DataPelago Inc. today unveiled what it calls a “universal data processing engine” that powers high-speed computing at massive scale by making better use of an organization’s underlying infrastructure.

The three-year-old startup also announced $47 million in new funding.

The company’s cloud-based software framework is independent of the operating system and leverages all available CPU, graphics processing units, floating point units and data processing units to apply to generative artificial intelligence and “data lakehouse” analytics across large volumes of unstructured data.

The engine is based on two intelligent layers called the DataVM and the DataOS. Together they process data in the most efficient way possible based on the characteristics of the data and the hardware resources available, the company said.

DataPelago said the DataVM uses a domain-specific instruction set for data operators while dynamically abstracting accelerated computing elements. DataOS maps data operations to underlying processing resources to optimize performance. The result is an analytics engine that the company claims can process data up to 10 times faster than traditional computing platforms at one-half to one-third the cost.

Configure at runtime

“If you have a sea of GPU, FPGA or CPU servers available, the system will automatically decide where to run which workload,” said co-founder and Chief Executive Rajan Goyal. “We wanted to build an engine that fits into the new ecosystem without disrupting it to deliver the promise of accelerated computing without even a single line of change in the application.”

Goyal was previously a distinguished engineer at Cavium Inc., which built multicore processors for domain-specific use cases and sold to Marvell Technology Group Ltd. for $6 billion in 2018. He was also chief technology officer at chipmaker Fungible Inc., whose data processing units are used in data centers to offload networking, storage and security tasks. It was acquired by Microsoft Corp. in early 2023.

DataPelago’s architecture is based on three technology pillars, Goyal said. The first is a virtual machine that abstracts underlying instruction sets so developers don’t need to write operating system kernels for different kinds of hardware. “It generates the code at runtime for the target CUDA kernels, or the ROCm kernels in AMD, or the field-programmable gate arrays,” he said, referring to platforms from Nvidia Corp. and Advanced Micro Devices Inc.

The DataOS, “when given a query plan, decides at runtime and dynamically maps operators to run CPUs, [single instruction, multiple thread] processors, vectorized machines, GPUs and FPGAs,” he said. “We not only look at the data affinity to generate the tasks but also the capability of the commuting element to map the task to the right source.”

DataOS derives workload statistics from metadata and sends a sideband signal to the runtime engine about the processing elements needed. “That’s what we mean by reconfiguring the computing process,” Goyal said.

The third piece is a composable architecture that can be integrated with query engines and analytic frameworks such as Apache Trino and Apache Spark without changes to the upper processing layer.

Goyal said the company has been testing its technology with customers that are looking to control spiraling costs for analytics hardware or want to train artificial intelligence models more quickly. He claimed DataPelago’s technology is able to achieve 10-fold performance improvements while reducing total cost of ownership by a factor of two to three times.

The new round was led by Eclipse Ventures LLC, Taiwania Capital Management Corp., Qualcomm Inc.’s venture capital arm, Alter Venture Partners Management Inc., Nautilus Ventures Advisors US LLC and Silicon Valley Bank.

Photo: photopin cc