The Embodied AI

NVIDIA’s upcoming Blackwell GPU architecture is highly anticipated, with rumors circulating about its potential to revolutionize AI, machine learning, and gaming. Expected to succeed the current Hopper architecture, Blackwell is set to introduce cutting-edge technology in both data centers and consumer GPUs.

What Are GPUs and Why Are They Important?

A Graphics Processing Unit (GPU) is a specialized electronic circuit designed to accelerate the processing of visual data. Originally developed to render complex graphics for video games, GPUs have evolved far beyond gaming. Today, they are integral to high-performance computing (HPC), artificial intelligence (AI), machine learning (ML), and deep learning (DL) applications. Their ability to handle parallel tasks makes them much faster than traditional CPUs for certain types of computations, such as matrix operations that are common in AI workloads.

Modern GPUs, like those from NVIDIA, also feature Tensor Cores and Ray Tracing Cores, which make them exceptionally powerful for tasks such as deep learning, real-time ray tracing, and generative AI models. This rapid advancement in GPU technology has driven innovations across industries, from medical research to autonomous vehicles, all of which rely on massive data processing and real-time computations.

NVIDIA Blackwell: What’s New?

Blackwell is expected to bring several key improvements over its predecessor, Hopper. These include:

1. Increased AI Capabilities: Blackwell is anticipated to build on Hopper’s advancements in Tensor Cores, which handle the matrix operations crucial for deep learning and AI workloads. While Hopper introduced the FP8 (8-bit floating point) format, Blackwell could go further by improving mixed precision capabilities, enabling faster training and inference for large language models (LLMs) and generative AI. These enhancements will likely focus on improving tensor processing speed and accuracy, critical for the demands of modern AI applications, such as natural language processing (NLP) and large-scale recommendation systems​.
2. Energy Efficiency: Power consumption is a growing concern as AI workloads scale, particularly in data centers. Blackwell is expected to improve performance-per-watt over Hopper, which already made strides in energy efficiency. One likely area of enhancement will be in the process node; Blackwell could adopt a 3nm fabrication process, improving transistor density, reducing leakage, and optimizing power consumption. This would lead to more efficient computation for AI and high-performance computing (HPC) tasks, reducing both the operational costs and environmental footprint of running large GPU clusters.
3. Advanced Memory and Multi-Chip Design: NVIDIA is expected to introduce GDDR7 memory in Blackwell, delivering up to 36 Gbps of memory bandwidth, a significant improvement over Hopper’s GDDR6X, which maxes out at 23 Gbps. This increase will dramatically enhance data transfer speeds, crucial for memory-bound applications like AI training and real-time data analytics. Additionally, Blackwell may feature multi-chip module (MCM) architecture, which allows multiple GPU dies to be interconnected on a single package. This design would significantly improve scalability and throughput, enabling Blackwell to handle much larger workloads with increased efficiency, similar to AMD’s use of MCMs in their CPUs and GPUs.

Conclusions

The first Blackwell GPUs are expected to launch mid-to-late 2024 for data centers, with consumer GPUs like the RTX 5090 anticipated in early 2025. There’s also speculation that the highest-end consumer cards could drop by the end of 2024.

While Blackwell is likely to be more energy-efficient than previous architectures, the growing demand for AI and HPC workloads could lead to higher overall energy consumption. NVIDIA’s efforts to improve efficiency are critical, but the tech industry must balance innovation with sustainability to reduce its environmental footprint​.