AMD launches Ross, an agentic AI assistant

On 30 September 2026 the company put out an agent that talks to Vivado and Vitis HLS. The language model is the customer’s choice.

On 30 September 2026 AMD launched AMD Ross, an agentic AI assistant for people who build on FPGAs (chips you reconfigure after they leave the factory), adaptive SoCs (a whole system on one chip), embedded x86 processors, and edge AI platforms. The same day, @AMD published “Meet AMD Ross™” on X, with a video and a product-page link. Marco Chiappetta at Forbes and Hilbert Hagedoorn at Guru3D covered the launch.

Ross is not an AMD model. It is a layer that connects the model the customer already picked to AMD’s own tools. For anyone already living in Vivado, the useful test is whether the agent shortens the flow or only wraps the vendor’s own docs and playbooks.

First: what an agentic AI assistant is here

An agentic assistant here is software that calls tools, looks up documentation, and follows a specialist’s written playbook. Picture a junior engineer who already knows how to open Vivado, find an application note, and run a checklist. The real mechanism is different: MCP servers (Model Context Protocol, an open standard so a model can talk to other programs), a vectorized knowledge base with RAG (it retrieves useful passages before answering), and agent skills, markdown files that capture how someone already did the job.

The picture stops at judgment. AMD’s FAQ says AI workflows are non-deterministic. The newsroom footer says the user still has to review the output.

What is available today

According to AMD, launch day includes the Knowledge Base with RAG, Vivado Design Suite (the FAQ says all versions), and Vitis HLS from version 2025.2 onward. Vitis HLS is the high-level synthesis tool: instead of drawing logic gates, you describe the behavior in C++. Power Design Manager, ChipScope, and Vitis AI are listed for the fourth quarter of 2026, with other pieces on a monthly cadence. The product page names VS Code, Cursor, Devin, Claude Code, Copilot CLI, and Codex CLI.

There is no extra Ross license on top of AMD tool licenses. The FAQ does not spell out the customer’s model or cloud bill. Downloads include MCP servers for Windows and Linux, a VS Code extension (.vsix), skills and examples on GitHub at xilinx/ross-ai-assistant, and an air-gapped option. The support page and the repo mention local embedding models, Qwen3 and Gemma, for that offline setup.

Salil Raje, SVP and GM of AMD Embedded, said in the release that embedded work had grown more complex across hardware, debug, software, AI inference, and system design, and that Ross brings together AMD Embedded tools, company knowledge, and workflows written by specialists.

The numbers, according to AMD

The product overview PDF claims “800+ Customers Participating in Early Access.” The survey after use cited in the same file had 34 complete responses. AMD also says the Knowledge Base holds 500+ user guides, application notes, and articles.

In one image-sensor demosaic pipeline, targeting 300 Mpixels/s, AMD compares C++ written by a specialist with a Ross skill in Vitis HLS: equivalent throughput, 1,801 LUTs versus 1,177 LUTs (about 35% fewer; (1,801 - 1,177) / 1,801 = 34.65%), and “2 hours vs. 2 days.” A LUT is the small logic block inside an FPGA. The same material labels the traditional MATLAB/Python to C++ HLS, pragma, and RTL path as something that “Traditionally Takes 3-6 Weeks.” That 3 to 6 week window is the full cycle of porting the algorithm, not the two-day comparison. This is a vendor demo, not a third-party benchmark. Another timing-skill example says it classified “735 failures” and went from 735 to 0 failing paths. Also according to the company.

Geetha Govindaraj, associate director of the FPGA SOM BU at iWave Global, said in AMD’s materials that Ross “has been a valuable addition… reducing the time and effort required during the debugging and bring-up stages.”

Limits, cost, and who it is for

The newsroom describes a wide scope: PCB, schematic, software, AI, deployment. The PDF limits day one shipping to the knowledge base, Vivado, and Vitis HLS. Guru3D noted that the press release does not quantify overall time or performance gains. We have not tested Ross. AMD sells the tools the agent plugs into. The footer asks users not to send confidential, personal, or regulated data without authorization.

If there is a gain, it comes from having Vivado, documents, and skills in the same loop, not from a new chip or an AMD model. The workflow also stays tied to the AMD/Xilinx stack.

It fits people who already hold Vivado or Vitis HLS licenses and want an agent in VS Code, Cursor, or a CLI, including in an air-gapped setup. It does not fit anyone waiting on an AMD model, or for ChipScope and Vitis AI on day one, or for treating the LUT and time figures as independent proof.

What we still don’t know

We still don’t know the real cost of the customer’s model and cloud, whether the 800+ early access names are active users, or whether Power Design Manager, ChipScope, and Vitis AI arrive in the fourth quarter of 2026 as listed.

If you already sit on that stack, the product page and the GitHub repo are the next step. The rest still needs a human review and a test of your own.

Sources

Official

Press

First: what an agentic AI assistant is here

What is available today

The numbers, according to AMD

Limits, cost, and who it is for

What we still don’t know

Sources

Sources

Markers