Google has released its latest artificial intelligence (AI) chip, Trillium, designed to power advanced AI workloads. The sixth-generation model in Google Cloud’s custom AI chip lineup, Trillium offers substantial advancements over its predecessor, the TPU v5e.
Key features of Trillium include:
– 4x faster training and 3x faster inference performance compared to TPU v5e
– Enhanced HBM and ICI bandwidth for large language models
– Capacity to scale up to 256 chips per pod, ideal for extensive AI tasks
Trillium boasts improved energy efficiency with a claimed 67% increase over previous generations. Benchmark testing has demonstrated substantially improved performance, with 4x faster training speeds and 3x greater throughput for inference tasks.
Google’s new TPU also features scalable design, allowing up to thousands of chips to be linked together in a single pod. The Multislice software ensures consistent performance across hundreds of pods.
This development is tied to the upcoming launch of A3 Ultra VMs featuring Nvidia H200 Tensor Core GPUs, offering Google Cloud customers a high-performance GPU option within the tech giant’s AI infrastructure.
Source: https://www.techradar.com/pro/google-puts-nvidia-on-high-alert-as-it-showcases-trillium-its-rival-ai-chip-while-promising-to-bring-h200-tensor-core-gpus-in-november-2024