NVIDIA Blackwell Ultra Sets World Record in AI Model Training Performance
NVIDIA announced that its Blackwell Ultra GPU set a new world record by reaching 1,648 TFLOPs in pretraining the DeepSeek-V3 671B model, tripling the performance of its previous generation thanks to innovations in design and software optimization.
What Happened
NVIDIA posted on its official Twitter account that its new graphics processor, the Blackwell Ultra GPU, set a world record in performance for pretraining the artificial intelligence model DeepSeek-V3 with 671 billion parameters (671B). The GPU achieved 1,648 TFLOPs of performance per unit, approximately three times the performance delivered by the previous generation of its AI hardware. This breakthrough was achieved through an extreme collaborative design between hardware and software, accompanied by continuous optimizations.
In parallel, Intel highlighted its role in the semiconductor industry, mentioning that it designs and manufactures CPUs and advanced packaging for data centers specialized in AI. Additionally, the company operates its own factories through Intel Foundry, being one of the few U.S. manufacturers of cutting-edge chips, which implies significant support for the critical infrastructure for the development and adoption of AI technologies.
Why It Matters
NVIDIA's achievement in the performance of its Blackwell Ultra GPU is a significant milestone for AI infrastructure, especially in training large foundational models like DeepSeek-V3. The increased processing capacity allows accelerating development and innovation cycles, enabling more efficient and complex training that can produce more robust and accurate models.
Likewise, Intel's role as an advanced chip manufacturer and its manufacturing infrastructure in the United States strengthen the critical technology supply chain for AI evolution. Local availability of cutting-edge semiconductor production can help mitigate issues of international dependency and ensure continuity in developing high-performance hardware for AI.
Both advances highlight the importance of the co-evolution of hardware and software in the AI ecosystem, where chip performance and software optimization are indispensable elements to sustain the exponential growth of foundational model capabilities.
What Remains to Be Confirmed
Based on available publications, there are no specific technical details about the internal architecture of the Blackwell Ultra GPU nor information on when it will be commercially available to customers or its costs. There is also no information about explicit collaborations between NVIDIA and Intel or how their developments integrate into joint solutions for AI infrastructures.
## Sources - NVIDIAAI. https://x.com/NVIDIAAI/status/2079582939373863353 - Solana. https://x.com/Solana/status/2079561569932083390
*This note is based on public posts from official accounts of technology companies. The information requires further verification and does not imply financial recommendations or market projections.*