NVIDIA Blackwell Accelerates AI Workloads
NVIDIA's Blackwell family of AI accelerators is seeing rapid adoption, with Russian firm YADRO integrating the latest GPUs into its servers for AI and machine learning tasks. NVIDIA itself claims its hardware has nearly doubled the real-time inference speed for OpenAI's GPT-OSS 120B model. In automotive, the 2000-TOPS DRIVE Thor SoC is setting a new standard for autonomous vehicle compute power.
- The Blackwell architecture, officially unveiled at NVIDIA's GTC on March 18, 2024, is named after David Blackwell, the first African American scholar inducted into the National Academy of Sciences. - A key architectural innovation in Blackwell is its dual-die design, where two large dies are connected by a 10 TB/s interconnect, allowing them to function as a single, unified GPU. This design packs 208 billion transistors and is manufactured using a custom TSMC 4NP process. - Compared to the previous Hopper architecture, the Blackwell B200 GPU offers more than double the memory capacity (192GB vs. 80GB HBM3) and 2.4 times the memory bandwidth (8 TB/s vs. 3.35 TB/s). - The DRIVE Thor SoC leverages the Blackwell architecture and is the first to integrate an Arm Neoverse V3AE CPU, a combination designed to consolidate functions like automated driving, infotainment, and driver monitoring into a single system. - Automakers including BYD, Hyper, XPENG, and Li Auto have announced plans to build their future vehicles on the DRIVE Thor platform. - Intel's Gaudi 3 accelerator is positioned as a competitor, with Intel claiming better price/performance in certain precisions, though it lacks support for the FP4 format found in Blackwell. - AMD's upcoming Instinct MI400 series is expected to compete with NVIDIA's next generation of chips, with AMD claiming the flagship MI400X will be significantly more powerful than their current MI300X. - YADRO's integration involves the NVIDIA RTX PRO 6000 Blackwell Server Edition in their G4208P G3 server, targeting a mix of AI computing and visualization workloads.
