Software

AI Stack for Token Serving Zero friction: your existing models run with no changes on vLLM and SGLang, PyTorch underneath, and ONNX Runtime.

HPC Acceleration Stack Program the device directly from Linux with GCC, LLVM and OpenMP.

Your models. Your software. No lock-in.

Powerful AI hardware shouldn't mean rewriting your software.

Open, modular software stack on Semidynamics hardware

Semidynamics works with mainstream AI frameworks and development tools, making it easier to bring existing models and applications to our memory-centric inference platform.

Our software is built on two pillars: the Aliado AI stack for token serving, which plugs into the frameworks you already use, and an HPC acceleration stack that lets any developer program the device directly from Linux.

AI Stack for Token Serving

A hardware-agnostic layer that integrates with the AI ecosystem, enabling performant cross-framework support.

Aliado provides framework plugins for PyTorch, vLLM, SGLang and ONNX Runtime, so developers focus on their models while Aliado connects them to Semidynamics hardware.

Models such as Kimi K3, GLM, DeepSeek, Mistral or any other model run through vLLM, PyTorch or ONNX Runtime and the Aliado Dispatcher on the host CPU, which sends the work to the XPUs running the Aliado Kernel Library

Designed to scale

Aliado dispatches kernels across cores and devices automatically.

Workloads distributed from one orchestrator to many accelerators

SCCL collectives implement sharding across many XPUs, so models larger than a single device run across the rack.

Integration with Kubernetes and model-aware request routing supports larger deployments, including enterprise AI and shared inference infrastructure.

HPC Acceleration Stack

The set of tools that enables any user to program the device directly from Linux.

Code running directly on Semidynamics hardware

GCC and LLVM toolchains for our RISC-V ISA, with C and C++, for custom AI and high-performance computing applications.

OpenMP device offload lets developers use standard code, with no GPU primitives. OpenMP is an open standard for expressing parallel work in C, C++ and Fortran, so existing parallel code comes to our platform with less porting effort.

Visibility across the stack

Every layer of the software stack inspected down to the device

Aliado provides visibility from framework integration through to device execution, helping developers understand workloads, identify opportunities for optimisation and manage complex deployments.

Semidynamics hardware and software as a complete inference platform

Hardware and software, working together

Semidynamics brings processors, memory, interconnects and software together as a complete inference platform. Easier development, deployment and scaling of AI infrastructure.