Macnica and Fixstars partner to support Tenstorrent AI deployment

Macnica and Fixstars have partnered to provide comprehensive support for Japanese companies looking to implement Tenstorrent AI products. The collaboration covers the entire lifecycle of AI deployment, including hardware selection, performance evaluation, model porting, infrastructure construction, and ongoing operational support.

To demonstrate the effectiveness of this partnership, the companies conducted a joint Proof of Concept (PoC) using the Tenstorrent TT-QuietBox AI workstation. They focused on optimizing the inference performance of the OpenAI large language model ‘gpt-oss-120b’.

The PoC yielded significant performance improvements, including a 3.68x speedup in the expert matrix multiplication for Mixture of Experts (MoE). Additionally, the time per generated token (TPOT) was reduced by 31.3%, decoding speed per user increased by 45.7%, and the average end-to-end response time was shortened by 23.0%. The companies intend to expand this support to various other AI models and workloads to facilitate broader corporate AI adoption.