AI needs more power. Perhaps in response to this reality, modular and composable data analytics company Voltron Data has acquired Claypot AI, a real-time AI platform startup. The move is designed to increase AI engine power as we seek ways to build AI more powerfully, more rapidly and more comprehensively. Voltron Data processes ‘large batch’ data workloads on Graphical Processing Units (GPUs) and the company will now be incorporating Claypot AI’s real-time capabilities onto GPU workloads with the aim of making AI builds faster.
As we know then, there’s a race on to provide AI processing power, so why is the data preparation challenge part of this equation so tough?
AI preparation work
When it comes to preparing and provisioning for AI, organizations are constrained by factors such as data preprocessing on Central Processing Units (CPUs) (i.e. using traditional chips to get a headstart on AI), the Extract, Transform, Load (ETL) tasks associated with getting raw data into areas where AI models can work with it… and the work AI teams need to provide so-called ‘feature engineering’, a process that AWS defines as the extraction and transformation of variables from raw data – such as price lists, product descriptions and sales volumes – so that we can use features for training and prediction in a downstream statistic model serving an AI workload.
What all that means is that businesses looking to use AI often can’t get Machine Learning (ML) related capabilities and functionalities forward efficiently or quickly as they struggle to build out big data CPU clusters (groups of chips, designed to be orchestrated so that they work together). As a result, the GPUs that are used for AI/ML workloads are on standby waiting for the preprocessed data from CPU clusters.
Accelerator-native engines
Voltron Data recently launched Theseus, a technology described as an ‘accelerator-native’ distributed query engine. Theseus is built to help take advantage of GPUs and full system hardware accelerators such as high bandwidth memory, accelerated networking and storage. It enables both data preprocessing and AI/ML workloads on GPUs which is said to unify data analytics and AI pipelines on the same infrastructure. The company also touts its ability to lower vendor lock-in risk, energy consumption and carbon footprints.
Theseus is available to enterprises and government agencies with the largest datasets as well as through channel partners. HPE is the first partner to embed Theseus as its accelerated data processing engine as part of HPE Ezmeral Unified Analytics Software.
“I couldn’t be more excited to bring on the entire Claypot AI team [onto our staff]. Together we’re going to be able to accelerate our real-time and MLOps product roadmap with state-of-the-art features for our customers,” said Josh Patterson, co-founder and CEO of Voltron Data. “By acquiring Claypot AI, Voltron Data can enable real-time analytics, feature engineering and MLOps capabilities powered by Theseus as well as open source products Apache Arrow, Ibis and Substrait.”
Nominative determinism?
Claypot AI was founded in 2022 by Chip Huyen, a computer scientist who has written books on designing machine learning systems book and her co-innovator Zhenzhong Xu, who led the streaming data platform team that serves thousands of data use cases at Netflix.
“AI strategies have to start from data strategies. I am thrilled about this new chapter with Voltron Data and the incredible opportunities it presents. We are going to continue to push the boundaries of what’s possible in real-time, feature engineering and MLOps for enterprises,” said Huyen
There’s clearly a lot to talk about in AI. A good portion of the discussion is (unsuprisingly) focused on the use of Large Language Models (LLMs) that provide the data repositories for AI to learn from, there is also much talk of vector databases due to their inherent suitability (many would say essential need rather than suitability) for AI model support… and then there is AI bias & hallucination before we get on to how human workflows are going to be automated, accelerated and improved in the real world. But below all of those discussions is the engine processing and data management power we’re also trying to create to drive AI in the first place.
Whether we now cook up AI with a modern air fryer or an old-school claypot, this stuff is hot… so please use a handle or holder.
