Google Unveils Trilium TPU & NVIDIA Blackwell for Cloud AI Dominance

According to Wccftech, Google recently held the I/O 2024 conference, where it unveiled its latest sixth-generation cloud TPU, “Trilium,” which has already been integrated into Google’s AI supercomputer. Additionally, Google announced the inclusion of NVIDIA’s Blackwell architecture GPUs into the company’s “AI arsenal,” set to support Google Cloud server solutions and expected to be available in early 2025.

Google claims that the sixth-generation cloud TPU “Trilium” boasts a doubling of HBM capacity and interconnect bandwidth compared to the previous generation TPU (TPU v5p), enabling more efficient AI model simulation and development. Trilium also offers a 67% improvement in energy efficiency, making it more suitable for large-scale data centers. Google has partnered with tech company Hugging Face to simplify model training and deployment, allowing developers to easily fine-tune and run open-source models on Google’s AI infrastructure, including TPUs.

Regarding the GPUs based on NVIDIA’s Blackwell architecture, Google asserts that this will elevate their computational performance to unprecedented levels. From the live demonstration, it appears that Google will utilize NVIDIA’s GB200 NVL72 computing platform, a full-rack solution comprising eighteen 1U servers. This platform delivers 720 petaflops of FP8 performance, 1440 petaflops of FP4 computational power, and can handle AI LLM models with 270 trillion parameters. Each server features two GB200 Grace Blackwell Superchips, totaling 36 Grace CPUs and 72 Blackwell GPUs. These compute nodes are equipped with 1.7TB of HBM3E memory, 32TB/s memory bandwidth, and are entirely liquid-cooled using MGX packaging.