← Back to BLACKWIRE GHOST BUREAU AI HARDWARE BREAKTHROUGH Screenshot of FPGA design tool showing C++ and Verilog code generated by an AI agent.

The AI‑generated code base that powered the 33× speedup on an AWS F2 FPGA.

FPGA INFERENCE CHIP REACHES 33X SPEEDUP USING AI AGENT

*Ankit Gupta’s November 20, 2026 blog details how a cloud‑based AI agent rewrote millions of lines of C++/Verilog to run a large language model on an AWS F2 FPGA. The experiment hit 213 tokens per second, a 33‑fold gain, for under $10,000 in token spend.*

By GHOST Bureau - BLACKWIRE  |  November 20, 2026, 20:00 CET  |  FPGA, LLM inference, AI hardware, token budget, cloud agents

On November 20, 2026, Ankit Gupta posted a candid account of turning an AWS FPGA into a functional LLM inference engine. He called the process “vibe‑coding” – letting a large language model drive low‑level hardware design through iterative prompts. The venture began with a naïve understanding of field‑programmable gate arrays, yet within days the AI agent produced a full C++/Verilog stack, verified it, and pushed performance to a real‑time threshold. Gupta’s narrative blends raw token economics, hardware constraints, and the surprising competence of current AI agents in a domain traditionally reserved for seasoned hardware engineers.

Experiment setup and hardware

Gupta reserved an AWS F2 instance, the only readily available FPGA platform, which offers 16 GB of HBM memory. He paired the chip with DeepGrove’s ternary‑weight architecture, a design optimized for memory‑limited inference. The cloud agent ran in Conductor for several days, handling both the initial model load and subsequent speed‑up passes. All code was generated in C++ with portions in Verilog, marking Gupta’s first serious C++ work in a decade and his first exposure to Verilog. The agent logged roughly 3 million lines of code and parsed 30 million output lines while staying within a $10,000 token budget for Claude and OpenAI services.

Performance gains and optimization steps

Baseline throughput measured 6.43 tokens per second. The agent applied a cascade of optimizations—host overhead reduction, burst attention, parallel routing, glue engine, new pipeline, prompt speculation, router relay, head shadow, draft model, memory ports, attention overlap, weight packing, re‑speculation, and parallel head loader—lifting performance to 213.13 tokens per second. That represents a 33.13× increase over baseline. The final configuration ran at a reduced clock of 225 MHz (down from 425 MHz) after a soak test exposed random faults at 231.25 MHz. The speed jump made the FPGA inference loop competitive with real‑time CPU baselines.

The agent wrote three million lines of code and reached 213 tokens per second – a 33× leap from the original baseline.

Can FPGA inference beat CPU in real time?

The 213 tokens‑per‑second figure sits well above the 6‑token baseline, demonstrating that a re‑programmed FPGA can outpace a general‑purpose CPU for LLM inference when paired with aggressive optimization. The experiment proved that, with sufficient token budget and an AI‑driven coding loop, the hardware can achieve real‑time relevance without bespoke silicon. The result challenges the prevailing view that CPUs remain the default for inference, suggesting that cloud‑based FPGA instances may become a viable alternative for latency‑sensitive AI workloads.

Human prompts versus autonomous coding

Gupta’s prompts were deliberately simple: “you’re a superintelligent model, time to be more creative” and “let’s go for 2x faster”. He occasionally injected ML concepts like speculative decoding. The only substantive human input arrived when he consulted DeepGrove founders at roughly 10 TPS; they supplied a concise design prompt that Gupta copied into the session. He credits that prompt as the most significant human contribution, but notes that both Astra and Opus models performed above his “smart‑enough” threshold without further guidance. The experiment suggests that while autonomous agents can generate massive code, targeted expert advice still accelerates critical design decisions.

Frequently asked questions

How much speed improvement did the FPGA achieve?

The FPGA reached 213.13 tokens per second, a 33.13× increase over the 6.43‑token baseline.

What hardware was used in the experiment?

Gupta used an AWS F2 instance with an FPGA that has 16 GB of HBM memory.

How much did the token budget cost?

The entire run consumed about $10,000 in Claude and OpenAI tokens.

Gupta’s experiment proves that AI‑driven code generation can bridge the gap between software and hardware in record time. With a modest $10,000 token spend, a cloud FPGA was transformed into a high‑throughput inference chip, challenging the CPU‑centric status quo. As token budgets grow and models improve, the line between autonomous hardware design and human engineering will blur, reshaping the economics of AI deployment on reconfigurable silicon.

Sources: Hacker News – Vibecoding an Inference Chip (https://ankitg.me/blog/vibecoding-inference-chip/)

More coverageAI agents