Skip to content

Anush Elangovan

Vice President of AI Software at AMD

  • 3 episodes on Chain of Thought
  • 6 topics covered
  • 3 full transcripts
  • Jun 2025 first on the show
Anush Elangovan

Anush Elangovan is Corporate VP of AI Software at AMD, where he leads AI software strategy after AMD acquired his company Nod.ai in 2023. He appeared on Chain of Thought first to make AMD’s open-ecosystem case against NVIDIA, then returned to walk through his agent-first workflow of running 10 to 12 Claude Code agents in parallel and burning billions of tokens a week.

Cover art for Tuning GPU Performance with AI Agents | AMD’s Anush Elangovan on ROCm 10

Latest episode with Anush Elangovan · EP 72

Tuning GPU Performance with AI Agents | AMD’s Anush Elangovan on ROCm 10

Anush Elangovan, AMD ·

Show notes & transcript → All listening options →

Episodes

What the conversations cover

These notes summarize the episodes featuring Anush Elangovan, including other speakers in panel conversations. Follow a source link to read the discussion in context.

EP 72 Tuning GPU Performance with AI Agents | AMD’s Anush Elangovan on ROCm 10

  • ROCm 10 moves the first interaction with AMD’s software stack into an agentic workflow. A developer can ask Claude Code or Codex to install ROCm, then ask it to serve a model with vLLM; the agent reads the documentation, pulls in the relevant skills, resolves dependencies, and sets up the service. Episode takeaways
  • Hyperloom treats GPU performance tuning as an iterative search problem. It profiles a model to locate kernel or fusion hotspots, then GEAK tests kernel changes while preserving numerical accuracy. Elangovan says AMD used this approach to optimize 14,000 models in one pass when enough compute was available. Episode takeaways
  • Optimization work can continue when the hardware would otherwise sit idle. Elangovan describes Hyperloom running performance experiments in parallel, including while the user sleeps, until the workload reaches a stronger result for its specific deployment. Episode takeaways

EP 56 250,000 Lines of Code/Week: Inside an AMD VP's Agent-First Workflow | Anush Elangovan

  • Anush Elangovan starts coding-agent work with the test plan. He describes defining the test set before implementation and requiring it to pass on each commit, rather than judging progress from generated code alone. Transcript at 14:12
  • Elangovan uses engineering experience to narrow an agent’s task before it starts. In his GPU demand-paging example, he asks for test cases first and inspects them, including noticing repetitive cases. Transcript at 34:43
  • Elangovan favors parallel agents with explicit jobs over unrestricted agent spawning. He describes an issue-monitoring agent with a scheduled task and keeps responsibility for problems it cannot resolve. Transcript at 38:17

EP 28 AMD's Challenge to NVIDIA: The Open Ecosystem Bet | Anush Elangovan & Sharon Zhou

  • Anush Elangovan frames AMD’s MI350 GPU series, built on the CDNA 4 architecture, as delivering 20 petaflops of FP4 performance, which he calls “mind blowing.” He credits AMD’s chiplet design for NUMA load balancing and power savings, since unused chips can be turned off and only what a workload needs is powered on. Episode takeaways
  • CDNA 4 introduces new AI data types — both FP4 and FP6 — and Anush highlights that FP6 throughput is nearly that of FP4. That lets data scientists step models down from FP8 to FP6 as an intermediary before reaching FP4, which he frames as where things are headed. Episode takeaways
  • Anush says the MI350 series reaches 288 GB of memory, enough that “500,000,000,000 parameter models that can run on one GPU,” with at-scale deployments going far larger. He cites deep partnerships with seven of the ten top AI companies as evidence the generational hardware investment is paying off. Episode takeaways

In their words

we can block PRs that go in and are always tested on AMD

Anush Elangovan Tuning GPU Performance with AI Agents | AMD’s Anush Elangovan on ROCm 10 · 9:13

the models themselves are not just aware of the documentation, they're also aware of the architecture ahead of time

Anush Elangovan Tuning GPU Performance with AI Agents | AMD’s Anush Elangovan on ROCm 10 · 21:37

Transcript excerpts; ellipses mark omitted words. Some automatically selected quotes have filler words removed.

Essays featuring Anush Elangovan