This site is fictional demo content. It is not real news or affiliated with any real organization. Do not treat it as fact or professional advice.

NextPaper Logo
nextpaper.ccCYBER CHRONICLE
RSS
EN中文

Full article

FULL TEXT

View this issue
BriefAI

NVIDIA Releases Blackwell Ultra with HBM4 Memory

NVIDIA unveiled Blackwell Ultra at GTC 2028, featuring next-gen HBM4 memory with 8TB/s bandwidth. A single card can support 10-trillion-parameter model inference. Expected to ship in Q3.

Content

NVIDIA officially unveiled Blackwell Ultra at GTC 2028, a major upgrade following the Blackwell architecture.

Blackwell Ultra uses TSMC 3nm process and SK Hynix HBM4 memory, with memory bandwidth reaching 8TB/s—approximately 4x that of H100. The new architecture supports FP4 precision, with energy efficiency ratio improving over 3x in specific AI inference scenarios.

NVIDIA CEO Jensen Huang demonstrated: a single Blackwell Ultra can support real-time inference of 10-trillion-parameter LLMs, meaning future flagship model deployment thresholds will significantly decrease.

Blackwell Ultra's thermal design power is 1000W, requiring new liquid cooling solutions. NVIDIA simultaneously released the DGX Blackwell Ultra server, which with 8-GPU interconnect can support 100-trillion-parameter-level model training.

New products are expected to ship in Q3 2028, with pricing not yet announced. Analysts predict early Blackwell Ultra pricing may exceed $40,000, maintaining NVIDIA's dominance in the AI training chip market.

Boundary

This is fictional content for entertainment only.