Using GIGABYTE's AI TOP ATOM, Australia's Reasoning AI Trailblazer Cognitivo Tackles Artificial Intelligence's "Last Mile"

The spark of inspiration that changes the world can come from anywhere, especially in the age of artificial intelligence. Case in point: the Australian AI software company Cognitivo, established in 2017. Founder and CEO Alan Hsiao, a 20-year veteran of the fintech sector, positions Cognitivo as a pioneer in the field of reasoning AI for commercial use. The billion-dollar question on his mind is what's known as the "last mile" conundrum in AI implementation. His solution: "Neuro-Symbolic AI", capable of following chains of logic, resolving ambiguity, and explaining its outputs. It was created through the coupling of large language models (LLMs) and knowledge graphs (KGs). To incubate Neuro-Symbolic AI's next stage of development, especially as it pertains to scaling, Cognitivo turned to GIGABYTE's AI TOP ATOM personal AI supercomputers.

To date, Cognitivo has helped clients innovate in banking, financial advice, and healthcare, so that they might overcome the last mile of AI adoption and arrive at meaningful AI integration. Cognitivo has worked in partnership with UNSW and other Australian research organizations to develop many state-of-the art techniques in knowledge graph integration in retrieval-augmented generation (RAG) pipelines. This research has been commercialized through working with innovation teams within major Australian financial institutions on multiple applications, such as product recommenders, financial advice assistants, and coding agents for legacy software migrations.

The secret ingredient that Cognitivo injects into LLMs is the knowledge graph—a graph-structured data model which allows for multi-hop reasoning and contextual intelligence. Whereas conventional RAG responds to queries by combing its vector database for entries with the closest vector similarity, a knowledge graph engages in hybrid retrieval combining semantic vector search and graph context to take attributes like relationships and causality into account. The knowledge graph has another advantage over RAG: it retains long-term memory of user context due to its dynamic read-write architecture, as opposed to RAG's read-only limitation. The resulting KG-augmented LLM is less prone to hallucinations, and it is more likely to provide accurate and actionable answers. Additionally, the LLM is able to demonstrate its line of reasoning, as opposed to traditional AI's "black box" thought process. Employees now have the opportunity to discuss with the AI about its outputs, rather than blindly following its recommendations or dismissing the AI as "unhelpful". In turn, managers can review every step of the workflow and foster better workplace accountability.

Hsiao offers another example of how Cognitivo's custom-built AI outshines publicly available LLMs. One of its customers is an Australian healthcare provider specializing in mental health. The client uses AI for speech-to-text transcription, notetaking, risk analysis, and medical data compilation. None of the sensitive patient information can be uploaded to the cloud, of course; furthermore, topics such as suicide and self-harm are serious red flags for general-purpose LLMs. Cognitivo engineered AI according to industry-specific demands and helped advance smart medicine.

Building custom AI for diverse applications front-loads a lot of the heavy lifting on Cognitivo, however, and so it went shopping for a developer kit that not only has the juice to deliver competitive performance for AI model fine-tuning, but also the intelligence-per-dollar suitable for continuous feedback loops and data at scale. Hsiao and his team ultimately selected GIGABYTE's AI TOP ATOM, a personal AI supercomputer based on NVIDIA DGX Spark and powered by NVIDIA® GB10 Grace Blackwell Superchip.

"We are pushing technological boundaries with AI TOP ATOM's GB10 Superchip," beams Hsiao. "And we are grateful for GIGABYTE's market presence here in Australia, as well as the local team's impeccable response time and after-sales service."

Learn More:
《Power of AI: How to Benefit from AI in the Healthcare & Medical Industry》
《Case Study: When Quantum Algorithms Meet GIGABYTE AI TOP 100》
《GIGABYTE Supports Nemotron 3.5 Lightning｜Reshaping the Era of Autonomous Agentic AI with AI TOP ATOM》

Cognitivo purchased a number of AI TOP ATOMs and deployed them as multipurpose developer kits in its New South Wales office. The 1.2-kilogram desktop solution fits neatly in the palm of one's hand, measuring only 15 by 15 centimeters long and wide, and about five centimeters high. Not only can AI TOP ATOMs operate as individual workstations, but they can also cluster up in groups of four to form a single supercomputer; there's a third configuration in which one AI TOP ATOM assumes command as the control node while three others function as compute nodes. Hsiao's team highlights four key advantages to charting the course to their next big breakthrough in reasoning AI with assistance from GIGABYTE's AI TOP ATOMs:

● Petascale Performance Perfect for Localized AI Workloads

In terms of performance, AI TOP ATOM blows the competition out of the water, serving up one petaflop of AI compute at FP4 precision and 128GB of coherent unified system memory. The GB10 Grace Blackwell Superchip combines a single 20-core ARM-based CPU with fifth-generation Tensor Cores and NVIDIA NVLink™-C2C to engage in prototyping and fine-tuning on 200B-parameter generative AI models—upgradeable to 405B parameters if the user interlinks two AI TOP ATOMs through an AI TOP QSFP Cable, which offers high-bandwidth, low-latency communication enabled by NVIDIA® ConnectX® 7 NIC. What this means for Cognitivo is that it has control of computing prowess that can go toe-to-toe with the world's eminent AI software developers—both today and tomorrow.

Hsiao's team reports that it relies on AI TOP ATOM's high-performance, high-memory qualities to simulate user scenarios in low-throughput settings, in order to prep the reasoning AI for high-throughput user environments in real life.

● Energy-efficient, Token-effective Alternative to Bare Metal

AI TOP ATOM is extremely energy-efficient, matching its compact form factor with a modest 240W power draw that lets it deploy en masse in any run-of-the-mill office space. The cost savings are not limited to energy—Cognitivo has discovered that moving from cloud to deskside development can improve OpEX in terms of both man-hours and tokens, as well.

"Even when you purchase cloud services, you have to contend with queue times and availability," explains Hsiao. "You also burn through tokens as you fine-tune and test different LLM configurations. By running everything from vLLM to docker on AI TOP ATOMs, we have found that it's more cost-effective and accessible than anything on the cloud—including renting bare metal servers from CSPs."

● Best-in-class Demo Kit for On-prem Reasoning AI

AI TOP ATOM is the proverbial pint-sized powerhouse which the Cognitivo team can effortlessly bring to sales pitches to conduct live demonstrations. It also acts as a PoC for clients interested in incorporating on-premises reasoning AI: if the software can run without a hitch on a desktop platform, integration into the larger AI infrastructure should not pose an issue.

● Total Solution Doubles as Launchpad for Future Growth

Hsiao says purchasing AI TOP ATOMs was a "great investment"—partly because the solutions come bundled with GIGABYTE's world-class customer support and service, and partly because the quick lead time and budget-friendly pricing accelerated Cognitivo's pursuit of market leadership in the field of KG-enhanced LLMs for enterprise use. As the team continues to introduce reasoning AI tools to more and more verticals, Hsiao is contemplating upscaling the package to include agentic functionalities, or even to provide it on the neocloud. He sees the age of autonomous businesses as only being two to three years away: when the time comes, the solution provider who already has a foot in the door, who has invested in channels and customers and capabilities across multiple sectors in the lead-up to the paradigm shift—that's the AI champion who's coming out on top. In this context, it is imperative for first movers to embrace the latest innovations in production—not only so they don't fall behind the curve, but also to help maintain a holistic, up-to-date view of the technological changes reshaping the market.

"In AI, there's no killer use case, the lag effect is in everything, and the value curve to ROI can be long. What we can do is write as much software as we can to secure our slice of the market," Hsiao says. "Australia may not be the biggest market, but a great idea goes a long way toward catching up with the world's AI leaders, and we're excited to actualize our great ideas on AI TOP ATOM."

Thank you for reading our case study about the Australian AI software trailblazer Cognitivo's endeavors to tackle the "last mile" of enterprise AI with reasoning AI built on GIGABYTE's AI TOP ATOM. We hope this article has been helpful and insightful. For further consultation on how you can incorporate the AI TOP series in your AI development, we welcome you to reach out to us.