Nvidia announces Blackwell Ultra and Vera Rubin AI chips

Nvidia announces Blackwell Ultra and Vera Rubin AI chips

  • 18.03.2025 18:47
  • nbcnewyork.com
  • Keywords: AI, GPU

Nvidia announced new AI chips, Blackwell Ultra and Vera Rubin, at its GTC conference. Blackwell Ultra offers enhanced performance for cloud providers, while Vera Rubin, a next-gen GPU, promises higher processing power and memory for AI tasks. These advancements highlight Nvidia's focus on meeting growing demand for efficient AI solutions.

Nvidia ProductsNvidia ReportsNvidia NewsNvidia ServicesNVDAsentiment_satisfied
open_in_newOriginal article

Estimated market influence

Nvidia

Nvidia

Positivesentiment_satisfied
Analyst rating: Strong buy

Announced new AI chips and GPU designs, leading to increased revenue and market dominance.

OpenAI

Positivesentiment_satisfied
Analyst rating: N/A

Their ChatGPT transformed Nvidia's business, increasing sales by over sixfold.

DeepSeek

Neutralsentiment_neutral
Analyst rating: N/A

Nvidia uses DeepSeek R1 model to benchmark new products; no direct financial impact mentioned.

Context

Nvidia Announces Blackwell Ultra and Vera Rubin AI Chips

Key Announcements

  • Blackwell Ultra: New chip family shipping in H2 2023, designed for high-performance AI model deployment.
  • Vera Rubin: Next-generation GPU system expected to ship in H2 2026, featuring a custom CPU (Vera) and advanced GPU design (Rubin).

Business Insights

  • Revenue Growth: Nvidia's sales have surged over sixfold since late 2022, driven by demand for its GPUs in AI training.
  • Market Dominance: Nvidia leads the market for AI model training with its "big GPUs," capturing significant cloud provider investments.

Vera Rubin GPU System

  • Performance:
    • Inference Speed: 50 petaflops (vs. 20 petaflops for current Blackwell chips).
    • Memory Capacity: 288GB of fast memory, a key metric for AI developers.
  • Architecture:
    • Vera is Nvidia's first custom CPU design, based on the Olympus core.
    • Rubin combines two GPUs into one system, with plans for a "Rubin Next" chip in H2 2027 that doubles speed (four GPUs combined).

Blackwell Ultra Chips

  • Performance:
    • Delivers more tokens per second, enabling faster content generation.
    • Cloud providers can achieve up to 50x revenue increase compared to the Hopper generation (2023).
  • Deployment Options:
    • Available in versions with two GPUs paired with an Arm CPU (GB300) and standalone GPU (B300).
    • Rack version features 72 Blackwell chips.

Competitive Dynamics

  • DeepSeek R1 Model: While initially seen as a threat, Nvidia is using it to benchmark its new products.
  • Reasoning Models: Vera Rubin and Blackwell Ultra are optimized for reasoning models, which require more computing power during deployment.

Strategic Considerations

  • Annual Release Cadence: Nvidia aims to release new chips annually, up from every other year pre-AI boom.
  • Custom CPU Development: Vera marks a shift toward custom CPU designs, potentially offering better performance than off-the-shelf options.

Long-Term Effects and Market Implications

  • AI Training Growth: The "scaling law of AI" suggests sustained demand for high-performance chips.
  • Cloud Provider Investments: Top cloud companies are deploying three times as many Blackwell chips as Hopper chips, indicating continued focus on Nvidia's technology.
  • Global Competition: While models like DeepSeek R1 pose challenges, Nvidia's focus on efficiency and advanced architectures positions it for long-term leadership.

Regulatory Considerations

  • No direct mention of regulatory impacts in the text.