TAOS: Tsavorite Agentic Operating Stack
TAOS, the Tsavorite Agentic Operating Stack, is a complete, self-improving software platform for the OPU. It runs AI models, agents, and harnesses unchanged, from PyTorch and vLLM to Kubernetes, and continuously optimizes performance from kernels to fleet.
Your agents and models just work
TAOS meets developers where they are. The Fusion Compiler and runtime need no developer optimization, the SDK automatically converts CUDA kernels, and TAOS libraries expose a unified memory hierarchy with no data movement, so existing workloads run out of the box.
The TAOS Advantage
One stack, zero friction
From firmware to frameworks, TAOS delivers everything needed to develop, deploy, and operate AI on the OPU, with security and observability built in at every layer.
Run everything unchanged
Deploy PyTorch, Triton, vLLM, Ollama, and Kubernetes workloads as-is. No code changes and no proprietary lock-in.
Continuously self-improving
An agentic optimizer plans, experiments, and validates tuning automatically, so performance improves over time.
Secure and observable by design
Secure boot, trusted execution, and confidential compute, with full-stack telemetry, profiling, and Redfish management.
A complete AI operating stack
TAOS spans every layer, from firmware and a hardened Linux OS to the Fusion Compiler, runtime, and platform software. User code in PyTorch or OpenAI Triton compiles through MLIR and LLVM IR to AI OmniCores and Arm cores, while the runtime handles memory allocation, graph scheduling, and eager execution automatically.
Fusion Compiler and runtime
The Fusion Compiler lowers standard framework code through MLIR and LLVM IR to the OPU’s AI OmniCores and Arm cores with no developer optimization required. Custom Triton kernels are supported, and the SDK automatically converts CUDA kernels. TAOS libraries provide linear algebra, collectives, and deep learning primitives over a unified memory hierarchy with no data movement.
Agentic optimization and observability
A closed loop observes, diagnoses, optimizes, validates, deploys, and learns, continuously tuning kernels, scheduling, and system behavior against your workload intent and SLAs. The result is higher tokens per second, lower P99 latency, better tokens per watt, and faster deployment, with fleet-level learning that shares improvements across systems. Full visibility and profiling are built in.
Use cases
Whether you are serving models in production, fine-tuning privately, or orchestrating agents across a fleet, TAOS supports the complete AI lifecycle with the tools and frameworks your team already uses.
Built for developers, ready on day one
Zero
code changes required
1 -click
model deployment
15 +
frameworks supported
10 +
TAOS developer tools
Resources
Our architecture provides unmatched computational efficiency, allowing intricate AI models to operate with low latency and high throughput.
Beyond GPUs: why we built Tsavorite
Our view is that AI has outgrown the way the industry builds computers for it. Adding more of the same hardware is no longer the answer. AI needs an architecture designed for it from the start, one that becomes more efficient as it grows rather than less.
Tsavorite Scalable Intelligence Sets New Frontier for AI Efficiency and Scale with the Omni Processing Unit
Bring your models. Keep your code.
See how TAOS runs your models, agents, and frameworks unchanged on the OPU. Talk to us about SDK access and early programs.
