Tenstorrent's Blackhole Servers Generate Video
- Tenstorrent demonstrated an optimized AI model running on Blackhole servers that produces short videos faster than real time. - The setup reportedly generates a 5-second video in just 2.4 seconds, highlighting massive inference speed improvements. - The result underscores Tenstorrent's server advantage for media workloads and merits further independent benchmarking (wccftech.com).
Making video with artificial intelligence usually takes longer than the clip itself. Tenstorrent said its Blackhole-based system can turn a text prompt into a five-second video in as little as 2.4 seconds. (eetimes.com) EE Times reported that a hands-on demo on April 21 generated a 720p clip with 81 frames and 40 sampling steps in about three seconds. Tenstorrent said its best run on the same setup reached 2.4 seconds. (eetimes.com) The demo used an optimized version of Wan2.2-14B, a 14 billion-parameter video model, built with partner Prodia. Tenstorrent ran it on four Galaxy servers using 128 Blackhole chips, according to EE Times. (eetimes.com) A video model works by predicting one frame after another from a text prompt, then refining the result over repeated passes. Those repeated passes are why video generation has been slower and more expensive than text generation on most artificial intelligence hardware. (eetimes.com) Tenstorrent is trying to sell a different kind of artificial intelligence stack: chips, servers, and software built around RISC-V processors and an open-source toolchain. On its Galaxy product page, the company says the system is built for both training and inference and “leads the industry” in video generation, decode, and prefill benchmarks. (tenstorrent.com) The Blackhole chip is the company’s current high-end processor for this push. Tenstorrent says each Blackhole board has 120 Tensix cores, 16 “big RISC-V” cores, up to 32 gigabytes of GDDR6 memory, and up to 664 teraFLOPS in Block FP8 math. (tenstorrent.com) At the server level, Tenstorrent says one Galaxy Blackhole system packs 32 Blackhole application-specific integrated circuits, 1 terabyte of accelerator memory, and 23 petaFLOPS of Block FP8 performance. The list price starts at $110,000. (tenstorrent.com) Tenstorrent’s software pitch is as important as the hardware pitch. The company says TT-Forge, its machine-learning compiler, is open source, and its public inference-server repository shows active work on media and model-serving tools. (tenstorrent.com) (github.com) The speed claim is still a vendor claim. EE Times said Tenstorrent compared the 2.4-second result with “other leading hardware,” but the article did not publish a like-for-like benchmark table with identical model settings, power limits, and cost. (eetimes.com) That leaves the main takeaway narrower but clear: Tenstorrent now has a public demo showing Blackhole servers generating short 720p video at faster-than-real-time speed, and the next question is whether outside testers can reproduce it under the same conditions. (eetimes.com)