NVIDIA CEO Jensen Huang gave an hour and 45-minute keynote talk just before the start of the 2025 CES. He gave many announcements, including announcing a desktop AI machine called Project DIGITS, shown above, that should be available in May, in partnership with MediaTek.
He also talked about AI computer graphics with the Blackwell GeForce RTX 50 and AI foundation models for RTX AI PCs. He also talked about their Cosmos Foundation Model Platform to accelerate physical AI development, generative physical AI with NVIDIA’s Omniverse, vehicle company partners for next-generation highly automated and autonomous vehicle fleets and NVIDIA automotive safety and cybersecurity milestones for AV development. In this article we will talk about storage and memory and their use in some of the products discussed in Jensen’s keynote talk.
The RTX Blackwell device that powers the Blackwell GeForce RTX 50 desktop and laptop GPUs is shown below, with Jensen holding the desktop version. It has 1.8TB/s memory bandwidth with up to 32GB of VRAM (video random access memory for storing graphics data). Jensen pointed out that G7 (GDDR7) memory was sourced from Micron, one of two call-outs to Micron memory products in his presentation (the other was to HBM memory used in the Grace Blackwell NVLink72 shown below).
The GeForce RTX 50 Series GPUs run creative generative AI models up to 2X faster in a smaller memory footprint compared than the previous generation product. The GPUs support FP4, a quantization method to decrease image model sizes, somewhat similar to file compression. As a result, FP4 uses less than half of the memory and twice the performance and this is done with virtually no loss in the quality of the images.
NVIDIA showed real time image generation of very high-resolution streaming video using this GPU during the presentation. This GPU also enables 3D and virtual reality (VR) video and AV1 ultra high-quality mode and greater color depth than is available on most consumer cameras.
Jensen showed an image of the Grace Blackwell NVLink72. This was introduced by NVIDIA in March of 2024. As shown below, this large chiplet includes 72 Blackwell GPUs and 36 Grace CPUs. It supports 1.4 exaflops per second, EFLOPS of Tensorflow 4-bit Floating Point processing.
The 576 high bandwidth memory chips connected to the GPUs provide about 14TB of memory with 1.2PB/s aggregate bandwidth. The CPUs have up to 17TB of LPDDR5X memory with up to 18.4TB/s performance. These chiplet devices are mounted in racks that provide power and connectivity to support massive AI training.
At the very end of his talk Jensen introduced NVIDIA Project DIGITS with a new Grace Blackwell Superchip in a package that provides a desktop AI engine as shown below and at the beginning of this article. The GB10 Grace Blackwell Superchip is a system-on-a-chip (SoC) and provides a petaflop of FP4 AI computing performance that can be used for creating and running large AI models. The box shown also includes a 4TB SSD and 128GB of LPDDR5X memory.
The press release says that the GB10 Superchip features an NVIDIA Blackwell GPU with the latest-generation CUDA cores and with fifth-generation Tensor Cores, connected via NVLink-C2C chip-to-chip interconnect to a high-performance NVIDIA Grace CPU, which includes 20 power-efficient cores built with the Arm architecture. MediaTek, a market leader in Arm-based SoC designs, collaborated on the design of the GB10, contributing to its best-in-class power efficiency, performance and connectivity.
Jensen Huang from NVIDIA gave the first CES 2025 keynote talk and announced a slew of enterprise as well as consumer GPU products. Memory and storage play an important role in AI training and inference and his presentation showed how they enable modern AI solutions.
Read the full article here
