Tenstorrent’s Enhanced AI Model on Blackhole Servers Produces 5-Second Videos in Only 2.4 Seconds

Tenstorrent’s Enhanced AI Model on Blackhole Servers Produces 5-Second Videos in Only 2.4 Seconds

Tenstorrent has unveiled its revolutionary AI model, designed to leverage the power of Blackhole servers for exceptionally rapid video generation.

Revolutionary AI Video Generation with Tenstorrent

Traditionally, generating AI-driven videos is a time-consuming process. However, the speed of processing largely depends on the underlying hardware. While advancements have certainly improved efficiency, Tenstorrent is setting a new benchmark.

In a recent demonstration, Tenstorrent showcased its latest AI model, referred to as Wan2.2-14B. This model has been meticulously trained on a vast dataset of 14 billion parameters, utilizing a network of hundreds of Blackhole accelerators to achieve phenomenal performance.

During a live event reported by EETimes, the model successfully generated a 720p video, consisting of 81 frames over 40 steps, in just five seconds. Impressively, this generated video was also five seconds long. Tenstorrent further claims to have achieved a record-breaking time of 2.4 seconds for a similar video clip, outpacing competitors by a remarkable 10 times.

A chart titled 'Tenstorrent: 10x Faster Real-Time AI Video Gen' illustrates the performance of the Wan 2.2 A14B, showing significantly improved fps compared to other platforms.
Image Source: EETimes

The Wan2.2-14B AI model operates on an optimized framework created by Prodia, running across four Galaxy servers equipped with a total of 256 Blackhole accelerators. Each accelerator boasts 16 RISC-V cores and 32 GB of GDDR6 memory, ensuring superior AI processing capabilities. This setup represents the pinnacle of Tenstorrent’s AI technology.

Crucially, Tenstorrent has achieved these groundbreaking results by adopting a unified architecture strategy. This innovative approach eliminates reliance on proprietary technology, favoring an open compute platform where cores, memory, and networking collaboratively enhance the performance of large-scale AI models.

This demonstration is just a glimpse of what’s to come. Tenstorrent is gearing up to present additional performance showcases next week in anticipation of launching its next-generation cluster-scale systems, making it an exciting time for AI enthusiasts and industry experts alike.

Source & Images

Leave a Reply

Your email address will not be published. Required fields are marked *