Eine GPU-Start-up-Firma taucht auf, gegründet von Raja Koduri und basierend auf RISC-V.

Recently, the legendary GPU architect Raja Koduri, who previously worked at ATI, AMD, Apple, and Intel, announced the establishment of his GPU company, Oxmiq Labs.

As Raja Koduri explained on his social media, Oxmiq Labs Inc. is a new startup for GPU software and IP that has emerged from obscurity. "We've assembled a world - class team of GPU and AI architects with a combined experience of over 500 years. OXMIQ's licensable IP rebuilds the GPU from the ground up," said Raja Koduri.

"We want to become the next - generation Arm," Raja Koduri said in an interview with Reuters.

Oxmiq Labs stated that the company takes a software - first approach to develop its chip designs and has developed a tool that allows software programs written for Nvidia's CUDA to run on non - Nvidia hardware "without changing the code or recompiling."

"We might be the first new GPU startup in Silicon Valley in over 25 years. GPUs aren't easy," said Raja Koduri.

A GPU Based on RISC - V

According to Oxmiq Labs, modern computation has fundamentally shifted towards multimodal experiences where text, audio, video, images, and 3D environments interact seamlessly. GPU architecture is the cornerstone of this change. Unlike fixed AI accelerators for specific tasks, the GPU provides the necessary general computing flexibility for these different modalities while remaining deeply integrated into mainstream operating systems via standardized APIs and a unified memory model. This architectural advantage makes the GPU an indispensable computing machine for current applications and the emerging field of multimodal AI, where heterogeneous workloads need to be processed in a coordinated manner.

Based on this understanding, OXMIQ focuses on developing GPU hardware and software IP and licenses it to interested parties.

Oxmiq explained that the licensable GPU IP re - architects the GPU from first principles and integrates groundbreaking technologies, including silicon nano - agents that utilize RISC - V cores, near - memory and in - memory computing, and optical transmission. The solutions offered by OXMIQ to customers achieve a balance between multimodal computing flexibility and the fundamental performance boost required for the next generation of graphics and AI workloads.

In the hardware area, Oxmiq offers a GPU IP core called OxCore, which is based on the RISC - V instruction set architecture (ISA). This core integrates scalar, vector, and tensor computing units in a modular architecture and supports near - memory and in - memory computing functions.

Oxmiq also offers the chiplet - based System - on - a - Chip (SoC) builder OxQuilt, which allows customers to quickly and cost - effectively create SoCs that contain integrated Computation - Cluster - Bridge (CCB, possibly integrated with OxCores), Memory - Cluster - Bridge (MCB), and Interconnect - Cluster - Bridge (ICB) modules, depending on specific workload requirements.

For example, an inference AI accelerator for edge applications may contain one or two CCBs and an ICB, an inference SoC needs more CCBs, MCBs, and ICBs, while a large SoC for AI training may potentially contain dozens of chiplets. Oxmiq has not yet revealed whether its OxQuilt only supports the creation of multi - chip set System - in - Package (SiP) or can also be used for the assembly of single - die processors.

As the company explained on its website, OXMIQ also offers a complete GPU hardware IP stack that supports various silicon solutions from edge devices and autonomous robots to enterprise - edge infrastructures and zettabyte data centers. Their scalable GPU core OXCORE integrates scalar, vector, and tensor computing units in a modular architecture that can be tailored to specific workloads to support nano - agents, native Python acceleration, and compatibility with the SIMT/CUDA paradigm.

OXCORE can be seamlessly scaled via the OXQUILT chiplet architecture, from a single core for compact physical AI applications to thousands of cores for data centers. With OXQUILT, customers can configure the optimal ratio of computing power, memory, and interconnect according to their needs, significantly shorten the time to market, and reduce research and development costs. Compared to current industry standards, the time to market is greatly reduced.

Software Is the Core

Oxmiq Labs emphasizes that OXMIQ has learned from decades of industry experience and adopts a software - first strategy. Through a comprehensive software stack that is compatible with silicon based on OXMIQ IP as well as third - party GPU and AI accelerator platforms, the developer experience is prioritized.

As reported, this software package is supposed to abstract the complexity of heterogeneous hardware and enable the deployment of AI and graphics workloads on different hardware platforms, not just those using the company's IP. OXCapsule is the core of the entire software ecosystem, a unified software ecosystem of OXMIQ that abstracts hardware complexity and enables seamless deployment on different computing platforms by eliminating the configuration problems traditionally encountered in heterogeneous environments.

Specifically, this is a unified runtime and scheduling layer that manages workload distribution, resource utilization, and hardware abstraction. This layer encapsulates applications into independent environments that the company calls "heterogeneous containers." These containers are designed to run independently of the underlying hardware, allowing developers to develop for CPU, GPU, and AI accelerators without changing their codebase or worrying about the underlying configuration.

OXPython is the flagship component of OXMIQ. It enables Python - based NVIDIA CUDA AI applications to run seamlessly on non - NVIDIA hardware without changing the code or recompiling. OXPython will be first introduced on Tenstorrent's AI platform later this year and is currently being integrated with several providers. OXPython demonstrates OXMIQ's determination to break down hardware silos and accelerate the spread of high - performance computing across the industry.

In other words, it is a compatibility layer that converts CUDA - centric workloads into Oxmiq's runtime and allows Python - based CUDA applications to run on non - Nvidia hardware without changes and without recompilation.

It is known that OXPython will not be initially released 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 of Oxmiq's hardware, which is a key part of their strategy.

Jim Keller, CEO of Tenstorrent, also said: "We are very excited to work with OXMIQ on their OXPython software stack. OXPython can bring CUDA's Python workloads to AI platforms like Wormhole and Blackhole, which is very beneficial for developer portability and ecosystem expansion. This aligns with our goal of keeping and owning the entire AI stack open for developers."

The company emphasizes on its website that Oxmiq Labs focuses on re - architecting the entire GPU stack architecture from atoms to agents to meet the requirements of the next generation of games, graphics, and multimodal AI. The company develops licensable GPU hardware and software IP and achieves a balance between flexibility and performance by integrating groundbreaking technologies, including RISC - V - based silicon nano - agents, near - memory and in - memory computing, advanced optical transmission, and other innovative technologies. The OXMIQ architecture is designed to be seamlessly scalable from physical AI devices to data - center scale.

MediaTek Has Invested

Besides Raja Koduri, the company says it has brought together an extremely competent team.

It's worth noting that Oxmiq Labs doesn't manufacture consumer GPUs like AMD Radeon or Nvidia GeForce. Unlike Arm or Imagination Technology, Oxmiq also doesn't develop all the necessary IP modules for GPU manufacturing: it doesn't support all complete consumer graphics functions from the start (e.g., texture units, rendering backend, display pipeline, ray - tracing hardware, DisplayPort or HDMI output). So if Oxmiq licensees want to build a GPU, they have to implement these functions into the chip themselves.

Based on this leading team and experience, the company has already received $20 million in a seed round from renowned tech investors (including strategic players in the mobile phone and artificial intelligence fields like MediaTek) and has already generated first - time revenue from software. OXMIQ emphasizes that with its license - first model, it avoids the high capital requirements of chip startups that rely on expensive EDA tools and physical chip manufacturing, thus achieving excellent capital efficiency.

Lawrence Loh, Senior Vice President of MediaTek, said: "OXMIQ has an impressive vision and a world - class team. The company's GPU IP and software innovations will usher in a new era of computing flexibility across different devices - from mobile phones to cars and edge AI."

This article is from the WeChat public account "Semiconductor Industry Observation" (ID: icbank), Author: Editorial, published by 36Kr with permission.