HeyGen moved its talking-head AI from Nvidia GPUs to Google TPUs. The new Trillium port is 1.86× faster than its first, with no Nvidia comparison.

HeyGen ported Avatar IV, its flagship talking-head video model, from Nvidia GPUs to Google's Trillium TPUs and says the result is roughly 1.86× faster. Avatar IV is the engine behind HeyGen's product that turns a single photo and a voice clip into a moving, talking video, and it carries more than 18 billion parameters across two transformers plus a VAE decoder that turns the model's internal math into pixels.

The team kept the production model code in PyTorch and routed it through torchax, a PyTorch frontend that runs on Google's JAX compiler. Hardware-specific engineering concentrated where the two chips actually diverge: each attention shape was rewritten as a Pallas kernel, and XLA compiled the full pipeline end to end. The post identifies three walls they knocked down on an eight-chip Trillium host: exposed all-to-all collectives in the mesh, partial blocks in the sparse attention grid, and a serial dependency inside the softmax loop.

The 1.86× is a HeyGen and Google Cloud joint measurement against HeyGen's first working version on the same Trillium host, with identical output-quality gates. The post does not compare against a current Nvidia GPU baseline, and it does not show cost or time-to-first-frame. One shipped workload is portable; the AI video market is not declared portable.