AMD wants an agent to take over the tedious parts of timing closure, project setup, and board-level debugging. Ross wraps Vivado and Vitis HLS as MCP services and codifies analysis flows as Skills; AMD says more than 800 customers have joined early trials. For engineers in China, the open standards and local deployment are the draw, but the LLM is user-selected and results vary by project.
AMD engineer Shengjie Li wrote about the Ross Agentic AI Assistant, aimed at AMD embedded development and working as an agent. AMD groups everyday FPGA pain points into four categories: repeated analysis of violating paths during timing closure, repeated IP and Block Design configuration during project creation, board-level debugging that depends on individual experience, and senior engineers' know-how that is hard to capture. Ross has four parts: the LLM is user-selected; the MCP Server wraps Vivado and Vitis HLS as MCP services and connects to actual projects; the knowledge base collects public documents such as UGs, PGs, white papers, and ARs into a vector database; and Agent Skills are written by AMD experts and follow the Agent Skills open standard. AMD describes the division of labor as "MCP provides information, Skills provide the ability to act." That means AMD did not build its own model; instead it turned the toolchain and documentation into interfaces an agent can call.
AMD gives an example: a user issues "analyze the timing violations in this post-route DCP." Ross first has the LLM determine the task type, loads the timing-methodology-checks Skill, then executes commands in a Vivado session through MCP, queries the knowledge base when documentation is needed, and finally gives violation classifications and fix suggestions with citations such as UG949; the user decides whether to adopt them. AMD compares Ross with general-purpose AI on five points: general-purpose AI relies on training data and may give deprecated attributes without sources; Ross retrieves official documentation and cites document numbers. General-purpose AI requires manually copying thousands of lines of reports; Ross extracts the clock relationships, logic levels, and fanout of the worst paths on demand. General-purpose AI gives scripts for a person to run; Ross can execute directly, read errors when a Block configuration does not match, adjust, and retry, and asks for confirmation before modifying constraints. Skills fix the analysis flow and keep output format consistent. Prompts shrink from hundreds of words to a single instruction. The real issue is that these comparisons are AMD's own account, and actual results vary by project and model.
Ross does not restrict the client: VS Code, Claude Code, Cursor, Devin, OpenCode CLI, and other MCP-capable clients can be used. It does not restrict the model: cloud models such as Claude, GPT, Gemini, Qwen, and Llama, or local open-source models, are all acceptable. The knowledge base and MCP Server can be deployed locally, suiting intranets or environments with high data-security requirements. Neither MCP nor Agent Skills is an AMD proprietary format, so the related knowledge is transferable. Agent Skills and design examples are published in AMD's official GitHub repository, currently covering several directions, with each example complete on its own and categorized by Skill. According to AMD, more than 800 customers have joined early trials. For teams in China, local deployment and open standards lower the barrier to keeping data inside the intranet, but user-selected LLMs mean model capability and cost fall on the user, and a complete list of Skill coverage has not been published.
For teams in China working on high-speed interfaces and boards, there is no need to replace existing Vivado flows for Ross in the short term; it adds an agent layer on top of existing projects. What really needs evaluation is three things: whether the knowledge base and MCP Server can be deployed locally, which determines whether intranet projects can use it; whether the chosen LLM is on the available list, since cloud models involve cross-border data transfer while local open-source models require bearing compute costs; and whether Skill coverage for areas such as timing and IP configuration matches the devices and tool versions at hand. AMD has not published a complete Skill list or trial feedback, so it is advisable to first verify output consistency on a small project before considering team-wide rollout. For domestic EDA toolchains, MCP and Agent Skills are open standards and the wrapping approach is transferable, but Ross itself targets only AMD tools.