Tenstorrent's Blackhole Servers Speed Video AI

Video AI works by predicting one frame after another, and that repeated cleanup step usually makes a short clip take longer than the clip itself to render. Tenstorrent said this week it generated a five-second video in 2.4 seconds on its new Blackhole-based servers. (eetimes.com) EE Times reported seeing the demo on April 21 and got a five-second, 720p clip with 81 frames and 40 steps in about three seconds from a text prompt. Tenstorrent said its best run was 2.4 seconds. (eetimes.com) The demo used an optimized Wan2.2-14B model from Prodia, a company that sells image and video generation application programming interfaces. Wan2.2 is an open video model family published on Hugging Face and GitHub, where its maintainers describe 720p and 24-frames-per-second generation modes. (eetimes.com) (prodia.com) (huggingface.co) (github.com) Tenstorrent told EE Times the run happened on four new Galaxy servers built around Blackhole chips. Tenstorrent’s own Galaxy product page lists 32 Blackhole application-specific integrated circuits, or custom AI chips, per server, which implies a 128-chip cluster for the demo system EE Times described. (eetimes.com) (tenstorrent.com) Those servers are new enough that Tenstorrent is still filling in the product stack around them. The company’s site lists Galaxy Blackhole servers starting at $110,000 and says the machines target both training and inference, the stage where a trained model actually produces an answer or a video. (tenstorrent.com) Blackhole is also the name of the chip family inside those systems. Tenstorrent’s hardware pages say each Blackhole card 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, a lower-precision format widely used for AI inference. (tenstorrent.com) (docs.tenstorrent.com) The pitch is speed without proprietary networking. EE Times said Tenstorrent’s architecture ties compute, memory, and networking together so large systems can run one program across many chips, while Tenstorrent says Galaxy scales from one chip to thousands over Ethernet under one programming model. (eetimes.com) (tenstorrent.com) Tenstorrent senior fellow Jasmina Vasiljevic told EE Times that video jobs had stalled partly because they needed too much compute, and CEO Jim Keller said the company is already working with video-focused customers, including some large frontier AI labs. Those claims have not yet been backed by public third-party benchmarks beyond the EE Times hands-on report. (eetimes.com) If more independent tests match the 2.4-second result, the practical change is simple: a model could make a clip faster than the clip would play back. That would lower wait times for video tools and give Tenstorrent a concrete benchmark as it tries to sell Blackhole systems into cloud and enterprise AI deployments. (eetimes.com) (tenstorrent.com)