A analysis challenge run by Columbia College professors on GPU servers positioned greater than 5,000 miles away in Paraguay has produced a outcome that HIVE Digital Applied sciences (Nasdaq: HIVE) is asking a proof of idea for the way forward for distributed AI infrastructure.
HIVE Digital, a publicly listed information heart and digital asset mining firm based in 2017 and traded on each the TSX and Nasdaq, introduced on June 22 that it has accomplished its first AI analysis collaboration with Columbia College’s Division of Industrial Engineering and Operations Analysis.
The analysis has been submitted to the Convention on Neural Info Processing Programs (NeurIPS), one of many three most influential machine studying conferences on the earth, alongside ICLR and ICML.
What the analysis really discovered
The Columbia group ran iterative AI coaching experiments remotely on HIVE’s A40 GPU cluster positioned in Asunción, Paraguay, whereas working from New York Metropolis.
The examine targeted on neural community pretraining, the foundational course of by which massive AI fashions study from information earlier than being fine-tuned for particular duties.
The headline discovering is critical. After two months of code optimisation particularly tailor-made to HIVE’s A40 {hardware}, the Columbia researchers discovered that efficiency matched that of Nvidia’s H100 GPUs, a more moderen and significantly dearer technology of chip extensively thought to be the gold normal for AI coaching workloads.
The comparability was made after normalising for every {hardware} platform’s uncooked efficiency.
The group additionally examined serving throughput and latency on a 1.4 billion parameter language mannequin, working normal benchmarks on LLaMA fashions as a part of the examine.
“In our use case of pretraining LLMs of as much as 1.4B parameters, our outcomes match these of H100s after normalizing for every {hardware}’s uncooked efficiency,” a Columbia researcher from the IEOR division mentioned.
“This work advances our understanding of recent neural community optimization, together with matrix-aware optimizers corresponding to Muon and associated scale-invariant strategies.”
Trending on TheStreet Roundtable:
- Wall Road analyst predicts 400% upside for MicroStrategy inventory
- Billionaire reveals jaw-dropping proposal for Satoshi’s Bitcoin
- Mark Zuckerberg eyes one other billion-dollar market
Why Paraguay and what comes subsequent
The Asunción outcomes now function a efficiency baseline for a bigger infrastructure challenge HIVE is constructing in Yguazú, Paraguay. The corporate has a 100 megawatt substation below building there, with civil works already full.
The substation is anticipated to be energised in September 2026, with building on a brand new Tier-III information centre starting within the autumn and a ready-for-service date focused for the second half of 2027.
HIVE’s wager on Paraguay is rooted within the nation’s power profile — low cost, considerable hydroelectric energy that makes it a horny location for power-intensive computing workloads.
HIVE government ehairman Frank Holmes mentioned, “Paraguay has the ability, the strategic location, and now the proof level.”
“Excessive-performance computing doesn’t have to be restricted by geography. Our imaginative and prescient is to convey the appropriate energy, information heart design, software program stack, and execution, so by means of HIVE, Paraguay can take part immediately within the world AI economic system.”
HIVE’s broader story
HIVE was among the many first publicly listed corporations to mine Bitcoin utilizing inexperienced power when it was based in 2017. Immediately it operates Tier-I and Tier-III information centres throughout Canada, Sweden, and Paraguay, serving each Bitcoin mining purchasers and high-performance computing clients.
The corporate has beforehand collaborated with Intel Company on customized mining {hardware} and have become certainly one of Sweden’s largest demand-response contributors, serving to stability the nationwide electrical grid.
The Columbia collaboration represents HIVE’s clearest step but into AI infrastructure as a industrial enterprise line, transferring past crypto mining into the broader marketplace for GPU compute.
“Innovation is in our DNA,” HIVE President and CEO Aydin Kilic mentioned. “We see a vivid future for AI in Paraguay.”
Kilic added that nice engineering can unlock important worth. HIVE will proceed to put money into communities whereas bringing developments in our tech stack to our information facilities globally, he added.
