Kernelize offers a hardware‑agnostic AI inference platform built on the Triton kernel language, letting developers write a single high‑level kernel that the Triton compiler translates into optimized code for any supported accelerator. The solution integrates seamlessly with PyTorch, supports custom kernels and operator fusion, and enables incremental onboarding of new chips without modifying application code or rebuilding toolchains.
Funding
Funding not disclosed
Founders
Product
Problem
AI inference pipelines are often tightly coupled to a single accelerator, requiring custom kernels for each new chip. This hardware lock‑in forces engineers to rewrite and retest models whenever the underlying silicon changes, fragmenting workflows and inflating development costs.
Solution
Kernelize delivers a hardware‑agnostic inference platform built on the Triton kernel language. By encapsulating device‑specific optimizations in Triton Extensions, the platform lets developers write a single, high‑level kernel that the Triton compiler translates into efficient code for any supported accelerator. The approach preserves model portability while extracting peak performance on each architecture, enabling incremental adoption of new chips alongside existing NVIDIA deployments. Integration points align with PyTorch and other standard ML frameworks, so teams can upgrade hardware without altering application code or rebuilding toolchains. Kernelize also provides a unified release cadence synchronized with official Triton and PyTorch versions, ensuring consistent compatibility across heterogeneous clusters.
Target Audience
The primary customers are machine‑learning engineers, AI infrastructure teams, and cloud service providers that run inference workloads across heterogeneous accelerator fleets and need a unified, maintainable software stack.
Features
- Stable, high‑level Triton compiler language that separates model logic from hardware‑specific code
- Triton Extensions layer to define and isolate chip‑specific optimization strategies without modifying higher‑level software
- Hardware‑aware compilation that generates device‑native kernels, supporting operator fusion and custom hand‑written kernels (e.g., matrix multiplication, SoftMax)
- Seamless PyTorch integration via upstream Triton support, eliminating the need for forks or proprietary APIs
- Incremental hardware onboarding that allows new accelerators to be added to existing clusters without workflow disruption
- Release alignment with official Triton and PyTorch versions to guarantee binary compatibility and up‑to‑date performance improvements
- Tooling for portable kernel development, including tensor descriptor abstractions that abstract away device memory layouts