Industry executives and experts share their predictions for 2024. Read them in this 16th annual VMblog.com series exclusive.
GPU Databases Come of Age in 2024
By Chad Meley, Chief Marketing Officer, Kinetica
Graphic Processing Units (GPUs) are widely acknowledged as the driving force behind the AI revolution, expediting intricate neural network calculations and facilitating breakthroughs in machine learning and deep learning applications. However, their impact is now extending further into big data analytics. The maturation of GPU database architectures signals a notable transformation in data management practices, presenting several compelling reasons reshaping how organizations manage and get value from their data.
To understand the trend, it is important to go back in time to when GPU databases were first released into the market eight years ago. The primary value proposition was (and still is) faster query execution and complex computations on extreme datasets. However, mainstream adoption was hindered by several items: I/O bottlenecks, exotic hardware, and lacking key enterprise capabilities. All of which have been overcome recently.
From Bottleneck to Free Flow
One of the primary drivers behind the adoption of GPU database architectures is their remarkable speed and efficiency. GPUs are purpose-built for parallel processing and handling large data sets, making them exceptionally efficient in handling data-intensive tasks. Recent hardware breakthroughs by NVIDIA, including faster PCI buses and more VRAM, have addressed a key bottleneck that has improved overall system performance and responsiveness.
In contrast to big data analytics, the data-to-compute ratio for AI is way more skewed towards a high level computation for each data element. Analytics workloads have a much lower compute intensity but a much higher I/O intensity. So once you’ve transferred the data into GPU memory, it’s going to blast through whatever analytic operation you’ve set up for it. The tricky part is can you get the right data there or already cached there fast enough.
For a prolonged period, most GPUs faced limitations due to the sluggish evolution of the PCI Express (PCIe) system bus. The reign of the PCIe 3.0 era seemed unending until recent years when a breakthrough emerged. Enterprise hardware vendors have swiftly embraced newer PCI standards, ushering in a remarkable era of significantly elevated throughput and transfer rates for GPUs.
This hardware evolution alone forecasts a promising era for GPU analytics. However, when coupled with another understated hardware trend-the surge in VRAM capacity-it signifies an impending golden age. Consider Nvidia’s H100 NVL dual-GPU cards, tailored for housing large language models (LLM), boasting a staggering 188 GB total VRAM. These dual trends collectively address the primary bottlenecks for GPU analytics-streamlining data delivery to the GPU and ensuring ample proximity for data storage, thereby alleviating these longstanding limitations.
From Specialty to Commodity
The prevalence of GPU database architectures in the cloud is another key factor driving their adoption. When GPU databases hit the scene, they were fighting the cloud megatrend. Enterprise decision makers were understandably reluctant to bring exotic GPU hardware into their data center when the priority was to move to commodity servers in the cloud.
Leading cloud service providers are integrating GPU capabilities into their infrastructure, making it easier for organizations to harness the power of GPUs without the need for large capital investments in on-premises hardware. All of the major cloud service providers (CSPs) such as AWS, Google, and Azure offer GPU cloud computing services. This democratizes access to GPU-accelerated databases, enabling businesses of all sizes to leverage their benefits while reducing risk.
From Nascent to Mature
Revolutionary ideas or groundbreaking engineering feats often spark the inception of databases, heralding a new era of possibilities. Yet, their true impact and widespread adoption might take a surprising amount of time-often a decade or more-to reach maturity. Initially, these databases emerge with unique features or visionary concepts that show promise but might lack the essential functionalities vital for widespread acceptance.
For instance, a database might boast unparalleled processing speeds or innovative data structures, but it’s the gradual integration of critical elements like robust security measures, ANSI SQL compliance, and high availability that eventually propel its ascent. Over the years, the demands of the mainstream market begin to crystallize around fundamental needs: reliable backup and recovery systems, standardized ingest and egress protocols, and support for tiered storage. These become the linchpins determining a database’s broader adoption and success.
What to Look For in 2024
When assessing a GPU database, numerous pivotal factors take center stage. Among these, the database’s capacity to scale effectively through a distributed architecture stands paramount, especially in managing expansive and ever-expanding datasets. Enterprise-grade functionalities emerge as crucial, encompassing robust security measures, tiered storage solutions, high availability features, and seamless integration with widely used tools. Adherence to industry standards, exemplified by compatibility with PostgreSQL, becomes imperative, ensuring seamless connectivity with prevailing systems. Establishing strong partnerships with industry leaders like NVIDIA holds immense significance, granting access to cutting-edge GPU technology and invaluable engineering resources, thereby augmenting the database’s performance and capabilities significantly.
##
ABOUT THE AUTHOR
Chad Meley, Chief Marketing Officer, Kinetica
Chad Meley is CMO for Kinetica. Chad’s experience includes more than 20 years as a leader in SaaS, big data, advanced analytics, and data-driven marketing, strategy and planning for early-stage software companies and large, established leaders alike. Prior to joining Kinetica, Chad was VP of Product Marketing at Teradata, where he played a key role in repositioning Teradata during the rise of big data and the cloud to its current leadership position. Chad has also held a variety of leadership roles centered on data and analytics with Electronic Arts, Dell and FedEx. Chad holds a doctorate from the University of Florida where his dissertation was on Applied Artificial Intelligence, an MBA from the Rawls College of Business at Texas Tech University and a B.A. in Economics from the University of Texas.
