Google announces Trillium, its sixth-generation TPU, with 4.7x the peak compute per chip of TPU v5e

Google DeepMind/Tensor Processing Unit (TPU)

Google announced Trillium, its sixth-generation TPU, delivering 4.7x peak compute, 2x memory capacity and bandwidth, and 2x interchip interconnect bandwidth versus TPU v5e. The chips are 67% more energy-efficient; models including Gemini 1.5 Flash announced at I/O were trained and served on TPUs.

• Supports up to 256 TPUs in a single pod; hundreds via multislice; tens of thousands via Google Jupiter network.
• Third-generation SparseCore accelerator handles ultra-large embeddings for ranking and recommendation workloads.
• 2x HBM capacity and bandwidth enable larger models with more weights and faster memory access.
• Available later in 2024; represents the most sustainable TPU generation to date.