Fujitsu's PHOTON architecture is up to 475x faster than Transformer for multi-query tasks

Fujitsu in Japan introduced a new AI architecture called PHOTON (Parallel Hierarchical Network for Top-down Processing). It aims to fix Transformer models' slowdowns during long text or multi-thread tasks caused by heavy memory operations. PHOTON processes at the semantic level instead of token-level, reducing computing complexity and improving parallelism. For multi-query tasks, it decides by majority or best pick in one go. Tests on smaller models (600M to 1.2B parameters) show higher throughput and lower memory use. The 1.2B model performs up to 475 times better than Transformer, with slightly lower quality. It also cuts KV Cache needs, boosting max iterations and potentially lowering GPU costs.