastest exact CPU comparison reached 129.5 tokens per second at almost the same hourly price.
  • The budget: about $10,000 in model tokens. A smaller CPU option was slower but cost $0.24 an hour against the FPGA’s $1.98.
  • The testing tradeoff: the agent’s retrospective says shortening simulations accelerated iteration, but may have let hardware hangs through that cost 38 hours later. That causal explanation remains unproven.

Gupta’s account shows how a measurable correctness target can guide agents through unfamiliar engineering. It also records the limits: the final revision improves short-context speed while roughly tying the previous version on a full-length run.

Caveat: these are author-reported results on a small workload, with unequal optimization effort across hardware alternatives.

The 64-comment thread on Hacker News adds verification and reproducibility concerns.

What the thread adds

  • mikeayles — reports a separate FPGA experiment where external-memory bandwidth constrained a larger model to about 14 tokens per second. That is a commenter-reported experience, not a replication of Gupta’s setup.
  • reuben364 — describes agent-built projects at varying stages of completeness, worries that missing expertise can hide defects, and says testing capacity is now their bottleneck.
  • chalkycrimp and datadrivenangel — ask for code and flag the missing model type/size specification, respectively. Replies disagree about the model; the fetched discussion leaves those reproduction details unresolved.

Handles are pseudonymous; HN publishes no per-comment scores. HN’s ranking shaped this sample, which is a slice of the thread, not a consensus.