Raja Koduri, the former chief architect of Intel’s GPU department, is embarking on an exciting new venture at Oxmiq Labs. With a vision of revolutionizing AI GPU development, he is assembling an elite team of architects to create a fresh GPU ecosystem entirely from scratch.
After retiring from Intel and with a rich background at AMD, Koduri is steering Oxmiq Labs toward producing groundbreaking AI GPUs. The company, built on 500 years of collective expertise and armed with numerous patents, plans to innovate radically within the GPU landscape.
Rather than simply adapting existing designs, Oxmiq Labs aims for a comprehensive overhaul. With its “Atoms to Agents” approach, the initiative focuses on both low-level hardware and high-level AI systems, emphasizing flexibility and efficiency.
Given the growing demand for GPUs capable of processing diverse data in real time, Oxmiq’s new architecture seeks to enhance performance and efficiency significantly. By introducing innovations like the OXCORE, a modular GPU core accommodating Scalar, Vector, and Tensor operations, the company plans to support everything from nano-agents to Python acceleration and CUDA/SIMD compatibility. This scalability will benefit devices ranging from small gadgets to vast data centers.
The OXCORE will leverage the inventive OXQUILT chiplet architecture, facilitating the creation of bespoke GPUs by blending compute, memory, and interconnect components efficiently. This approach promises reductions in research, development, and production costs compared to conventional GPUs.
A cornerstone of Oxmiq’s strategy is its “Software First” approach, ensuring a user-friendly software stack. This stack will be compatible with Oxmiq’s GPU IP as well as third-party AI hardware like Tenstorrent.
OXPython, Oxmiq’s flagship software, is set to transform NVIDIA CUDA applications, enabling them to run on non-NVIDIA hardware seamlessly, initially targeting Tenstorrent AI chips. The unified software ecosystem, OXCapsule, will simplify AI app deployment across various platforms by abstracting hardware complexities.
Oxmiq Labs is adopting a licensing-first model, offering IP licenses without the need to produce complete GPUs, thus sidestepping the costly tape-out process. With $20 million in seed funding from backers like Mediatek, the future looks promising. Raja Koduri’s initiative marks the first GPU startup in Silicon Valley in over two decades, setting the stage for a new era of computational flexibility.
