NVIDIA's Blackwell AI Accelerators See Global Adoption
NVIDIA's Blackwell-generation AI accelerators are being integrated into new server platforms globally. Russian company YADRO is deploying the new GPUs in its servers for AI and machine learning workloads. In partnership with OpenAI, NVIDIA also demonstrated a nearly 2x acceleration for real-time inference on a 120B parameter model, showcasing the hardware's performance. Meanwhile, companies like Reinventy are joining the NVIDIA Inception Program to build out their edge AI roadmaps.
- The Blackwell B200 GPU is built on a custom TSMC 4NP process and features 208 billion transistors, a significant increase from the 80 billion in the previous Hopper generation. It achieves this by connecting two separate dies to create a single, unified GPU. - For AI workloads, the Blackwell architecture introduces a second-generation Transformer Engine and support for new, smaller data formats like FP4 and FP6, which can double the performance and the size of models that can be supported while maintaining accuracy. - A key configuration is the GB200 Grace Blackwell Superchip, which combines one Grace CPU with two Blackwell B200 GPUs via a 900GB/s NVLink-C2C interconnect, designed to eliminate data transfer bottlenecks. - Compared to the prior H100 GPU, the B200 offers up to 4 times the training throughput and is stated to be up to 2.5 times faster overall, with up to 25 times greater energy efficiency in some workloads. - For high-performance computing (HPC) tasks that rely on double-precision formats, Blackwell GPUs provide a 30% performance increase in FP64 and FP32 fused multiply-add (FMA) operations over the Hopper architecture. - A rack-scale system, the GB200 NVL72, connects 36 Grace CPUs and 72 Blackwell GPUs, functioning as a single massive GPU to deliver up to 30 times faster real-time inference for trillion-parameter large language models. - The NVIDIA Inception Program is a free, virtual accelerator that provides startups with technical training, access to SDKs, cloud computing credits, and networking opportunities with investors and partners. - New security features in the Blackwell architecture include NVIDIA Confidential Computing, which protects AI models and data with hardware-based security without compromising performance.
