AMD's AI software VP on how agentic loops are supercharging chip performance and eroding Nvidia's moat

AMD's AI software VP on how agentic loops are supercharging chip performance and eroding Nvidia's moat

Anush Elangovan joined AMD about two and a half years ago through its acquisition of Node.ai, where he had spent five or six years building ML compilers. He now leads AMD's software strategy with the goal of giving AMD as strong a software story as it has a hardware one.

The agentic acceleration

Until December 2024, Elangovan describes progress as linear, a grind. Since January 2025, agentic AI has materially changed AMD's execution speed. The key mechanism is automated performance loops: when a customer tries a model on AMD silicon, an agent spins up immediately and optimizes that model continuously, without waiting for an engineer to investigate. That compounds on AMD's existing open-source advantage. Every frontier model has already seen AMD's source code in pretraining, which means AI models can rewrite AMD specs and generate optimized kernels from the jump. During a dev day contest AMD ran with GPU Mode, AMD generated more Triton and HIP kernels than had existed on the internet at the time, feeding those back into pretraining data and reinforcing the loop.

The result, Elangovan argues, is that agentic AI has become a great equalizer. Abstractions like Triton were supposed to close the gap with Nvidia's CUDA ecosystem, but agentic loops running nonstop against known rooflines are doing it faster. Overnight agents now scan bugs and pull requests and fix them automatically, with humans in the loop where needed.

“Every frontier model that I use has already seen every bit of AMD source code... Since January, it's just like supercharged our ability to execute. We have automated performance loops that just run — as soon as the customer tries the model, we start an agent that's just nonstop optimizing the customer's model.”

Heterogeneous compute on the edge

AMD's Strix Halo laptop integrates a CPU, GPU, and NPU, and Elangovan says AMD's Agent DKR runtime and compiler can now dynamically shift workloads between them based on what is actually being computed. Tool-calling that isn't GPU-bound routes to the CPU; voice models run on the NPU. Elangovan runs a local voice transcription setup on a Strix Halo using a hacked version of Codex, with heavier reasoning offloaded to cloud models. The near-term goal is making that switching seamless and elastic for developers.

Forward-deployed engineering

AMD has several hundred forward-deployed engineers it calls FDEs, a model started roughly two years ago. Elangovan frames them as the "backward pass" to software developers' "forward pass": developers execute from a product spec forward, FDEs execute from customer problems backward, working the same codebase from opposite directions.

Community feedback and the George Hotz lesson

Elangovan monitors AMD-related keywords on social media personally and says he treats specific GitHub issues as actionable bugs, not just noise. He cites George Hotz and Tiny Corp as a concrete example of open-source feedback driving real change. About a year ago, after community complaints that AMD had dropped support for certain hardware and had poor Windows coverage, Elangovan ran a public poll, ranked the systems developers actually needed, and systematically re-enabled community-supported versions across Windows, Linux, and now macOS.

AMD's open-source posture is Elangovan's throughline. Closed ecosystems can't benefit from AI models trained on their own specs because those specs aren't public. AMD can, and that compounds over time as more AMD code enters training data.

• AMD Developer Day: San Francisco, April 30
• Key figures: Anush Elangovan (VP AI Software, AMD); George Hotz (Tiny Corp) referenced as a catalyst for community-driven software fixes