NVIDIA’s Blackwell platform, launched at GTC 2024, has doubled the performance of large language model (LLM) training benchmarks in MLPerf Training v4.1, showcasing significant gains over previous architectures. The platform, now in full production, features seven types of chips, including GPU, CPU, and DPU, and is designed to support the development of next-generation LLMs.
In MLPerf Training benchmarks, the Blackwell platform outperformed its predecessor, Hopper, across all tests, with a 2x increase in performance for GPT-3 pre-training and a 2.2x boost for Llama 2 70B low-rank adaptation (LoRA) fine-tuning. The company plans to leverage the GB200 NVL72 system for further performance improvements.
The Blackwell platform’s advancements represent a significant leap in AI capabilities, paving the way for more complex and capable AI systems.
Source
<p>The post NVIDIA’s Blackwell Platform Doubles Large Language Model Training Performance first appeared on CoinBuzzFeed.</p>