Beyond GPU IP: How Oxmiq Labs Is Opening Up Custom AI Compute

Licensing a CPU core is a familiar starting point for engineers developing a custom chip. But what happens when the application needs high-performance GPU computing? In this ipXchange interview, Adam Yap speaks with Raja Koduri, founder and CEO of Oxmiq Labs, about the gap his company aims to fill, and why the answer involves much more than silicon IP.

Koduri draws a distinction between established mobile graphics IP and the GPU compute architectures used for demanding AI workloads. His ambition is to give engineering teams access to configurable compute IP, alongside the design tools and software needed to turn it into a usable system.

At the centre of that approach is OxCore. It combines scalar, CPU-style computing with parallel GPU execution and tensor processing for AI. Rather than offering one fixed balance of these capabilities, Oxmiq allows customers to configure the architecture around their workloads. The aim is to give teams a foundation for custom silicon without requiring them to develop the entire compute architecture themselves.

That core is only one part of the design challenge. Engineers also need to decide how much memory their application requires, how quickly data must move and how the different parts of the system connect. OxQuilt addresses these questions through chiplet design tools and modelling. In the interview, Koduri uses a robot as an example: its algorithms determine the compute and bandwidth requirements, while real-time targets help establish whether a proposed configuration can deliver.

This also introduces one of the interview’s strongest points: the best architecture on paper may depend on components that a team cannot obtain. Koduri argues that engineers must account for supply availability alongside performance, power and cost. An alternative memory technology may bring compromises, but designing around those limitations could provide a more practical route to production. OxQuilt helps teams explore the combinations of memory, compute, packaging and interconnects behind those decisions.

The software story extends Oxmiq’s relevance beyond companies building chips. Through OxCapsule, the company is working to make AI workloads easier to deploy across different hardware platforms. Koduri describes a demonstration in which a single command launches a large AI model across a network of PCs. This matters because useful compute resources may already exist across several machines, yet coordinating them can be a substantial task. It also means customers can benefit from Oxmiq’s software without first adopting its silicon IP.

At the time of the interview, Koduri reports that OxCore had reached FPGA validation, with AI models executing without code changes. This represents progress towards silicon implementation, while SDKs, documentation and evaluation platforms form part of the planned support for customers.

Across the conversation, the common thread is engineering choice: how to configure a chip, which components to build around and where to run the software. Oxmiq’s proposition connects those decisions, bringing GPU IP, chiplet design and AI deployment into the same discussion.

Watch the full ipXchange interview with Raja Koduri to hear how Oxmiq Labs is approaching each part of that challenge.