To support data-intensive workloads, Speedata introduced its Analytics Processing Unit (APU). Dutch neocloud Nebul will be the first to integrate the hardware for Apache Spark workloads. According to both parties, the speed gain is a hundredfold.
The APU is now available in Nebul’s sovereign cloud infrastructure. This IT infrastructure is aimed at organizations that want to run advanced analytics and AI data processing on a large scale, but must remain within national borders.
Europeans are eager for additional AI capacity, according to the two parties. This is also evident from research, which concluded that demand for this infrastructure has tripled in the past year. At the same time, it will have escaped no one’s attention that digital sovereignty has become a key issue at both national and continental level. A viable alternative is needed to house data-intensive workloads for AI outside the US hyperscalers. Speedata’s APU should help Nebul to be that alternative.
100x faster than CPUs and GPUs
The APU is designed for large amounts of data and advanced analytics workloads. It is a specialized ASIC that runs Apache Spark SQL natively. Complex queries, joins, aggregations, and transformations run directly on the silicon instead of an abstraction layer operating via memory. In standard benchmarks, the APU achieves up to 100 times better performance than CPUs and GPUs. That’s no surprise: GPUs are primarily parallel data processors, and CPUs are designed to be jacks-of-all-trades.
This specialization of the APU also translates into costs. Speedata CEO Adi Gelvan states that a customer was able to consolidate from 38 servers to 3, resulting in a 90 percent cost reduction. This demonstrates the significant overhead of the data layer, with a considerable portion of conventional hardware solely engaged in data-intensive tasks. However, bringing storage to compute requires little from the APU, which is designed to make precisely that step.
Although the Speedata chip is designed specifically for a particular task, the complexity of that task should not be underestimated. In addition to traditional batch ETL workloads, the APU accelerates AI data preparation and cleanup. Real-time query-oriented RAG and TAG (Table Oriented Generation) are also supported, allowing structured data and data from tables to be delivered to LLMs at an accelerated rate. This acceleration also helps other hardware, such as GPUs, by allowing these devices to spend more of their time performing calculations rather than waiting for data.
Sovereignty as an operational requirement
Nebul hosts the APU technology in its European data center network. It runs on green energy. As a preferred partner in EMEA, Nebul offers the APU within its Data Platform.
“Sovereignty is not a value judgment, it is an operational requirement,” says Arnold Juffer, CEO of Nebul. “True data sovereignty goes beyond GDPR compliance and EU-based data centers. It encompasses ownership structure, operational control, and jurisdiction. Our customers need to know who owns their infrastructure, who operates it, and who can be compelled to access it.”
That urgency is palpable. Earlier, the Dutch government had lawyers investigate AWS Sovereign Cloud for legal vulnerabilities. The conclusion was as follows: US legislation makes access to metadata technically possible, even within European sovereign clouds of hyperscalers. This means that even the most sovereign solutions from US cloud providers are only limitedly sovereign.