-
Google’s Ironwood TPU: A Leap in AI Technology: Google launched its 7th generation Tensor Processing Unit (TPU), Ironwood, at the Next ’25 conference.
-
Inference-Focused Design: Ironwood is designed for AI inference, enabling AI models to make predictions based on learned data.
-
Proactive AI Systems: Ironwood signals a shift to proactive AI systems that independently generate insights, marking the “age of inference.”
-
Performance Boost: Ironwood features a peak compute of 4,614 teraflops (TFLOP) per chip.
-
Scalability: Ironwood can scale to clusters of 9,216 liquid-cooled chips, delivering up to 42.5 exaflops of computing power.
-
Superior Computing Power: Ironwood’s throughput is claimed to be over 24 times that of the world’s largest supercomputer, El Capitan.
-
Expanded Memory: Each Ironwood chipset offers 192GB of memory, six times more than its predecessor, Trillium.
-
Next-Gen AI Support: Engineered for Large Language Models (LLMs) and Mixture of Experts (MoEs), essential for advanced reasoning.
-
Advanced Interconnect: Features a cutting-edge Inter-Chip Interconnect (ICI) network for rapid data transfer and synchronization.
-
SparseCore: Includes an advanced SparseCore component for applications like recommendation systems and financial modeling.
-
Pathways Software: Integrates Google’s Pathways software for distributing AI workloads across thousands of TPUs.
-
Improved Efficiency: Delivers double the performance per watt compared to Trillium, due to optimized design and liquid cooling.
-
High Bandwidth Memory (HBM): Features 192 GB of HBM per chip with 7.2 terabits per second bandwidth, reducing latency.
-
ICI Network Bandwidth: Enhanced ICI network with 1.2 terabits per second bidirectional bandwidth for efficient communication.
-
Broadened Applications: Supports scientific and financial domains in addition to AI-specific applications.
-
Availability: Set to be available later in 2025, setting a new benchmark for AI performance and efficiency.
