News Overview
-
GB300 AI GPU Introduction: NVIDIA is set to unveil the GB300 “Blackwell Ultra” AI GPU, featuring 288GB of HBM3E memory and a 1.4kW thermal design power (TDP), promising a 50% performance increase over its predecessor.
-
Rubin Architecture Preview: The upcoming Rubin R100 AI GPU will incorporate dual logic chips, 384GB of HBM4 memory, and a 1.8kW TDP, with production anticipated to commence in late 2025.
Original article link: TweakTown
In-Depth Analysis
GB300 “Blackwell Ultra” AI GPU
-
Memory Capacity: Equipped with 288GB of HBM3E memory, utilizing 12-high stack technology, significantly enhancing data throughput.
-
Power Consumption: The GPU’s TDP is rated at 1.4kW, indicating substantial power requirements for its enhanced performance capabilities.
-
Performance Enhancement: Anticipated to deliver a 50% increase in FP4 computing performance compared to the previous B200 model.
-
Availability: Shipments are expected to commence in the third quarter of 2025.
Rubin R100 AI GPU
-
Dual-Chip Design: Features a dual logic chip structure, with each chip fabricated on TSMC’s N3 process node, aiming to enhance computational efficiency.
-
Advanced Memory: Incorporates eight stacks of HBM4 memory, totaling 384GB, a 33% increase over the GB300’s HBM3E memory capacity.
-
Increased Power Demand: The TDP is projected to be around 1.8kW, reflecting the GPU’s advanced capabilities.
-
Networking Enhancements: Expected to adopt a 1.6T network configuration, utilizing two ConnectX-9 Network Interface Cards (NICs) to improve data transfer rates.
Commentary
NVIDIA’s forthcoming GB300 and Rubin R100 AI GPUs signify substantial advancements in AI computing, focusing on increased memory capacities and enhanced performance metrics. The GB300’s 50% performance boost and the Rubin R100’s integration of cutting-edge technologies like HBM4 memory and dual-chip architecture position NVIDIA to address the escalating demands of AI workloads. However, the elevated power requirements, with TDPs reaching 1.4kW and 1.8kW respectively, necessitate robust cooling solutions and may influence data center infrastructure considerations. These developments underscore NVIDIA’s commitment to pushing the boundaries of AI hardware capabilities.