NVIDIA Unveils Blackwell Ultra GPU Built on 3nm Process — 2x AI Performance Leap
NVIDIA announced the Blackwell Ultra architecture, its next-generation GPU built on a 3nm process delivering twice the AI inference performance of the previous generation.
In Brief
NVIDIA's Blackwell Ultra GPU, built on a 3nm TSMC process, delivers 2x AI inference performance with 288GB HBM4 memory. The 208-billion transistor chip targets enterprise AI workloads with 4x faster LLM training vs Hopper. Cloud providers have already committed to deployments.
NVIDIA has officially unveiled Blackwell Ultra, its next-generation GPU architecture built on a cutting-edge 3nm process node, during a special keynote at the Siggraph 2026 conference. The new architecture delivers a 2x improvement in AI inference performance over the existing Blackwell lineup.
The Blackwell Ultra architecture features 288GB of HBM4 memory with 8TB/s bandwidth, enabling training of models with over 2 trillion parameters on a single server node. NVIDIA claims the new architecture achieves 4x faster training times for large language models compared to the Hopper generation.
Information
Blackwell Ultra uses TSMC's N3E process with 208 billion transistors — the most complex chip ever built.
Enterprise customers including Microsoft Azure, Amazon AWS, and Google Cloud have already committed to deploying Blackwell Ultra-based instances in Q1 2027. The chip is expected to power the next wave of AI infrastructure expansion.
“Blackwell Ultra doesn't just push boundaries — it redefines what's possible in AI computing. The 3nm process gives us transistor density we've only dreamed of.”
— Jensen Huang, CEO of NVIDIA
The announcement sent NVIDIA shares up 5% in after-hours trading. Data center GPUs now account for over 78% of NVIDIA's revenue, and Blackwell Ultra is positioned to extend that dominance through 2028.
Key Takeaways
- 01Blackwell Ultra is built on TSMC's 3nm N3E process with 208 billion transistors
- 022x AI inference performance improvement over the previous Blackwell generation
- 03288GB HBM4 memory with 8TB/s bandwidth enables 2-trillion parameter model training
- 04Major cloud providers committed to Q1 2027 Blackwell Ultra deployments
- 05NVIDIA's data center GPU revenue now represents 78% of total company revenue
Frequently Asked Questions
When will Blackwell Ultra GPUs be available?
NVIDIA expects Blackwell Ultra-based products to ship to cloud providers in Q4 2026, with customer deployments beginning in Q1 2027.
How does Blackwell Ultra compare to Hopper?
Blackwell Ultra delivers 4x faster LLM training and 2x AI inference performance compared to the Hopper architecture, with 3x the memory bandwidth.
Sources
Topics
Stay Ahead with TECHRADAR360
Get the latest tech news delivered to your inbox every morning.
Related
AMD Ryzen 9000 Series Breaks Gaming Benchmarks with Zen 6 Architecture
Intel Arrow Lake-S Desktop CPUs Launch with Hybrid Core Design and New LGA1851 Socket
