Agentic AI in vendor FPGA tools: AMD Ross and Lattice Prompt

Summary

Public material on AMD Ross and Lattice Prompt read side by side, with risks and a community experiment protocol; no own measurements.

Overview

Why now. Within two days in late September 2026 two FPGA vendors described the same pattern: an MCP server that lets an AI agent drive the vendor design tool, plus a bundle of skills and vendor documentation. AMD released Ross on 30 September; Lattice announced on 29 September that it will demo Lattice Prompt and run a hands-on workshop at FPGA Horizons in London on 6-7 October (see /events/fpga-horizons-london-2026). This article reads the public material for both, separates what the sources state from what they leave open, and ends with a protocol the community can use to test the claims. We have run no measurements: everything below is a reading of documents, and our own conclusions are marked as editorial conclusions or hypotheses.

How AMD Ross is built. AMD describes four components: MCP servers that connect an agent to Vivado, Vitis HLS and the other AMD Embedded tools; the AMD Knowledge Base, a vectorized set of user guides, application notes, white papers and answer records, available in the cloud or offline; expert-written agent skills in markdown; and ready-to-run design examples. The release calls Ross client-agnostic, working with any LLM, IDE or command-line client. The Model Context Protocol is, in its own documentation, an open-source standard for connecting AI applications to external systems.

What is actually in the repository. The public repository xilinx/ross-ai-assistant (project card /projects/amd-ross-ai-assistant) is MIT-licensed and holds skills, docs, examples and agent definitions, plus plugin manifests for several clients. Its skills folder lists 50 directories. The README groups the main skills into Vivado (simulation, RTL lint, timing methodology checks, revision control, IP configurator), Vitis HLS (architecture, optimization, MATLAB to C++, build flow), hardware debug (ILA, VIO) and Vitis AI; the rest are helper skills. The changelog records 2026.9.1, dated 30 September 2026, as the initial synced release. The Vivado MCP server is not in the repository: the README sends users to AMD download page for the server or the Vivado AI Extension. The tool reference shows what that server exposes: starting a session, executing Tcl, reading log messages, and running Vivado over SSH or on an LSF cluster.

What a skill looks like. A skill is a SKILL.md file with front matter and step-by-step instructions. The timing skill, vivado-timing-methodology-checks, runs more than 55 methodology checks from UG906, asks the agent to write results into a results folder (violation data, constraints before and after, an HTML diff of the constraints, run and agent logs), has a section on creating waivers and declares Vivado 2026.1 or newer. The hls-optimize skill demands a measured baseline (cosimulation latency, post-route clock period, resource use), one change per attempt and a commit per attempt, and moves from C simulation and synthesis to cosimulation only when synthesis shows at least a 10 percent improvement or a structural change. Editorial conclusion: these are process rules written in plain text, which makes them readable, portable and auditable.

Documentation access and design confidentiality. Documentation search is a separate MCP server, amd-doc-search. The README contrasts the hosted service with a local knowledge base for teams that want documentation search to stay on their own machine. The local option is described as an air-gapped retrieval database with a local embedding model, delivered with either EmbeddingGemma-300M (Gemma Terms of Use, about 3.1 GB) or Qwen3-Embedding-0.6B (Apache 2.0, about 4.4 GB); AMD says a retrieval-quality comparison is not yet published. AMD is explicit that this covers only retrieval: the model that writes the final answer can be a cloud model or a local one. The release also says users should not submit confidential, personal or regulated data unless it is authorized for the model, tool and deployment configuration. Editorial conclusion: an offline knowledge base alone does not make the workflow private; the answering model and the agent client decide that.

Vendor agent or generic agent with a home-made MCP. The question is often posed as AMD Ross against a general agent such as Claude or Codex with a self-written MCP wrapper around Vivado. The sources suggest a different split. AMD FAQ points to existing coding agents (Cursor, Claude Code, Copilot or similar) and to a current large-context coding model that your company already allows, so Ross is a layer for existing clients rather than a separate agent. What AMD adds is a curated documentation corpus, skills written for its tools and a maintained MCP server. A home-made server can in principle run Tcl in the same way, since Tcl execution is one of the core tools in AMD reference, so the difference is likely to sit in the documentation and the skills rather than in the execution channel. This is a hypothesis of the editors, and the protocol below is designed to test it.

Portability of skills. According to AMD FAQ the skills follow the open Agent Skills standard, and the repository ships manifests for several clients, so the skills can be loaded into a generic agent. The limits are practical. Skills that need live Vivado require the Vivado MCP server, while the HLS skills work through command-line tools. Version statements differ: the timing skill declares Vivado 2026.1 or newer, the FAQ says the server supports Vivado 2020.2 or later and is tested on 2026.1, and the product page says Vivado is supported in all versions (Vitis HLS from 2025.2). Editorial conclusion: treat these statements as unreconciled until you test on your own Vivado version.

What is known about Lattice Prompt. The Lattice page describes an AI design assistant that interfaces with Lattice Radiant for compilation, simulation and validation and runs as an open MCP server compatible with any capable IDE and major LLMs. Compilation and simulation flows run through Radiant locally or over the network, and answers are said to be grounded in Lattice documentation. The page offers a free download of the MCP server and Skills bundle for Windows and Linux alongside an existing Radiant installation, with a one-command installer; software version 1.0.1 and user guide 1.0 are dated 17 September 2026. Lattice will demo the tool on 6 October and use it in the 7 October Nexus workshop at FPGA Horizons, where AMD also lists an agentic AI workshop.

What is not known about Lattice Prompt. From the public pages we could not establish the list of skills, the license of the bundle, whether documentation retrieval is hosted or local, which device families and Radiant versions are covered, or how design data is handled; the PDF user guide was not reviewed for this article. The page quotes productivity gains of 10X and more from early users without a method, so we treat it as a vendor claim. We do not infer how Lattice Prompt is built beyond what the page says.

Risks. First, constraints and timing advice. The AMD FAQ confirms that the agent can change the project inside the session, for example adding IP or changing constraints; the timing skill has a waiver section and produces a constraints diff, which is the right artifact to review. Hypothesis of the editors: a plausible but wrong constraint or waiver can make a report look clean while the design is wrong. Second, verifiability: AMD states that AI workflows are non-deterministic, so one successful run proves little. Third, licenses and terms: the repository is MIT, Vivado and Radiant keep their own licenses, the AMD FAQ says Ross needs no extra license beyond the AMD tools, the local knowledge base ships embedding models under different licenses, and the license of the Lattice bundle is not stated on its page. Fourth, vendor numbers: AMD publishes none, and Lattice gives 10X without a method.

Experiment protocol, setup. This is a protocol, not a report; nobody at FPGA.camp has run it yet. Take a real design your team may share: an open design, or a closed one only with a local knowledge base and a local model after the data policy has been checked. Fix the tool version (Vivado or Radiant), device, constraints, strategies and seeds. Compare three arms on the same tasks: (A) a vendor package, that is AMD Ross skills with the vendor MCP server, or Lattice Prompt; (B) a generic agent with the same model and a self-written Tcl MCP and no vendor skills; (C) an engineer without an agent as the human baseline. Use the same model in A and B where the client allows it, and repeat each agent arm several times, because AMD itself warns that results vary between runs. Open cores and projects from the catalog, for example /cores/vexriscv or /projects/litex, can be candidates if their build flow fits the vendor tool; that has to be checked first.

Experiment protocol, tasks and metrics. Task 1, timing closure: start from a design with failing setup timing after route; the agent may analyze reports and propose or apply RTL or constraint changes. Task 2, HLS pragma tuning: take a kernel with a testbench; the agent tunes pragmas and loops against a stated target such as latency or resource use. Task 3, lint: run RTL lint and fix reported issues without changing behavior. Record WNS and TNS (and hold slack where relevant), LUT and FF (plus DSP and BRAM where relevant), post-route clock period and, for HLS, cosimulation latency and true latency as cycles times period, wall-clock time per iteration and the number of tool runs, the number of manual edits by the engineer, the number of agent changes rejected on review, and whether simulation still passes.

Interpretation and publication. Count a result only if functional simulation or cosimulation still passes and a human has reviewed the constraints diff; a timing gain from a loosened constraint is a failure, not a result. Report the median and the spread over repeated runs, not the best run. Publish tool versions, model name, client, skill versions (release 2026.9.1 for the AMD repository), the full agent logs and the raw reports so that others can repeat the run. Threats to validity: one design is an anecdote, a model update in the middle of the experiment changes the arm, vendor skills may be tuned on the vendor examples, and prompt wording matters. Until the community collects such runs, this article makes no claim about which approach is faster.

Verified facts

Ross consists of MCP servers, a knowledge base, agent skills and design examples, and works with any LLM, IDE or CLI client. Source: https://newsroom.amd.com/news/amd-ross-agentic-ai-embedded-design-development/. Checked: 2026-10-02. Confidence: 5/5.

MCP is described by its documentation as an open-source standard for connecting AI applications to external systems. Source: https://modelcontextprotocol.io. Checked: 2026-10-02. Confidence: 5/5.

The repository is MIT-licensed, lists 50 skill directories and records release 2026.9.1 of 2026-09-30. Source: https://github.com/xilinx/ross-ai-assistant/tree/main/skills. Checked: 2026-10-02. Confidence: 4/5.

The Vivado MCP server is installed separately from the repository. Source: https://github.com/xilinx/ross-ai-assistant. Checked: 2026-10-02. Confidence: 5/5.

The MCP tool reference includes Tcl execution, log parsing and remote runs over SSH or LSF. Source: https://raw.githubusercontent.com/Xilinx/ross-ai-assistant/main/docs/reference/vivado-mcp-tools.md. Checked: 2026-10-02. Confidence: 5/5.

The timing skill runs 55+ UG906 checks, writes a constraints diff, covers waivers and declares Vivado 2026.1+. Source: https://raw.githubusercontent.com/Xilinx/ross-ai-assistant/main/skills/vivado-timing-methodology-checks/SKILL.md. Checked: 2026-10-02. Confidence: 5/5.

The hls-optimize skill requires a measured baseline, one change per attempt and a commit per attempt. Source: https://raw.githubusercontent.com/Xilinx/ross-ai-assistant/main/skills/hls-optimize/SKILL.md. Checked: 2026-10-02. Confidence: 5/5.

The local knowledge base is air-gapped for retrieval, while the answering model can be cloud or local. Source: https://raw.githubusercontent.com/Xilinx/ross-ai-assistant/main/docs/local-kb/README.md. Checked: 2026-10-02. Confidence: 5/5.

The two embedding packages carry different licenses: Gemma Terms of Use and Apache 2.0. Source: https://raw.githubusercontent.com/Xilinx/ross-ai-assistant/main/docs/local-kb/README.md. Checked: 2026-10-02. Confidence: 5/5.

The FAQ names Vivado 2020.2 or later with 2026.1 tested; the product page says all versions; the timing skill declares 2026.1+. Source: https://raw.githubusercontent.com/Xilinx/ross-ai-assistant/main/docs/faq.md. Checked: 2026-10-02. Confidence: 4/5.

AMD states AI workflows are non-deterministic and results may vary between runs. Source: https://www.amd.com/en/products/software/ross-agentic-ai.html. Checked: 2026-10-02. Confidence: 5/5.

Lattice Prompt is an open MCP server that works with Lattice Radiant; version 1.0.1 is a free download for Windows and Linux dated 2026-09-17. Source: https://www.latticesemi.com/Products/DesignSoftwareAndIP/FPGAandLDS/Lattice-Prompt. Checked: 2026-10-02. Confidence: 5/5.

Lattice Prompt page quotes 10X productivity gains from early users without a method. Source: https://www.latticesemi.com/Products/DesignSoftwareAndIP/FPGAandLDS/Lattice-Prompt. Checked: 2026-10-02. Confidence: 5/5.

Engineering benefit

Skills with explicit baselines, one-change-per-attempt rules and constraints diffs give engineers a checkable process for timing and HLS work. Editorial reading of the skill text; no measured effect is published. Source: https://raw.githubusercontent.com/Xilinx/ross-ai-assistant/main/skills/hls-optimize/SKILL.md. Checked: 2026-10-02. Confidence: 4/5.

The Lattice Prompt MCP server connects an agent to Radiant compilation and simulation flows. Skill list and device coverage are not stated on the page. Source: https://www.latticesemi.com/Products/DesignSoftwareAndIP/FPGAandLDS/Lattice-Prompt. Checked: 2026-10-02. Confidence: 4/5.

Commercial benefit

The AMD FAQ states Ross needs no license beyond the AMD tools, and the Lattice Prompt bundle is offered as a free download, so trial cost is mainly engineer time and model usage. Model usage cost and Lattice bundle license terms are not stated. Source: https://raw.githubusercontent.com/Xilinx/ross-ai-assistant/main/docs/faq.md. Checked: 2026-10-02. Confidence: 4/5.

Community benefit

Open markdown skills and an open MCP standard let the community compare, audit and port vendor workflows. Openness covers skills and protocol; vendor tools and some servers stay proprietary. Source: https://modelcontextprotocol.io. Checked: 2026-10-02. Confidence: 4/5.

Critical review

The article reads documents, not runs; vendor claims about acceleration remain untested here. Source: https://newsroom.amd.com/news/amd-ross-agentic-ai-embedded-design-development/. Checked: 2026-10-02. Confidence: 5/5.

Version statements for Vivado support differ between the product page, the FAQ and the timing skill. Source: https://raw.githubusercontent.com/Xilinx/ross-ai-assistant/main/docs/faq.md. Checked: 2026-10-02. Confidence: 4/5.

A local knowledge base does not make the workflow private if the answering model is a cloud model. Source: https://raw.githubusercontent.com/Xilinx/ross-ai-assistant/main/docs/local-kb/README.md. Checked: 2026-10-02. Confidence: 5/5.

Lattice Prompt internals, license and data handling are not documented on its product page. Source: https://www.latticesemi.com/Products/DesignSoftwareAndIP/FPGAandLDS/Lattice-Prompt. Checked: 2026-10-02. Confidence: 4/5.

Most claims come from vendor primary sources and repository files that anyone can re-check. Source: https://github.com/xilinx/ross-ai-assistant. Checked: 2026-10-02. Confidence: 4/5.

The protocol makes no result claims and lists threats to validity. Source: https://raw.githubusercontent.com/Xilinx/ross-ai-assistant/main/skills/hls-optimize/SKILL.md. Checked: 2026-10-02. Confidence: 4/5.

Practical recommendations

If you run the protocol, post tool versions, model, client, skill versions, logs and raw reports; a table without them is a rumour with columns.

Run the first pass on an open design: this keeps the data question out of the experiment and lets others repeat it.

Review the constraints diff yourself before accepting any timing result from an agent.

Official links

AMD newsroom release

AMD Ross product page

xilinx/ross-ai-assistant repository

Repository FAQ

Vivado MCP tool reference

Local knowledge base guide

Timing methodology skill

HLS optimize skill

Skills directory

Lattice Prompt product page

Model Context Protocol

FPGA Horizons London 26

Evidence

AMD introduced AMD Ross on 2026-09-30 with four components: MCP servers, the AMD Knowledge Base, expert-authored agent skills and design examples. Source: https://newsroom.amd.com/news/amd-ross-agentic-ai-embedded-design-development/. Checked: 2026-10-02.

The release calls Ross client-agnostic: any preferred LLM, IDE or command-line environment. Source: https://newsroom.amd.com/news/amd-ross-agentic-ai-embedded-design-development/. Checked: 2026-10-02.

The MCP documentation describes MCP as an open-source standard for connecting AI applications to external systems. Source: https://modelcontextprotocol.io. Checked: 2026-10-02.

The repository xilinx/ross-ai-assistant holds an MIT License file and its skills folder lists 50 skill directories. Source: https://github.com/xilinx/ross-ai-assistant/tree/main/skills. Checked: 2026-10-02.

The LICENSE file is the MIT License, copyright 2026 Advanced Micro Devices, Inc. Source: https://raw.githubusercontent.com/Xilinx/ross-ai-assistant/main/LICENSE. Checked: 2026-10-02.

CHANGELOG records release 2026.9.1 dated 2026-09-30 as the initial synced release. Source: https://raw.githubusercontent.com/Xilinx/ross-ai-assistant/main/CHANGELOG.md. Checked: 2026-10-02.

The README states the Vivado MCP Server or Vivado AI Extension is installed separately from the AMD Ross download page. Source: https://github.com/xilinx/ross-ai-assistant. Checked: 2026-10-02.

The Vivado MCP tool reference lists tools for session management, Tcl execution (vivado_execute), log parsing, and remote runs over SSH or LSF. Source: https://raw.githubusercontent.com/Xilinx/ross-ai-assistant/main/docs/reference/vivado-mcp-tools.md. Checked: 2026-10-02.

The vivado-timing-methodology-checks skill runs 55+ methodology checks (UG906), writes results including a constraints diff and logs, covers creating waivers, and declares compatibility with Vivado 2026.1+. Source: https://raw.githubusercontent.com/Xilinx/ross-ai-assistant/main/skills/vivado-timing-methodology-checks/SKILL.md. Checked: 2026-10-02.

The hls-optimize skill requires a measured baseline (cosimulation latency, post-route clock period, resource utilization), one change per attempt, a commit per attempt, and cosimulation only after csynth shows at least 10% improvement or a structural change. Source: https://raw.githubusercontent.com/Xilinx/ross-ai-assistant/main/skills/hls-optimize/SKILL.md. Checked: 2026-10-02.

The local knowledge base is an air-gapped retrieval database covering Vivado, Vitis, Power Design Manager, ChipScope, system software, example designs, wiki and answer records; the answering model may be a cloud or a local model. Source: https://raw.githubusercontent.com/Xilinx/ross-ai-assistant/main/docs/local-kb/README.md. Checked: 2026-10-02.

The local knowledge base ships with one of two embedding models: EmbeddingGemma-300M under the Gemma Terms of Use (about 3.1 GB) or Qwen3-Embedding-0.6B under Apache 2.0 (about 4.4 GB); a retrieval-quality comparison is not yet published. Source: https://raw.githubusercontent.com/Xilinx/ross-ai-assistant/main/docs/local-kb/README.md. Checked: 2026-10-02.

The AMD FAQ says Ross works with Cursor, Claude Code, Copilot or similar agents that load Agent Skills and call tools, and recommends a current large-context coding model. Source: https://raw.githubusercontent.com/Xilinx/ross-ai-assistant/main/docs/faq.md. Checked: 2026-10-02.

The AMD FAQ states Vivado 2020.2 or later is supported with 2026.1 as the tested version, and that the agent can change the project in the session (for example adding IP, changing constraints). Source: https://raw.githubusercontent.com/Xilinx/ross-ai-assistant/main/docs/faq.md. Checked: 2026-10-02.

The product page states Vivado is supported in all versions, Vitis HLS from 2025.2, no additional license beyond the AMD tools, and that AI workflows are non-deterministic. Source: https://www.amd.com/en/products/software/ross-agentic-ai.html. Checked: 2026-10-02.

The release states users must review generated outputs and should not submit confidential, personal or regulated data unless authorized. Source: https://newsroom.amd.com/news/amd-ross-agentic-ai-embedded-design-development/. Checked: 2026-10-02.

Lattice Prompt is an AI design assistant that interfaces with Lattice Radiant for compilation, simulation and validation and runs as an open MCP server compatible with any capable IDE and major LLMs. Source: https://www.latticesemi.com/Products/DesignSoftwareAndIP/FPGAandLDS/Lattice-Prompt. Checked: 2026-10-02.

The Lattice Prompt page offers a free download of the MCP server and Skills bundle for Windows and Linux alongside Radiant; software version 1.0.1 and user guide 1.0 are dated 2026-09-17. Source: https://www.latticesemi.com/Products/DesignSoftwareAndIP/FPGAandLDS/Lattice-Prompt. Checked: 2026-10-02.

The Lattice Prompt page cites productivity gains of 10X and more reported by early users, without a method. Source: https://www.latticesemi.com/Products/DesignSoftwareAndIP/FPGAandLDS/Lattice-Prompt. Checked: 2026-10-02.

FPGA Horizons London lists a Lattice Nexus FPGA hands-on and agentic AI workflow workshop with Matt Holdsworth, Senior FAE, on 2026-10-07 and a separate AMD-attributed agentic AI workshop. Source: https://www.fpgahorizons.com/london-26/. Checked: 2026-10-02.