FuriosaAI SDK 2026.3 drops with TCL
South Korean AI semiconductor startup FuriosaAI has unveiled SDK 2026.3, introducing the Tensor Contraction Language (TCL) framework designed to accelerate the deployment of new AI models on its flagship RNGD Neural Processing Unit (NPU). This release focuses on scalability improvements, addressing the rapid evolution of the open AI model ecosystem by optimizing tensor contractions—the mathematical core of modern deep learning—and simplifying the compile-to-deployment pipeline for developers.
By natively optimizing tensor contractions through a dedicated compiler framework rather than relying on static matrix multiplication libraries, FuriosaAI is positioning its RNGD NPU as a highly flexible and cost-effective alternative to NVIDIA GPUs for modern transformer models.
* The introduction of TCL helps bridge the gap between rapidly shifting model architectures and specialized hardware compiler requirements, speeding up time-to-market for new models.
* With RNGD's 48GB HBM3 memory and low 180W power envelope, the SDK 2026.3 enables high-efficiency inference of major open weights models.
* Competing with NVIDIA's CUDA software ecosystem remains FuriosaAI's biggest hurdle, making developer-friendly compiler tools like TCL crucial for adoption.
DISCOVERED
3h ago
2026-07-01
PUBLISHED
5h ago
2026-06-30
RELEVANCE
AUTHOR
SVTrivo