AMD AI DevDay 2026

Annmaria Antony

Student Researcher, Stanford University 

Annmaria Antony is a junior studying Computer Science (concentrations in computer systems and AI) at Stanford University. Previously, she worked as a researcher at Integrated Biosciences, a start-up backed by Sutter Hill Ventures that utilizes deep learning for small-molecule drug discovery. She has also held software engineering roles at NVIDIA and Intel on the VLSI and Network & Edge teams, respectively.

Tech Talk: Adapting Language Models for Low-Resource GPU Kernel Programming

Abstract
Large Language Models perform well on high-resource programming languages, but they struggle to generate low-resource, compiler-verified languages such as AMD HIP, where open-source training data is scarce and performance constraints are strict. In this work, we investigate methods for improving HIP kernel generation with (1) using RAG over HIP/ROCm documentation and (2) SFT on a closed-loop synthetic data pipeline built on top of KernelBench. As part of this work, we constructed a synthetic dataset of high-quality kernel examples containing the KernelBench task description, reference PyTorch architectures, generated HIP kernels, compilation and correctness results, speedup measurements, and full prompt histories across iterative attempts. We benchmark on AMD MI350X GPUs using KernelBench on 3 metrics: compilation success, functional correctness, and execution speedup relative to PyTorch baselines.

Vamsi Boppana

Senior Vice President, AI, AMD​

Vamsi Boppana is senior vice president of the Artificial Intelligence Group at AMD. In this role, he is responsible for the Data Center AI GPU business as well as the cross-company AI strategy, driving the AI roadmap across client, edge, and cloud, and for the company’s AI engines and AI software and ecosystem efforts.​

Boppana joined AMD from Xilinx and has served in a variety of leadership roles across the combined company. Prior to his role leading the AMD Artificial Intelligence Group, he served as senior vice president of the Central Products Group (CPG) at Xilinx responsible for developing the company’s adaptive and AI product portfolio. Boppana also led the development of industry-leading products including Versal™ and Zynq™UltraScale™+ MPSoC. Before joining Xilinx in 2008, he held engineering management roles at Open-Silicon and Zenasis Technologies, a company he co-founded.​

Boppana earned a Bachelor of Technology, with honors, in computer science and engineering from the Indian Institute of Technology, Kharagpur, and Master of Science and doctorate degrees in electrical engineering from the University of Illinois at Urbana-Champaign. The author and co-author of more than 40 technical papers, he has received  six patents for his designs. 

Daniel Han

CEO, Unsloth

Daniel Han is the CEO of Unsloth, an open-source AI startup on a mission to make AI accessible and accurate for everyone. Unsloth has grown rapidly, earning over 38,000 GitHub stars and surpassing 10 million monthly downloads on Hugging Face. The team has contributed critical bug fixes to major models including LLaMA, Phi, Gemma, and Mistral. Prior to founding Unsloth, Daniel worked at NVIDIA, where he achieved a 2000x speedup in t-SNE visualization.

Tech Talk: The Race for AGI - Open Source vs Closed Source

Abstract​
Open models like DeepSeek, Gemma-4, Kimi, GLM, and Qwen (and more!) are great, but closed source models like Claude and ChatGPT still dominate in performance and real-world use. If only a few companies lead the AGI race, they end up shaping its direction. So how do we close the gap between large labs and open source? And how do we make open models run seamlessly across all hardware, including AMD GPUs, to truly democratize and expand access to AI?

Jack Huynh

Senior Vice President and GM, Computing and Graphics Group, AMD

Jack Huynh is the senior vice president and general manager of the Computing and Graphics Group, overseeing our PC, graphics, and semi-custom businesses. As the leader of this group, he is responsible for all aspects of strategy, business, and engineering to advance our leadership in AI and high-performance computing across end-user devices. Jack brings over 25 years of experience driving innovation and delivering leadership CPU, GPU, and NPU technologies across gaming, consumer, and enterprise devices. 

Since joining AMD in 1998, Huynh has spanned a breadth of businesses and technologies while building deep, strategic partnerships that have transformed AMD’s product portfolio and business. He previously led the AMD semi-custom group as senior vice president and general manager, responsible for managing all aspects of strategy, business management, and engineering for high-performance AMD semi-custom solutions.

Junchen Jiang

CEO and Co-Founder of Tensormesh, Faculty Lead of LMCache Lab, and Associate Professor at University of Chicago

Junchen Jiang is CEO and Co-Founder of Tensormesh, Faculty Lead of LMCache Lab, and Associate Professor of Computer Science at the University of Chicago. He received his PhD degree from Carnegie Mellon University and Bachelor from Tsinghua University. He has received CMU Computer Science Doctoral Dissertation Award (2017), two Google Faculty Research Awards on ML Systems (2018, 2026), an NSF CAREER Award, and multiple best paper awards.

Tech Talk: KV Cache as the New AI Memory Abstraction​

Abstract​
Modern AI agents increasingly operate over long contexts—reading documents, executing multi-step plans, and maintaining evolving state—but today’s inference engines reprocess this growing context from scratch, and this causes low throughput, high cost and latency bottlenecks. The key to tackling these bottlenecks is to elevate KV cache as a first-class memory layer for agentic systems, allowing the LLMs to reuse their past reasoning instead of recomputing it. I present LMCache, an open-source industry-adopted KV caching library that provides agents with fast, persistent, and addressable AI-native memory across steps, tasks, and model invocations, which dramatically reduces prefill time and GPU cost. I will also highlight the roadmap that brings in more research to industry adoption, in particular, KV cache compression and KV cache blending. Together, they make KV cache much smaller and reusable in many more cases. In short, KV cache is the missing substrate for scalable, efficient, and long-context AI agents, and building KV-aware memory infrastructure will be essential to the next generation of AI systems.​

Laasya Konidala

Student Researcher, Stanford University

Laasya Konidala is a junior studying Computer Science at Stanford University, focusing on computer systems and AI. She currently does research with Stanford’s Scaling Intelligence Research Lab and has previously done research with Stanford’s Multiscale Architecture & Systems Team.

Tech Talk: Adapting Language Models for Low-Resource GPU Kernel Programming

Abstract
Large Language Models perform well on high-resource programming languages, but they struggle to generate low-resource, compiler-verified languages such as AMD HIP, where open-source training data is scarce and performance constraints are strict. In this work, we investigate methods for improving HIP kernel generation with (1) using RAG over HIP/ROCm documentation and (2) SFT on a closed-loop synthetic data pipeline built on top of KernelBench. As part of this work, we constructed a synthetic dataset of high-quality kernel examples containing the KernelBench task description, reference PyTorch architectures, generated HIP kernels, compilation and correctness results, speedup measurements, and full prompt histories across iterative attempts. We benchmark on AMD MI350X GPUs using KernelBench on 3 metrics: compilation success, functional correctness, and execution speedup relative to PyTorch baselines.

Ryan Lee 

Lead Engineer, Developer Relationships, Minimax

Ryan Lee is a R&D Engineer and Head of Developer Relations at MiniMax. He has extensive algorithm engineering experience at leading internet companies, and has long been committed to the implementation and application of AI technologies in efficiency scenarios.

Tech Talk: Best Practices for Agent Harness Based on MiniMax M2.7

Abstract

1. How Does MiniMax M2.7 Start Model Self-Evolution
2. How to Build an Agent Harness to Enable One Person to Valuably
3. Consume 1B Tokens per Day
4. Infrastructure Above LLM and Infrastructure Below LLM

Natalia Pahlavan

Student Researcher, Stanford University 

Natalia Pahlavan is a junior at Stanford University studying Computer Science, specializing in AI and Systems. She is a researcher at Stanford’s Translational AI (STAI) lab and has previously held software engineering roles at Point72, Scale AI, and Trellis Law.

Tech Talk: Adapting Language Models for Low-Resource GPU Kernel Programming

Abstract
Large Language Models perform well on high-resource programming languages, but they struggle to generate low-resource, compiler-verified languages such as AMD HIP, where open-source training data is scarce and performance constraints are strict. In this work, we investigate methods for improving HIP kernel generation with (1) using RAG over HIP/ROCm documentation and (2) SFT on a closed-loop synthetic data pipeline built on top of KernelBench. As part of this work, we constructed a synthetic dataset of high-quality kernel examples containing the KernelBench task description, reference PyTorch architectures, generated HIP kernels, compilation and correctness results, speedup measurements, and full prompt histories across iterative attempts. We benchmark on AMD MI350X GPUs using KernelBench on 3 metrics: compilation success, functional correctness, and execution speedup relative to PyTorch baselines.

Banghua Zhu

Co-Founder & CTO, RadixArk

Banghua Zhu is CTO and Co-founder of RadixArk, where he is building open AI infrastructure to bring frontier-level systems to a broader community. Previously, he was a principal research scientist at Nvidia, and co-founder at Nexusflow AI. He received his PhD in EECS from UC Berkeley.

Tech Talk: Building Next-Gen AI Infrastructure: Scaling Enterprise LLM Serving and Training with RadixArk

Abstract
Deploying generative AI in production requires a robust, scalable, and highly optimized infrastructure. Join Banghua Zhu, co-founder and CTO of RadixArk, to discover how to build an efficient LLM inference and training stack with SGLang and Miles. This session will cover the open-source innovations in SGLang and Miles, and enterprise-ready inference and training services with RadixArk. We will demonstrate how our advanced inference and training engines maximize compute utilization, and scale AI applications seamlessly across diverse environments—ultimately building the foundations needed to bring the power of frontier AI to everyone.

Iswarya Alex​

Member of Technical Staff, AMD

ML Engineer at AMD specializing in high-performance, on-device AI. My work focuses on fine-tuning, optimizing, and deploying machine learning models locally on Ryzen AI–powered systems, leveraging both GPUs and NPUs to deliver efficient, real-time experiences.

Mahdi Ghodsi

AI Solutions Architect, AMD

Mahdi Ghodsi is an AI Solution Architect at AMD, specializing in the integration of open-source AI projects and the development of advanced AI solutions optimized for AMD GPUs. Previously, he led the deployment of production AI systems for dozens of large companies worldwide, ensuring robust and scalable implementations. Mahdi holds a master’s degree in Computer Engineering with a specialization in Deep Learning for Computer Vision, from the University of British Columbia.

Alex He

AI Solutions Architect, AMD

Alex He is an AI Solutions Architect at AMD AIG, focusing on ROCm ecosystem expansion and GPU product application enablement. His current focus is on Physical AI applications on AMD products.

David Li

Product Engineer, AMD

David Li is a Product Engineer at AMD AIG team, focusing on LLM inference and Physical AI acceleration on GPUs. Previously as Deep Learning Solution Architect at NVIDIA, where specialized in self-driving simulation and GPU acceleration.

Eda Zhou

Software Development Engineer, AMD

Eda Zhou is a Software Development Engineer at AMD, focused on creating developer-facing materials such as courses, tutorials, and videos to empower the AI and software community. Previously, she was a founding engineer at Lamini, where she helped design and scale one of the first LLM training and inference platforms built on AMD GPUs. Eda holds a Master’s degree in Computer Science from Worcester Polytechnic Institute.

Liz Li

Principal AI Engineer, AMD

Liz Li is a Principal AI Engineer in the AMD AI group, specializing in enabling and optimizing cutting-edge AI models on AMD Instinct GPUs for both distributed inference and training. With over 10 years of experience in computer, graphics, and AI architecture, she has previously led cross-functional teams in delivering platform hardware and software architecture requirements and optimizations for a variety of AI use cases.

Naveen Purushotham 

Member of the Technical Staff, AMD

Naveen Purushotham is a Member of the Technical Staff at AMD University Program, part of AMD Research and Advanced Development, in San Jose, CA, where he collaborates with academic institutions on teaching and research, with a focus on efficient and applied deep learning for AMD AI platforms.

Yusheng Su

Member of Technical Staff, RadixArk

Yusheng Su is a Member of Technical Staff at RadixArk, where he builds AI systems for frontier models. Currently, he is a main contributor to SGLang and Miles. Previously, he worked at AMD and received his PhD in CS from Tsinghua University.

Tech Talk: Determinism and Scalability of the Miles RL Framework Across Hardware Platforms

Abstract
Building an RL framework requires careful consideration of both stability and efficiency during training. In this talk, Yusheng from RadixArk will present key challenges in RL systems and explain how the Miles framework ensures deterministic training while maintaining scalability. The session will highlight the core design principles and features of Miles, including its support for determinism, scalability, and fault-tolerant protection. These capabilities enable more reliable RL training at scale, ultimately addressing the needs of frontier labs and real-world production environments.

Timothy Carambat

Founder and CEO, AnythingLLM

Founder and CEO of AnythingLLM, an open-source platform transforming local LLMs into seamless, cloud-like experiences. A self-taught developer with a mechanical engineering background, Tim has spent over a decade building consumer-facing products and enterprise automation for Fortune 500 companies.

Tech Talk: Stop Throwing Tokens at Your Problems and Start Engineering

Abstract
This session explores how the next wave of AI innovation will be driven by thoughtful system design rather than sheer model scale. While the cloud offers virtually unlimited resources, building for on-device environments forces developers to engineer more efficient, cost-effective, and user-centric solutions. This talk argues that today’s models are already capable—what’s missing is better tooling, orchestration, and “cloud-like” local experiences that make them feel truly intelligent. By treating AI systems as a composition of smaller, specialized components, developers can unlock powerful capabilities without relying on massive infrastructure or token-heavy approaches. The session also highlights how hardware-aware development and practical constraints lead to better user experiences and more sustainable applications. Ultimately, the path forward isn’t bigger models—it’s smarter systems that bridge local and cloud seamlessly.

Chris Lattner

Co-founder and CEO of Modular

Chris Lattner is co-founder and CEO of Modular. Before that, he built infrastructure that millions of developers use every day: LLVM and Clang, the compiler toolchain underlying most modern software, and Swift, the language Apple shipped to hundreds of millions of devices. He led AI infrastructure work at Google, then co-founded Modular to fix a problem he kept running into: AI deployment is slow, hardware-specific, and harder than it should be. Modular's answer is MAX, an inference framework that gives teams full control across hardware stacks, and Mojo  , a programming language that unifies high-level AI development with low-level systems programming. When he's not building compilers or running a company, Chris enjoys woodworking and writing.

Tech Talk: Compiling the Full Diffusion Pipeline: 4x Faster Image Generation on MI355X

Abstract
FLUX.2 is a 32-billion parameter diffusion model split across four stages: DiT backbone, VAE decoder, text encoder, and scheduler. Most serving stacks run these as separate steps with Python overhead between each one. MLIR-based compilation collapses all four into a single fused execution graph, eliminating that overhead and generating hardware-optimized code for AMD Instinct MI355X. Modular's compiled pipeline runs 3.8x faster than torch.compile, generates 1024x1024 images in under 3.5 seconds, and ships in a container under 700MB (90% smaller than a typical inference stack). By compiling the same pipeline for both AMD Instinct MI355X and Blackwell GPUs, we show that AMD silicon delivers equivalent generation performance at a 5.5x total cost advantage.

Gaël Delalleau

Founder and CEO, Kog

Gaël Delalleau is the founder and CEO of Kog, where he pushes the limits of low-level GPU engineering and LLM architectures. A graduate of École Polytechnique and a four-time finalist at the Defcon CTF hacking competition, he has spent two decades breaking and building systems — from cybersecurity research and penetration testing of critical infrastructure to co-founding LeHack, France's largest hacker event. He is a serial entrepreneur whose previous ventures span AI, gaming, and the social web.

Tech Talk: Building a real-time inference stack on AMD Instinct GPUs

Abstract
In this talk, we share our vision for real-time generative AI, and the techniques we developed to achieve the fastest LLM inference on GPU ever, with a generation speed of 4000 tokens/s per request.
We first showcase our end-to-end stack optimized for minimal latency on AMD hardware, spanning model re-architecting, a single monokernel implementation, along with topology-aware algorithms.
In the second part, we focus on one of the defining challenges of megakernels, intra-GPU grid synchronization barriers and reduce/gather primitives. Using a chiplet-aware approach grounded in deep hardware insight, we are able to decrease the overhead from 1.5µs to 600ns.

Kevin Cochrane

CMO, Vultr

Kevin Cochrane is a 25+ year pioneer of the digital experience space. Kevin co-founded his first start-up, Interwoven, in 1996, pioneered open sourcecontent management at Alfresco in 2006, and built a global leader in digital experience management as CMO of Day Software and later Adobe. Kevin has also held senior executive positions at OpenText, Bloomreach, and SAP. Now at Vultr, Kevin is now working to build Vultr'sglobal brand presence as a leader in the independent Cloud platform market.

Tech Talk: Accelerating Business Outcomes with AMD Enterprise AI Suite

Abstract
In this talk, we explore how enterprises can move from experimentation to production applications.  We look at how AMD's full-stack AI infrastructure can empower developers to quickly build and deploy new AI-native applications that transform the customer experience and drive real revenue growth.  We'll explore the keys to unlocking developer productivity and global scalability powered by AMD EPYC, AMD Instinct, and AMD Enterprise AI Suite.

Quentin Anthony

VP of AI Engineering, Zyphra

Quentin Anthony is VP of AI Engineering at Zyphra, where he leads model AI infrastructure and systems optimization for large-scale foundation models. His research focuses on the co-design of distributed training and inference frameworks with hardware.

Tech Talk: Building Production Inference for Trillion-Parameter Models on AMD Instinct™ MI355X

Abstract
Zyphra built its foundation model training stack on AMD and is now extending that expertise to inference. In this session, Quentin Anthony walks through how Zyphra built Zyphra Cloud to serve frontier open-weight models — including DeepSeek V3.2, Kimi K2.6, and GLM 5.1 — on AMD Instinct™ MI355X GPUs.

Topics include ROCm kernel optimization, parallelism, and execution strategies for trillion-parameter MoE models, as well as systems approaches to long-context inference under real-world latency and throughput constraints.

Quentin shares lessons from training on AMD Instinct and how that co-design approach extends to inference, along with perspectives on the current optimization frontier for teams building on the AMD stack.

Anush Elangovan

Vice President, Software Development, AMD

Anush Elangovan leads the Artificial Intelligence Group (AIG) as Corporate Vice President of AI software and solutions.

Anush has 23 years of industry experience in AI, computer science, compilers, network security, operating systems, math, and its materialization on complex hardware systems. This co-founder and Chief Executive Officer of Nod.ai oversaw product strategy and the overall business until AMD acquired Nod.ai (see related article here) today.

Anush will lead the acceleration of deploying AI solutions optimized for AMD products while aligning with AMD’s AI growth strategy centered on an open software ecosystem. In the near term, he and his team will introduce the code generation (CodeGen) capabilities from the Nod.ai flagship software, Shark, to unlock customer engagements via the ROCm™ and Vitis™ AI platforms. Over time, Anush will lead the contributions of the Nod.ai team to the AMD Unified AI Stack.

Before starting Nod.ai, Anush was instrumental in the graphics stack on the first ARM Chromebook. He led the movement of the Chrome operating system from Debian to Gentoo Linux to enable Google to gain full control of the shipping software. Previously, he was Principal Engineer for Agnilux, which Google acquired. The Agnilux team became crucial to the Chrome OS team, building a fusion of Android and Chrome OS.

Previously, Anush was a technical lead at Cisco Systems in its Datacenter Group, creating the first distributed virtual switching platform. He has also been an early member of FireEye, where he led in-memory taint-check analysis for networking and security in virtualized environments. He started his career in an earlier stint at Cisco, contributing to metro ethernet initiatives. Anush holds a Master of Science in computer science from Arizona State University and a Bachelor of Engineering in computer science from the MepcoSchlenk Engineering College at Madurai Kamaraj University in India. He has earned 10 patents. In his spare time, he enjoys skiing, mountaineering, and trail running. Anush lives with his family, including three children and two dogs, in the East Bay of the San Franciso Bay Area.

Augustin Verneuil

GPU engineer, Kog

Augustin Verneuil is part of the GPU engineering team at Kog, where he works at the intersection of hardware reverse-engineering, and kernel forward-engineering. Previously, he received a PhD in Physics, awarded jointly by the Politecnico di Milano and UT Troyes. 

Tech Talk: Building a real-time inference stack on AMD Instinct GPUs

Abstract
In this talk, we share our vision for real-time generative AI, and the techniques we developed to achieve the fastest LLM inference on GPU ever, with a generation speed of 4000 tokens/s per request.

We first showcase our end-to-end stack optimized for minimal latency on AMD hardware, spanning model re-architecturing, a single monokernel implementation, along with topology-aware algorithms.

In the second part, we focus on one of the defining challenges of megakernels, intra-GPU grid synchronization barriers and reduce/gather primitives. Using a chiplet-aware approach grounded in deep hardware insight, we are able to decrease the overhead from 1.5µs to 600ns.

Jens Stapelfeldt

Global AI Lead for the AMD University Program, AMD

Jens Stapelfeldt is the Global AI Lead for the AMD University Program (AUP), where he works with universities worldwide to enable hands-on AI education using open-source software and high-performance computing platforms. His focus is on helping educators quickly deploy modern AI workflows—from large language models and agentic systems to GPU-accelerated computing—without the burden of building infrastructure from scratch.

Jens brings more than 30 years of experience in the semiconductor and microelectronics industry, spanning technical training, system architecture, and business leadership roles at AMD, Xilinx, Texas Instruments, and Doulos. He is passionate about bridging industry technology and academic learning to prepare the next generation of AI practitioners.