Tenstorrent has announced the general availability of its Galaxy Blackhole AI compute platform, introducing a new class of AI servers designed for high-performance computing.
Specifications and Pricing
The Galaxy Blackhole systems are housed in a 6U chassis and are equipped with 32 Blackhole accelerators. Each system boasts impressive specifications, including 1 TB of GDDR6 memory, 16 TB/s of memory bandwidth, and a performance capability of 23 petaFLOPS in dense FP8 operations. The price for each unit is set at $110,000, which positions it as a more affordable alternative compared to Nvidia’s offerings, which can cost three to five times more.
Scalability and Performance
Tenstorrent emphasizes the scalability of its Galaxy platform, which can be expanded to support larger models and higher throughput by adding more systems. The base Galaxy Supercluster configuration, priced at $440,000, includes four Blackhole systems and can scale up to 32 nodes with over a thousand chips. This flexibility allows users to optimize for various workloads, including interactive applications.
Software Enhancements
According to Jasmina Vasiljevic, a senior fellow at Tenstorrent, significant improvements have been made to the software stack since earlier tests. The previous limitations in model support and performance scaling have been addressed, enabling the Galaxy systems to process a 100,000 token prompt in under four seconds. The systems currently output up to 300 tokens per second per user, with plans to enhance this to 350 tokens through future software updates.
Market Adoption and Future Developments
Tenstorrent is positioning the Galaxy Blackhole not only for large language models but also for video generation tasks, claiming that a four-node supercluster can generate 720p video faster than real-time. The company has also developed a Python-based programming interface to facilitate the integration of new models. Notably, Tenstorrent claims that 90% of models from Hugging Face are compatible with its platform. The hardware is already being adopted by several major datacenter and cloud providers, including Cirrascale, Equinix, and Japan’s ai&. More details are expected to be revealed during the upcoming TT-Deploy event on May 1.
This article was produced by NeonPulse.today using human and AI-assisted editorial processes, based on publicly available information. Content may be edited for clarity and style.
