摘要: On July 28, 2026, NVIDIA officially released the Blackwell Ultra GPU, built on TSMC's 3nm process, integrating over 200 billion transistors, with AI training performance 4x higher than the previous generation. This chip is optimized for large model training and inference, while promoting the deployment of edge AI and IoT applications, marking a new stage in smart chip technology.
On July 28, 2026, NVIDIA officially launched its next-generation flagship AI chip, the Blackwell Ultra GPU, at its annual technology conference in Silicon Valley. Built on TSMC's advanced 3nm (N3P) process, the chip integrates over 200 billion transistors, delivering 4x higher AI training performance and 3.5x higher inference performance than the previous Hopper architecture, making it one of the world's most powerful AI accelerators.
Architectural Breakthrough: From HBM4 to NVLink 6.0
Blackwell Ultra continues NVIDIA's dual-die packaging strategy but now features the new NVLink 6.0 interconnect technology, boosting communication bandwidth between the two dies to 1.8 TB/s, a 50% increase over the previous generation. For the first time, the chip is equipped with Samsung's HBM4 high-bandwidth memory, offering up to 384 GB capacity and bandwidth exceeding 10 TB/s, fully meeting the processing demands of large language models and massive datasets.
NVIDIA CEO Jensen Huang stated at the event: "Blackwell Ultra is not just about speed; it's a paradigm shift in computing. It reduces AI model training costs by 60% and energy consumption by 40%, laying the foundation for general artificial intelligence with trillion-parameter scales." He also revealed that the first DGX Ultra servers powered by this chip will be delivered in Q4 2026, with pre-orders already exceeding $5 billion.
Comparison with Previous Hopper: Leap in Performance
Compared to the H100 Hopper GPU released in 2024, Blackwell Ultra achieves significant leaps across key metrics:
- Transistor count: From 80 billion to 200 billion, a 150% density increase
- AI compute: FP8 Tensor Core performance jumps from 1979 TFLOPS to 8000 TFLOPS
- Memory bandwidth: From 3.35 TB/s (HBM3) to 10 TB/s (HBM4)
- Interconnect bandwidth: NVLink doubles from 900 GB/s to 1.8 TB/s
These improvements allow Blackwell Ultra to train GPT-4-level models in just 5 days, down from 3 weeks, with inference response times reaching millisecond levels, enabling real-time dialogue, video generation, and other complex applications.
Driving Smart Chip Applications in IoT and Edge Computing
Beyond the data center market, NVIDIA also unveiled an edge computing version based on the Blackwell Ultra architecture — Blackwell Ultra Edge — designed for industrial IoT, autonomous driving, and smart cities. This version uses a downclocked but more energy-efficient design, achieving 20 TFLOPS/W energy efficiency and supporting on-device large model inference.
"In the past, power consumption and form factor limited AI chip deployment at the edge. With advanced 3nm technology and sparse computing, Blackwell Ultra Edge keeps power under 150W while delivering compute performance equivalent to previous-generation data center chips," said Deepu Talla, NVIDIA's vice president of edge computing, in a follow-up interview. He cited an example in smart factories where the chip can simultaneously process defect detection from dozens of 8K video streams with decision latency under 1 millisecond.
Industry Impact: Intensifying Chip Manufacturing Competition
By adopting TSMC's 3nm process first, NVIDIA has widened its lead over competitors. AMD's Instinct MI400 series and Intel's Falcon Shores remain on 5nm/4nm nodes, with 3nm expected only by 2027. Meanwhile, Samsung and IBM's neuromorphic chip "NorthPole 2" boasts extremely low power but still cannot match Blackwell Ultra in absolute computing power.
TSMC revealed that 3nm process yield exceeded 95% in Q2 2026, with capacity prioritized for NVIDIA orders. The company also confirmed that 2nm (N2) mass production will begin in 2027, with NVIDIA already reserving initial capacity.
Market research firm IDC analysts commented: "The launch of Blackwell Ultra will accelerate AI adoption, especially in IoT. Demand for edge AI chips is set to explode. By 2028, the global AI chip market is projected to exceed $200 billion, with edge AI's share rising from 15% to 35%."
Future Outlook: From Cloud to Edge — A Smart Chip Ecosystem
NVIDIA simultaneously announced a new CUDA 13 software platform, optimized for Blackwell Ultra with enhanced AI framework support and a patented "auto-parallel compilation" feature that enables developers to fully leverage multi-die performance without manual tuning.
In the smart chip arena, Blackwell Ultra is not just a product but represents a technology roadmap — combining advanced manufacturing with system-level optimization. As AI extends from cloud to edge, NVIDIA's "super chip" builds a complete AI computing loop from training to inference, from data centers to IoT, further solidifying its leadership in the smart chip market.
Looking ahead, competition in smart chips will revolve around three dimensions: process technology, energy efficiency, and ecosystem. And Blackwell Ultra has undoubtedly set a new benchmark in this race.
