Luma moves production inference to AMD GPUs
Video AI lab Luma says thousands of AMD Instinct MI325X GPUs now handle inference for all its major models, a sign that AMD hardware is being used for frontier production workloads.
Luma says that AMD hardware, using compute from partners including Tensorwave, now handles inference for every major model it runs. Thousands of AMD Instinct MI325X GPUs serve text, image, video and multimodal requests across its platform. Those requests cover its flagship video model Ray3.2, its new multimodal reasoning model Uni-1, and Luma Agents.
The company explained the move through one pipeline, Ray3.2's video-to-video feature. Luma said that pipeline is hard to serve because it is built from several submodels, including VAEs, ControlNets, LoRAs and a denoiser backbone. According to Luma, the MI325X has enough memory to keep every submodel loaded at once. That avoids slow CPU offloading and the extra communication that comes with splitting models across more GPUs.
Luma trains on a separate cluster, and it said the inference code needed little porting. AMD's PyTorch backend ran about 90% of the code without changes, including FP8 work and custom Triton kernels. The one manual change was attention: Luma sent those calls to AMD's AITER library. For its multimodal model, which uses KV caching and a disaggregated setup, Luma relied on AMD's FlashInfer support. It said it needed help from AMD to reach its performance targets.
The video-to-video pipeline took two to three engineers about two weeks to reach production. The multimodal model took three engineers about four weeks. Luma said much of that time had nothing to do with AMD, because it was the first multimodal model the company had served to customers. Human evaluators found no drop in output quality.
This is a case study Luma published itself. It does not give cost, throughput or latency comparisons with its previous hardware, and it does not say what share of its total compute AMD now provides.