In this special guest feature, Ami Gal, CEO of SQream Technologies, discusses how GPUs and databases are a match made in IT heaven and how GPU databases are poised to take over high performance compute. The reasoning is simple: GPUs can read and process data at speeds far greater than CPUs and are increasing in performance at a rate of roughly 40% per year (equal to the growth rate of data). Ami brings over 20 years of technological and entrepreneurial leadership experience, with a great business approach, and strong technical background. Former VP of Business Development of Magic Software (MGIC) where he generated new growth engines around high performance and complex data integration environments. Ami co-founded Manov, later acquired by Magic Software, which completed a secondary PO around Manov’s solutions.
Even in computer circles, outside of a few core areas (such as deep learning, virtual reality, and autonomous vehicles) the capabilities of the graphics processing unit (GPU) remains relatively unknown. In fact, to most individuals, the thought of even using a GPU for non-graphical computing is mostly unheard-of. That oversight, however, is now changing, starting with GPU databases.
A brief history on GPUs
To appreciate why a GPU may serve as a better computation resource than a CPU, you’ve got to go back to the beginning – the birth of the GPU. GPUs were originally developed as an effective alternative to a CPU when engineers and architects realized that rendering complex images on a screen was a parallelizable problem, that is, instructions from a single instruction stream could be processed simultaneously.
The way they solved the problem was by creating a new processing unit with thousands of less sophisticated cores and faster, higher bandwidth cores. The result of their efforts was a processing unit which surpassed CPUs, both in reading and processing data. (Now you may start to see why GPUs would be desirable for a database, but there’s more to it than that!)
2016 was already the year of the GPU
Not surprisingly, until very recently GPUs were mostly used to do what they were named for: graphics processing. Since the CPU could handle most of the computer’s tasks, the graphics card didn’t see much usage outside of creating realistic gaming experiences.
But, in 2016 a tipping point was reached. See, data accumulation has been growing at a rate of roughly 40% per year, and is accelerating. Consider the Terabyte. You likely remember when the Terabyte was unheard of, but now we’re hearing about smartphones with 1 TB storage. Many organizations are now dealing with Petabytes of data. It may not be long before enterprises are dealing with zettabytes (a billion terabytes) of data.
The reason we hit a tipping point was because CPUs had only been increasing in performance by 10 – 20% each year. Last year, the rapid growth rate of data surpassed what CPUs could handle. This led to many looking for a new solution, and finding it in GPUs.
GPUs and Databases are a Match Made in IT Heaven
The shift towards GPU processing began with supercomputers almost a decade ago and has been trickling down to other areas. Now, in 2017, GPU databases are poised to take over high performance compute. The reasoning is simple: GPUs can read and process data at speeds far greater than CPUs and are increasing in performance at a rate of roughly 40% per year (equal to the growth rate of data).
Just look at the raw FLOPS a CPU and GPU could handle as of now:
- CPU – 200 gigaflops – 200 billion floating point operations per second
- GPU – 120 teraflops – 120 trillion floating point operations per second
Because GPUs can process many calculations simultaneously, they are able to divide and conquer. Large problems are broken down into smaller ones which are solved all at once. A CPU based database, on the other hand, is limited by the number of the cores on the CPU, and is optimized to run serialized computations. This results in a database which can only handle a few queries at once and works to resolve queries one after the other.
You’re likely starting to see just how revolutionary and powerful a GPU database can be.
On the other hand
GPU database contains thousands of cores for each GPU. This enables you to run thousands of queries at once, and with higher bandwidth memory to boot. Best off, the ROI is there for most businesses. Although a full GPU server could be expensive, it can match the computational power of dozens if not hundreds of traditional CPU servers.
This all adds up to create a very compelling argument for GPU databases, and one which will result in their takeover starting very soon.
Sign up for the free insideBIGDATA newsletter.
Speak Your Mind