How choosing the NVIDIA H200 and the De Novo Cloud accelerated LLM performance by up to 10 Times
We’ll explain the performance gains that switching from the NVIDIA H100 to the H200 can deliver, and under what conditions a Ukrainian cloud provider has an advantage over hyperscalers
Data Science UA has been developing AI/ML solutions for Ukrainian and international corporate clients for ten years. The company runs some of its GPU workloads on the De Novo cloud, where the development team has tested various accelerators on real-world tasks.
AI/ML Projects Require Powerful GPUs
Data Science UA has been working in the AI/ML field for about ten years, creating solutions for Ukrainian and international companies. The team regularly requires significant GPU resources for developing, testing, and training large language models (LLMs). Data Science UA runs some of its workloads on the De Novo cloud. In particular, one such task requires up to 25 million tokens per hour.
When the NVIDIA H100 Isn’t Enough
The team initially worked on a specific task using the NVIDIA H100 GPU. However, for certain large models, this GPU’s performance was no longer sufficient. Therefore, Data Science UA switched to the NVIDIA H200 — a more powerful GPU with increased memory capacity and higher memory bandwidth. In a real-world project, the performance gain turned out to be significantly better than expected.
Up to 10x Faster on a Real-World Workload
When working with the OpenAI GPT-OSS-120B model, the team was dissatisfied with the H100’s performance. After migrating the same workload to the NVIDIA H200, performance on certain tasks increased up to 10 times. This is significantly higher than the difference seen in synthetic tests and clearly illustrates why GPUs should be evaluated based on actual workloads.
How We Chose a Provider
Data Science UA compared three options: global cloud providers, European providers, and Ukrainian providers. The key criteria were:
- Cost of GPU computing;
- Ability to handle sensitive data;
- Compliance;
- Quality of technical support and responsiveness of the provider.
Why De Novo
The services of global cloud providers were significantly more expensive than those offered by European and Ukrainian providers—the difference was particularly noticeable for large workloads.
European providers may be competitive in terms of price, but transferring sensitive data abroad often requires additional approvals. Among Ukrainian providers, De Novo was chosen specifically for its technical expertise and quality of support.
“De Novo’s expertise, professionalism, and the responsiveness of its technical support were key factors in our decision,” said Kostyantyn Cherkashyn, Solution Architect at Data Science UA.
Currently, the Data Science UA team is working in the De Novo cloud with an NVIDIA H200 card and can scale GPU resources according to the nature of the workload.
Cases
We’ll explain the performance gains that switching from the NVIDIA H100 to the H200 can deliver, and under what conditions a Ukrainian cloud provider has an advantage over hyperscalers
“17,000 unregistered properties” — By the end of this year, Ukraine plans to update the Unified Tourism Registry
AI for Reviewing Construction Projects: How a “Digital Inspector” Analyzes Documents, Identifies Errors, and Automates Processes in Ukraine's Public Sector
The e-Permit service has received an AI module that automatically checks documents, relieving officials of routine tasks
This unique model could change cardiac diagnostics worldwide. Powered by biosensors and NVIDIA H100 capabilities, the system performs diagnostics in minutes instead of hours
This is a story about how artificial intelligence helps the ministry review applications for youth exchanges: it analyzes estimates, filters out vague goals, identifies unjustified expenses, and suggests areas that need clarification
We went to an AI conference for you, and now we're telling you a story about how buying a cheap camera almost ruined a project to implement computer vision at a company. And that's not the only thing that went wrong
Fields are too large, there are too few agronomists, and the weather is too unpredictable. Could AI be the answer? Kernel is implementing 12 ML models — from growth phase forecasting to agrologistics — and showing how to transform agribusiness through data
The Ministry of Youth and Sports of Ukraine has launched the country's first service based on large language models (LLM) on the Dія.Engine platform. The most powerful equipment in Ukraine was used to host the LLM
Every company goes through its own “digital transformation” journey, and each case is unique. This is especially true when it comes to national-scale organizations
The success of an agricultural company today largely depends on information technology. Increasing yields, optimizing logistics, achieving deep automation of routine operations, and ensuring full mobility of employees—all this is possible thanks to the effective use of IT
We discussed how “digital transformation” works in specific examples in the country with Yevhen Yentis, a person without whose participation it is difficult to imagine the success of such national-scale projects as Prozorro and “New Customs”
Interview with Andriy Begunov, Director of the Information Technology Department at PUMB Bank, which ranks among Ukraine's top seven banks in terms of assets (according to data from the National Bank of Ukraine)
The Ministry of Justice of Ukraine is creating an AI assistant that will answer legal questions 24/7. The project will run on LLM in the secure De Novo cloud