Nvidia details its GPU, CPU, and system roadmap for the next three years

Nvidia details its GPU, CPU, and system roadmap for the next three years

  • 18.03.2025 00:00
  • networkworld.com
  • Keywords: AI, GPU

Nvidia revealed its GPU, CPU, and system roadmap for the next three years, including the Rubin GPU in 2026, Rubin Ultra in 2027, and Feynman in 2028. The systems feature enhanced performance, new interconnects, and a focus on AI model efficiency.

Nvidia ProductsNVDAsentiment_satisfiedCSCOsentiment_satisfiedDELLsentiment_satisfied

Estimated market influence

Nvidia

Nvidia

Positivesentiment_satisfied
Analyst rating: Strong buy

Announced new GPU and system roadmaps, leading to increased revenue potential.

Cisco

Cisco

Positivesentiment_satisfied
Analyst rating: Buy

Collaboration with Nvidia on AI security solutions.

Dell

Dell

Positivesentiment_satisfied
Analyst rating: Strong buy

Partnership with Nvidia for developer tools.

Context

Analysis of Nvidia's GPU, CPU, and System Roadmap (2026-2028)

Key Product Launches

  • Vera Rubin NVL144

    • Release: Second half of 2026
    • FP4 performance: 3.6 exaflops
    • FP8 performance: 1.2 exaflops
    • Includes Vera CPU with 88 custom cores and 176 threads
    • GPU: Rubin (successor to Blackwell) with 50 petaflops of FP4 performance
    • Memory: HBM4e at 4.6PBps
    • Interconnect: NVLink 6 (260TB/second)
  • Rubin Ultra NVL576

    • Release: Second half of 2027
    • FP4 performance: 15 exaflops
    • FP8 performance: 5 exaflops
    • Includes 576 Rubin GPUs, 12,672 Vera CPU cores, and 72 Bluefield DPUs
  • Feynman GPU

    • Release: 2028
    • Details limited; next-generation HBM memory paired with Vera CPUs

Performance Metrics

  • Vera Rubin NVL144:

    • 3.6 exaflops FP4 (3x faster than predecessor)
    • 1.2 exaflops FP8
    • System power: 600 kilowatts, 2.5 million parts
  • Rubin Ultra NVL576:

    • 15 exaflops FP4 (4x faster than Vera Rubin)
    • 5 exaflops FP8

Power Consumption and Efficiency

  • Current systems:

    • Blackwell NVL72: 120 kilowatts (FP4)
    • Blackwell Ultra NVL72: Up to 135-140 kilowatts (estimates)
  • Efficiency Focus:

    • FP4 reduces power consumption during inferencing.
    • Huang: "Revenues are power limited."

Competitive Dynamics

  • Market Leadership:

    • Nvidia's roadmap positions it as a leader in AI and HPC systems.
    • Emphasis on extreme scale-up systems (e.g., 576 GPUs per rack).
  • Partnerships:

    • Collaborations with Cisco (Secure AI Factory) and Dell (AI model prototyping).

Strategic Considerations

  • Focus on FP4:

    • Reduces power consumption while maintaining performance.
    • Enables scalability for AI workloads.
  • Long-Term Effects:

    • Rubin systems aim to significantly reduce costs compared to Hopper generation.
    • Sustainability and efficiency will likely drive future innovations.

Market Implications

  • AI Growth Driver:

    • New GPUs enable faster reasoning models (100x more tokens generated).
    • Agentic AI adoption will be accelerated by these advancements.
  • Data Center Economics:

    • Power consumption directly correlates with revenue potential.
    • Nvidia's focus on performance per watt positions it for sustained growth in the data center market.

Regulatory and Industry Impact

  • No direct mention of regulatory impacts, but:
    • Sustainability concerns may influence future product development.
    • Higher power consumption could prompt stricter energy efficiency regulations.

Conclusion

Nvidia's roadmap underscores its commitment to advancing AI and HPC capabilities. The introduction of high-performance systems like Vera Rubin NVL144 and Rubin Ultra NVL576, coupled with a focus on efficiency (FP4), positions the company for sustained leadership in the AI market. Partnerships and ecosystem development will further reinforce its competitive edge.