NVIDIA is making a substantial push into the AI market with its new Blackwell GPUs, targeting significant performance gains. At a closed press event attended by Reporter, the company highlighted that the new Blackwell chips offer 2.2x better performance for Llama 2 70B models and double the performance for GPT-3 175B models compared to the previous Hopper GPUs.
The Blackwell architecture introduces enhancements to NVIDIA’s Tensor Cores, first launched with the Volta architecture in 2017 and later integrated into the RTX 20 series for consumers. These improvements focus on increased efficiency, coupled with HBM3e memory, which delivers higher bandwidth and speed, significantly boosting overall GPU performance.
To illustrate this leap in capability, NVIDIA shared a comparison showing that just 64 Blackwell GPUs can match the performance of 256 Hopper GPUs. This reduction by a factor of four in the number of required GPUs highlights the significant architectural advances and efficiency improvements made in this new generation.
“Training results from Blackwell GPUs follow a previous submission for MLPerf Inference 4.1, where Blackwell delivered up to 4x LLM inference performance compared to the Hopper generation. Leveraging the FP4 accuracy of the Blackwell architecture, coupled with the NVIDIA QUASAR quantization, presentation revealed powerful performance while meeting benchmark accuracy requirements.”
Evolution of Hopper GPUs and NVIDIA’s future in the segment
In addition to introducing the new Blackwell GPUs, NVIDIA highlighted ongoing optimizations for its current Hopper GPUs. These enhancements have significantly boosted performance: the Hopper series now delivers over 3x the speed for GPT-3 175B models compared to last year. For Llama 2, NVIDIA achieved a 26% increase in efficiency, demonstrating its commitment to improving existing hardware.
Looking ahead, NVIDIA announced the upcoming Blackwell Ultra series, which is expected to launch in 2025. This advanced lineup will feature expanded memory and enhanced bandwidth capabilities, further pushing performance. Beyond that, NVIDIA is already planning the next-generation Rubin GPUs, which will integrate HBM4 memory technology and are anticipated to arrive post-2026.
NVIDIA also revealed a strong network of partners adopting the Blackwell GPUs, including prominent industry names like ASUS, Dell, Lenovo, Microsoft Azure, Cisco, and Fujitsu. Although a specific launch date for the Blackwell series has not been disclosed, it is expected to be available before the end of this year.
