Legendary GPU architect Raja Koduri's new startup leverages RISC-V and targets CUDA workloads — Oxmiq Labs supports running Python-based CUDA applications unmodified on non-Nvidia hardware

Raja Koduri, a legendary GPU architect from ATI Technologies, AMD, Apple, and Intel, on Tuesday said he had founded a new GPU startup that emerged from stealth mode today. Oxmiq Labs is focused on developing GPU hardware and software IP and licensing them to interested parties. In fact, software may be the core part of Oxmiq's business as it is designed to be compatible with third-party hardware.

Another RISC-V-based 'GPU' for AI

Oxmiq develops a vertically integrated platform that combines GPU hardware IP with a full-featured software stack aimed at AI, graphics, and multimodal workloads where explicitly parallel processing is beneficial. On the hardware side, Oxmiq offers a GPU IP core based on the RISC-V instruction set architecture (ISA) called OxCore, which integrates scalar, vector, and tensor compute engines in a single modular architecture and can support near-memory and in-memory compute capabilities.

Oxmiq also offers OxQuilt, a chiplet-based system-on-chip (SoC) builder that enables customers to create their own SoCs that integrate compute cluster bridge (CCB, which probably integrates OxCores), memory cluster bridge (MCB), and interconnect cluster bridge (ICB) modules based on specific workload requirements in a rapid and cost-efficient manner. For example, an inference AI accelerator for edge applications can pack a CCB and an ICB or two, an inference SoC requires more CCBs, MCBs, and ICBs, whereas a large-scale SoC for AI training can pack dozens of chiplets. Oxmiq does not disclose whether its OxQuilt enables building only multi-chiplet system-in-packages (SiP), or is designed to assemble monolithic processors too.

Software is the key

Oxmiq's software stack is perhaps an even more important product that the company has to offer. The software package is designed to abstract the complexity of heterogeneous hardware and enable deployment of AI and graphics workloads across a range of hardware platforms, not just those using the company's IP. The core of the software stack is OXCapsule, a unified runtime and scheduling layer that manages workload distribution, resource balancing, and hardware abstraction. The layer encapsulates applications into self-contained environments, which the company calls 'heterogeneous containers.' These containers are designed to operate independently of the underlying hardware, enabling developers to target CPUs, GPUs, and AI accelerators without modifying their codebase or dealing with low-level configuration.

A standout component of this stack is OXPython, a compatibility layer that translates CUDA-centric workloads into Oxmiq's runtime and allows Python-based CUDA applications to run unmodified on non-Nvidia hardware without recompilation. OXPython will first launch not on Oxmiq's IP, but on Tenstorrent's Wormhole and Blackhole AI accelerators. In fact, Oxmiq's software stack is fundamentally designed to be independent from Oxmiq hardware, and that is a core part of its strategy.

"We are excited to partner with Oxmiq on their OXPython software stack," said Jim Keller, CEO of Tenstorrent. "OXPython's ability to bring Python workloads for CUDA to AI platforms like Wormhole and Blackhole is great for developer portability and ecosystem expansion. It aligns with our goal of letting developers open and own their entire AI stack."

What about graphics?

Having developed graphics processors at S3 Graphics, ATI Technologies, AMD, Apple, and Intel, Raja Koduri is primarily known as a GPU developer. In fact, he even positions Oxmiq as the first GPU startup in Silicon Valley in decades.

"We may be the first new GPU startup in Silicon Valley in 25+ years," wrote Koduri in an X post. "GPUs are not easy."

However, it should be noted that Oxmiq is not building a consumer GPU like AMD Radeon or Nvidia GeForce. In fact, it does not develop all the IP blocks necessary to build a GPU, unlike Arm or Imagination Technology: it does not support full consumer graphics features out-of-the-box (such as texture units, render back ends, display pipeline, ray tracing hardware, DisplayPort or HDMI outputs), so Oxmiq licensees must implement them in silicon themselves, if they plan to build a GPU.

Asset low strategy

Oxmiq has secured $20 million in seed funding from major tech investors, including mobile and custom AI silicon developer MediaTek, and has already recorded its first software revenue. By focusing on IP licensing instead of costly chip production or even actual silicon implementation, the company maintains high capital efficiency without relying on expensive EDA tools or tape-outs.

"Oxmiq has an impressive bold vision and world-class team," said Lawrence Loh, SVP of MediaTek. "The company's GPU IP and software innovations will drive a new era of compute flexibility across devices, from mobile to automotive to AI on the edge."

Follow Tom's Hardware on Google News to get our up-to-date news, analysis, and reviews in your feeds. Make sure to click the Follow button.