“`html
Artificial intelligence is transforming the manner in which enterprises store and access their information. This is due to the fact that conventional data storage solutions were constructed to manage simple commands from a limited number of users simultaneously, while current AI systems, equipped with millions of agents, necessitate ongoing access and processing of vast amounts of data concurrently. Conventional data storage solutions now present layers of intricacy, which hinders AI systems because data must navigate through multiple tiers before it reaches the graphical processing units (GPUs) that act as the neural cells of AI.
Cloudian, co-founded by Michael Tso ’93, SM ’93 and Hiroshi Ohta, is aiding storage in keeping pace with the AI surge. The company has created a scalable storage solution for enterprises that facilitates seamless data flow between storage and AI models. This system minimizes complexity by incorporating parallel computing into data storage, amalgamating AI functions and data onto a single parallel-processing platform that stores, retrieves, and processes scalable datasets, ensuring direct, high-speed transfers between storage and GPUs and CPUs.
Cloudian’s integrated storage-computing platform streamlines the process of constructing commercial-scale AI tools and offers businesses a storage foundation capable of keeping with the rise of AI.
“One aspect people often overlook about AI is that it fundamentally revolves around data,” Tso mentions. “Achieving a 10 percent enhancement in AI performance requires not just 10 percent more data or even 10 times more data — you actually need 1,000 times more data. The ability to store that data in a manageable way, allowing for computations to be integrated so that operations can run while the data is being gathered without having to move it — that’s the direction this industry is headed.”
From MIT to Industry
As an undergraduate at MIT in the 1990s, Tso was introduced to parallel computing by Professor William Dally — a computational methodology where numerous calculations occur simultaneously. Tso also collaborated on parallel computing projects with Associate Professor Greg Papadopoulos.
“It was an extraordinary period since most institutions had just one super-computing endeavor, while MIT had four,” Tso reminisces.
As a graduate student, Tso collaborated with MIT senior research scientist David Clark, a computing pioneer instrumental in shaping the early architecture of the internet, notably the transmission control protocol (TCP) which facilitates data transfer between systems.
“During my time at MIT as a graduate student, I focused on disconnected and intermittent networking functions for large-scale distributed systems,” Tso reflects. “It’s amusing — 30 years later, that’s precisely what I’m still engaged in today.”
After graduating, Tso joined Intel’s Architecture Lab, where he devised data synchronization algorithms utilized by Blackberry. He also developed specifications for Nokia that sparked the ringtone download market. He subsequently became part of Inktomi, a startup co-founded by Eric Brewer SM ’92, PhD ’94 that was at the forefront of search and web content distribution technologies.
In 2001, Tso established Gemini Mobile Technologies with Joseph Norton ’93, SM ’93 and others. The company eventually constructed the largest mobile messaging systems globally to manage the explosive data growth from camera phones. Then, in the late 2000s, cloud computing emerged as a robust method for enterprises to rent virtual servers while scaling their operations. Tso observed that the volume of data being accumulated was escalating far more rapidly than networking speeds, prompting him to pivot the company.
“Data is generated in various locations, and this data has its own gravitational pull: relocating it will incur costs and time,” Tso elaborates. “This implies that the ultimate solution is a distributed cloud that extends to edge devices and servers. You must transport the cloud to the data, rather than the data to the cloud.”
Tso officially launched Cloudian out of Gemini Mobile Technologies in 2012, with a renewed focus on supporting clients with scalable, distributed, cloud-compatible data storage.
“What we initially missed when we established the company was that AI would become the ultimate use case for edge data,” Tso reveals.
Despite Tso’s research at MIT beginning over twenty years ago, he perceives strong links between his past work and today’s industry.
“It feels as if my entire life is replaying because David Clark and I were involved with disconnected and intermittently connected networks, which are integral to every edge use case today, while Professor Dally was concentrating on extremely fast, scalable interconnects,” Tso says, noting that Dally is now the senior vice president and chief scientist at the leading AI firm NVIDIA. “Now, when you analyze the modern NVIDIA chip architecture and their interchip communication methods, Dally’s contributions are evident. With Professor Papadopoulos, I worked on accelerating application software by integrating parallel computing hardware without rewriting the applications, and that’s exactly the dilemma we are now addressing with NVIDIA. It’s serendipitous that all the work I undertook at MIT is materializing.”
Currently, Cloudian’s platform employs an object storage framework whereby various data types — documents, videos, sensor data — are stored as unique objects accompanied by metadata. Object storage can handle massive datasets within a flat file structure, making it suitable for unstructured data and AI systems, yet it has traditionally struggled to deliver data directly to AI models without first duplicating it into a computer’s memory system, leading to latency and energy inefficiencies for enterprises.
In July, Cloudian announced the enhancement of its object storage system with a vector database that organizes data in a manner immediately applicable by AI models. As data is absorbed, Cloudian is computing the vector representation of that data in real-time to fuel AI tools such as recommendation engines, search functionalities, and AI assistants. Cloudian also revealed a collaboration with NVIDIA allowing its storage system to interact directly with the AI company’s GPUs. Cloudian claims that this new system facilitates even swifter AI operations while lowering computing expenses.
“NVIDIA reached out to us about a year and a half ago since GPUs are only effective when provided with data that keeps them engaged,” Tso remarks. “Now that individuals are acknowledging that it’s simpler to shift AI to the data rather than transfer expansive datasets. Our storage systems incorporate many AI functionalities, enabling us to pre- and post-process data for AI close to where we collect and store the information.”
AI-First Storage
Cloudian is assisting about 1,000 enterprises globally in extracting greater value from their data, including large manufacturers, financial service providers, healthcare organizations, and governmental agencies.
For example, Cloudian’s storage platform is supporting a major automotive manufacturer in using AI to ascertain when each of its manufacturing robots requires maintenance. Cloudian is also collaborating with the National Library of Medicine to store research articles and patents, as well as with the National Cancer Database to archive DNA sequences of tumors — valuable datasets that AI models could analyze to aid in developing new treatments or gaining new insights.
“GPUs have been an exceptional enabler,” Tso asserts. “Moore’s Law doubles computing capacity every two years, but GPUs can parallelize operations on chips, therefore allowing you to connect GPUs together and surpass Moore’s Law. This advancement is propelling AI to unprecedented levels of intelligence, but the only way to fully leverage GPUs is to supply them with data at the same pace that they compute — and achieving that necessitates removing all barriers between them and your data.”
“`