Macnica and Fixstars have established a framework to jointly provide end-to-end support for Tenstorrent AI infrastructure, from validation including hands-on evaluation through performance optimization and post-deployment operations. The companies also disclosed the results of a joint PoC using a TT-QuietBox equipped with 8 Wormhole chips.
Macnica and Fixstars announced that they had established a framework to provide domestic companies considering the deployment of Tenstorrent AI products with end-to-end support, from hands-on evaluation of AI models through performance optimization, AI infrastructure development, deployment and operations. This will enable customers to coordinate validation before deployment with operations after deployment.
The joint PoC used the TT-QuietBox AI workstation, equipped with 8 Tenstorrent Wormhole chips and 96 GB of total GDDR6 memory, to run OpenAI's open-weight large language model gpt-oss-120b and analyze inference performance. Macnica provided the hands-on evaluation environment and support, while Fixstars was responsible for analyzing processing conditions and improving the software implementation.
The PoC accelerated the MoE expert matrix multiplication that had been a bottleneck in decoding by 3.68x. TPOT was reduced by 31.3%, from 144.85 ms to 99.49 ms, while decoding speed per user improved from 6.9 tokens/second to 10.05 tokens/second. Average response time across all requests was also reduced from 25.0 seconds to 19.3 seconds.
The two companies plan to expand the application to AI models and workloads other than gpt-oss-120b, supporting customers from pre-deployment validation through performance optimization, infrastructure development and operations. They also plan to pursue technical collaboration with Tenstorrent as necessary, contributing to the broader adoption of AI by domestic companies.