Partnering to Advance Performant Full-Model Inference

Infinity and d-Matrix today announced a partnership focused on enabling and optimizing performant full-model inference on d-Matrix Corsair.

Chip software matters: the gap between what inference hardware can deliver and what teams actually unlock in practice is where significant performance is often lost. Infinity and d-Matrix today announced a partnership focused on closing the developer’s gap, and enabled Infinity developers to optimize inference on d-Matrix Corsair, from first kernels to full-model inference in mere weeks.

The two teams have worked side by side, moving from the first tensor-parallel matrix operations to complete, stateful Qwen3 inference on one Corsair card. The resulting implementation spans the full 28-layer model, a prefill-decode program with multiple concurrent users, and persistent on-chip weight placement.

The partnership combines Infinity’s model-aware performance engineering and automated optimization systems with d-Matrix’s memory-centric Corsair accelerator and Aviator software platform.

What We’ve Achieved Together

Agentic, End-to-End Enablement

The collaboration began by enabling large matmuls with tensor parallelism. Those operations have regular shapes and enough arithmetic intensity to benefit from broad distribution across Corsair’s hardware gangs.

Full-model inference required a different strategy for each part of the workload. Together, the two teams:

• Split 28 decoder layers across the card’s two packages
• Distributed dense computation across 16 gangs
• Kept sequence-specific KV caches, masks, and runtime positions local to their owning slices

This hybrid approach spreads regular compute while localizing dynamic state. It creates deliberate communication boundaries instead of paying for data movement after every operation.

Orchestrating SRAM Inference, Together

Custom AI accelerators offer compelling performance per watt and latency advantages. But the developer friction has historically been the catch.

Teams often spend more time fighting the platform than optimizing for it. d-Matrix’s Corsair is designed around a different premise: the hardware’s architecture should be legible, and the software stack should make the hard parts tractable.

The architecture makes compute and dataflow a first-class, explicit part of model execution rather than something the runtime manages opaquely. The d-Matrix Model Builder SDK exposes this through a Python-level programming model that makes it easy for agentic systems to build and iterate on.

For Qwen3, we load the model’s weights into SRAM and keep them resident across prefill and decode. Persistent weights and KV state coexist with activation and scratch regions that can be safely reused as layers and iterations complete. The result removes recurring weight traffic from the latency-sensitive decode path.

Enhanced Tools and Ecosystem

The partnership also expands the reusable software path around Corsair. Full-model integration drove extensions to the d-Matrix kernel library for sharded operations and communication patterns, and to the Model Builder SDK for persistent weight bindings, runtime-controlled state, and prefill-decode execution.

Alongside this, the teams built and extended a supporting toolchain:

• Microbenchmarks and profiling to map real hardware behavior
• A model layout engine and its accompanying linter to plan and verify physical layouts
• An enhanced debugger to inspect intermediate state and localize numerical errors across the chip’s address spaces

These capabilities create a foundation for continued optimization, larger models, longer contexts, and faster enablement of future workloads on Corsair.

Together, Infinity and d-Matrix are making it faster and easier to deploy production-ready AI models on Corsair. By combining Infinity’s hardware-aware optimization systems with d-Matrix’s inference platform, customers can bring new models online in weeks instead of months.