NVIDIA Unveils H100 GPU and Grace CPU Superchip for AI Acceleration
NVIDIA introduces the H100 GPU and Grace CPU Superchip, delivering massive speedups for training large AI models like GPT-3. The new hardware features advanced architecture designs specifically tailored for modern transformer-based deep learning.
NVIDIA unveils its next-generation AI processors, the H100 GPU and Grace CPU Superchip, at the recent GTC conference. The H100 is built on the new Hopper architecture and stands out as the first GPU to support PCIe 5 and HBM3 memory. Meanwhile, the Grace CPU Superchip packs 144 Arm cores into a single-socket package connected by NVIDIA's high-speed NVLink-C2C technology.
The H100 GPU introduces a dedicated Transformer Engine designed to accelerate the training of large language models like GPT-3. This engine dynamically mixes 8-bit and 16-bit floating-point arithmetic to maximize performance, achieving up to a sixfold speedup for a 175B parameter GPT-3 model. Additionally, new dynamic programming instructions provide up to a sevenfold performance boost for complex algorithms like protein folding.
NVIDIA positions these new chips as the foundational engine for the world's AI infrastructure. Alongside raw performance gains, the Hopper architecture incorporates Confidential Computing technology to enhance security and privacy during data processing. Together, the H100 GPU and Grace CPU Superchip represent a significant leap forward for enterprises looking to accelerate their AI-driven businesses.