OXMIQ CEO Raja Koduri on Re-Architecting the GPU Stack: From Atoms to Agents

OXMIQ CEO Raja Koduri on Re-Architecting the GPU Stack: From Atoms to Agents

by Raja Koduri

Applied Sciences & Engineering Computing & Cybersecurity UYComputer scienceUYFComputer architecture and logic design
🎙 Raja Koduri 👥 5K 📅 February 4, 2026 ⏱ 128 min 👁 2K 📄 expert opinion 🧭 2026-08-16
Available in: English (current) Français

Keywords

GPUAIchipletsoftware stackbandwidth

Summary

In this episode of the Semiconductor Leadership Podcast, host Salah Nasri interviews Raja Koduri, founder and CEO of Oxmiq Labs, about re-architecting the GPU stack from atoms to agents. Koduri discusses the challenges and opportunities in AI silicon, emphasizing that bandwidth, not compute, is the true bottleneck. He introduces Oxmiq’s solutions: OXCORE, a unified computing core that integrates scalar, vector, and matrix processing, and OXPython, a Python-level intervention to address the ‘CUDA virus’ that locks software to Nvidia hardware. He critiques the overemphasis on FLOPS as a metric, advocating for real transistor utilization. He shares insights from his experience with Intel’s Ponte Vecchio, highlighting the complexities of chiplet design and the need for close collaboration. Koduri argues that the silicon budget for AI infrastructure is over $30 billion per gigawatt, necessitating cost reductions to make AI accessible globally. He discusses the importance of hardware-software co-design, the trade-offs between freedom and fragmentation in chiplet standards, and his commitment to spending transistors to simplify software. The conversation covers career lessons, the timing of Oxmiq, and his vision for the future of AI compute.

181 words

Critical Evaluation

Value of the Information & Strength of the Argument

The value of the information is high, offering unique insights from a leading industry figure on the future of GPU architecture and AI infrastructure. Koduri provides concrete examples and data, such as the $30 billion per gigawatt silicon cost and the projected token demand, to support his arguments. His argumentation is solid, grounded in his extensive experience at Intel and AMD, and he clearly articulates the challenges of hardware-software co-design and the limitations of current standardization efforts. He presents a compelling case for Oxmiq’s approach, though some claims are forward-looking and not yet fully validated.

Scientific Rigor, Source Quality, Title Accuracy

The discussion is rigorous in its technical depth, with Koduri referencing specific projects like Ponte Vecchio and partnerships like Tenstorrent. However, the content is primarily opinion and experience-based, with no formal citations to external sources. The title accurately reflects the content, focusing on re-architecting the GPU stack. The podcast format allows for a candid and detailed exploration of the topics, but the lack of verifiable sources limits its scientific rigor. The description provides chapter markers that help navigate the content, but no additional references are given.

196 words

Title / Content Match

The title accurately reflects the content, focusing on re-architecting the GPU stack from hardware to software, with a vision from atoms to agents.

Quality & Reliability

8/10

The speaker is a highly experienced industry leader with deep technical expertise in GPU architecture and software stacks. The discussion is grounded in practical experience and specific examples, though it is primarily opinion and forward-looking vision rather than peer-reviewed research.

Chapters

Contribution & Novelties

This podcast provides a unique perspective on the future of GPU architecture from a key industry insider. Koduri’s concept of OXCORE as a unified computing core and OXPython as a software-level solution to the ‘CUDA virus’ offers novel approaches to the challenges of AI hardware and software. His emphasis on bandwidth over compute and the need to reduce silicon costs are important contributions to the discussion on AI infrastructure scalability.

Pour aller plus loin :

  • CUDA — The proprietary parallel computing platform and API by Nvidia, central to the discussion.
  • Chiplet — The concept of modular semiconductor design, relevant to the chiplet discussion.
  • Ponte Vecchio — Intel’s data center GPU, which Koduri led, providing context for his insights.

118 words

Radar Profile

The radar profile shows high scores in quantity of information, quality, and technical level, reflecting the depth of the discussion. The reliability score is slightly lower due to the opinion-based nature of the content, but overall the podcast offers valuable insights for professionals in the semiconductor and AI fields.

Reliability 7/10