Groundbreaking startup founded by GPU mastermind Raja Koduri utilizes RISC-V architecture and aims to execute CUDA workloads, enabling direct operation of unaltered Python-based CUDA applications on hardware apart from Nvidia. Oxmiq Labs in action.
Oxmiq Labs, a Silicon Valley-based startup founded by renowned GPU architect Raja Koduri, is making waves in the GPU industry with its innovative chiplet-based system-on-chip (SoC) builder, OxQuilt. The company's modular and scalable architecture, OXCORE, promises to disrupt the traditional GPU market by reducing capital barriers, providing unprecedented architectural flexibility, and enabling vendor-agnostic AI software compatibility.
OxQuilt allows the integration of compute cluster bridge (CCB), memory cluster bridge (MCB), and interconnect cluster bridge (ICB) modules, creating a customizable SoC that caters to diverse workloads from edge devices to large data centers. The design leverages RISC-V nano-agents, near-memory and in-memory computing, and integrates scalar, vector, and tensor compute engines within a single modular GPU core. This approach eliminates the need for costly monolithic chip designs, potentially reducing R&D costs and time-to-market by over 90% compared to traditional GPU development.
Key software innovations include a software-first ecosystem (OXCapsule) with a standout feature, OXPython. OXPython enables unmodified Python-based NVIDIA CUDA applications to run seamlessly on non-NVIDIA hardware, breaking vendor lock-in and allowing cross-platform AI computing without code rewriting. This addresses a major limitation of traditional CUDA-dependent frameworks.
Differences from traditional GPU offerings include a licensing-first model, chiplet modularity, RISC-V based cores, near-memory and in-memory computing, and cross-vendor software compatibility. By licensing GPU IP and software stacks rather than producing full chips, Oxmiq avoids the huge capital investment typically required for cutting-edge GPU designs. The OXQUILT chiplet design allows assembling customized SoCs optimized for specific workloads, from single-core AI inference at the edge to multi-thousand-core data center accelerators.
The architecture uses RISC-V cores as nano-agents within the GPU, a departure from traditional proprietary ISA designs like NVIDIA’s CUDA cores. Integrating compute capabilities close to or within memory helps boost AI workload efficiency, which is less common in conventional GPUs. Cross-vendor software compatibility is achieved through OXPython, allowing running CUDA applications on different hardware without modification, disrupting the commonly proprietary software ecosystems that lock users to vendor hardware.
Oxmiq Labs' GPU innovations have garnered praise from industry leaders. Lawrence Loh, SVP of MediaTek, has expressed excitement about Oxmiq's GPU IP and software innovations, stating that they will drive a new era of compute flexibility across devices, from mobile to automotive to AI on the edge. Tenstorrent's CEO, Jim Keller, has also expressed excitement about partnering with Oxmiq on their OXPython software stack.
The company has already recorded its first software revenue and secured $20 million in seed funding from major tech investors, including mobile and custom AI silicon developer MediaTek. Oxmiq Labs is not developing a consumer GPU like AMD Radeon or Nvidia GeForce, but rather focusing on developing GPU hardware and software IP for licensing to interested parties.
With Oxmiq Labs, the GPU market is set for a significant shift, moving away from traditional monolithic, hardware-locked GPU models dominated by a few vendors towards a more open, flexible, and vendor-agnostic ecosystem.
Data-and-cloud-computing advancements are further propelled by Oxmiq Labs' innovative GPU architecture, as its modular design caters to diverse workloads in both edge devices and large data centers. The integration of artificial-intelligence capabilities, such as the cross-platform OXPython software, disrupts traditionally proprietary AI software ecosystems, promoting technology agnosticism in the data-and-cloud computing landscape.
