Viral Wire

HIVE Digital's A40 GPUs match H100 performance in Columbia study

A40 GPUs in Paraguay outperform expectations, rivaling Nvidia's H100s...

Deep Dive

HIVE Digital Technologies Ltd. (NASDAQ: HIVE) has completed a groundbreaking AI research project with Columbia University's Department of Industrial Engineering and Operations Research. The study, submitted to the prestigious NeurIPS conference, demonstrated that HIVE's A40 GPUs located in Asunción, Paraguay could match the performance of Nvidia's newer H100 GPUs after code optimizations. Researchers in New York remotely conducted iterative training runs on the Paraguayan GPU cluster, proving the viability of intercontinental AI training. The optimizations focused on pretraining algorithms for large language models up to 1.4 billion parameters, achieving throughput and latency comparable to H100s when normalized for raw hardware performance.

This proof-of-concept validates HIVE's strategy to build an HPC/AI Gigafactory in Yguazú, Paraguay, where a 100MW substation is under construction. Civil works are complete, with commissioning expected in September 2026 and a Tier-III data center ready for service by H2 2027. Executive Chairman Frank Holmes emphasized that this demonstrates high-performance computing isn't geographically limited, and Paraguay can participate in the global AI economy. CEO Aydin Kilic noted the research focuses on making AI training smarter and more efficient by improving the mathematical foundations of neural network learning.

Key Points
  • Optimized HIVE A40 GPUs matched H100 performance for LLM pretraining up to 1.4B parameters after code optimization by Columbia University researchers.
  • Research submitted to NeurIPS proves intercontinental AI training feasibility, with New York researchers remotely using GPUs in Asunción, Paraguay.
  • HIVE is building a 100MW substation and Tier-III data center in Yguazú, Paraguay, with substation commissioning in September 2026.

Why It Matters

HIVE's validation of older A40 GPUs against H100s could lower AI infrastructure costs and prove distributed, non-geographic HPC is viable.

📬 Get the top 10 AI stories daily