NVIDIA B200 Blackwell GPUs Become Accessible for Research
NVIDIA's B200 Blackwell GPU, featuring 192GB of HBM3e memory and up to 8 TB/s of bandwidth, is now available. In a move to accelerate AI development, Israel has opened access to its national supercomputer, offering discounted B200 GPU accelerators to high-tech companies and researchers. The B200 is said to deliver up to 15 times faster inference performance than the H100.
- The B200 GPU is built on the Blackwell architecture, which features a dual-die chiplet design packing 208 billion transistors, a significant increase from the 80 billion in the previous generation H100. This design uses a custom TSMC 4NP process and connects the two dies with a 10 TB/s link, allowing them to function as a single, unified GPU. - A key architectural improvement in the Blackwell series is the second-generation Transformer Engine, which adds support for new floating-point formats like FP4 and FP6. This allows for more efficient processing of AI models by reducing their size while maintaining high accuracy, contributing to a significant reduction in energy consumption per inference compared to the H100. - The Israeli national supercomputer initiative is part of the "Telem Program" for AI R&D Infrastructure and is operated by the cloud provider Nebius. The program will distribute 1,000 B200 accelerators, with 70% allocated to high-tech companies and 30% reserved for academic research, aiming to reduce dependence on foreign cloud infrastructure. - To gain access, industrial applicants must request a minimum of 16 B200 accelerators, while academic groups need at least eight, for periods ranging from one to six months. This initiative is a strategic move by the Israel Innovation Authority to ensure domestic R&D and intellectual property remain within the country. - The fifth-generation NVLink interconnect on the B200 provides 1.8 TB/s of bidirectional bandwidth per GPU. This is double the bandwidth of the NVLink used in the H100, enabling faster communication between GPUs for training increasingly complex and large-scale AI models. - Phala Network provides access to B200 GPUs combined with Trusted Execution Environment (TEE) technology. This creates a secure enclave for AI workloads by pairing Intel's TDX for CPU and memory protection with NVIDIA's Confidential Computing to encrypt data while it is being processed on the GPU.
