What are GPU Databases?
What makes GPU databases different?
The first item to address is what exactly is a GPU Database. It's actually quite simple. A GPU database is a database management system, relational or non-relational, that uses a GPU (graphical processing unit) to perform some or all database operations. While hardware accelerated databases are not new, the usage of GPUs specifically is still cutting-edge. Most GPU database systems are analytical database systems, and widely regarded as very fast and cost-efficient. Subsequently, GPU databases are also more flexible in processing many different types of data, or much larger amounts of data.
There are several types of GPU databases
Just as there are several types of database systems, GPU databases come in a variety of shapes and sizes:
- GPU-specific operation plugin
- Partially GPU-aware databases (SQL - relational)
- Full GPU database, in-memory
- Full GPU databases, larger-than-RAM (big data SQL - relational)
- Non-relational GPU databases (like graph databases)
Each of these has its own pros and cons. Identifying if a specific GPU database system is right for you is very dependent on what you're trying to accomplish. Check out the article titled "Which GPU database is right for me?" for a more thorough run-down of the types and their applications.
No longer just a tech demo
Until 2015, GPU databases were just a technology demo. Running actual big data SQL queries on GPUs was challenging, especially for OLAP workloads, where performance and SQL feature-coverage is key.
Not anymore! In the past two years, the field has evolved tremendously. GPU databases and GPU data warehouses are deployed in a variety of industries, from the world's largest telecoms, through ad-tech, and the finance industry.
GPU Database management systems can provide some crucial key benefits:
With upwards of 5,000 cores per-card, GPUs can bring petaflops and 80,000 cores to a standard rackmount server
Up to 16 GPUs can be installed in a single server, packing more compute per square inch than any other processor
Supercomputer capabilities in a small package results in lower running costs, and lower overall TCO
Learn how to gain an edge in analytics
Download a free whitepaper, which covers how you can turn your existing BI pipeline into a more capable, GPU-accelerated big data SQL analytics system, with topics including:
- Gaining an edge with raw data -Why raw data is much more powerful than pre-aggregated and cubed data, which limits your capabilities
- Query latency at scale - Distributed databases can't cope with multi-table and multi-key joins. Learn how to cope, without cutting out the JOIN operation
- Future-proofing your solutions - How GPU-acceleration offers value-added big data SQL capabilities that scale with your business
How GPU databases work
When databases run analytic workloads, a CPU alone may not be up to the task. As queries grow in complexity, diversity, and the amount of data being processed - acceleration becomes crucial. However, a GPU can't function by itself, as it lacks functionality in several areas, including network and disk I/O. GPU databases often combine CPUs and GPUs to run fast SQL on big-data. The combination of these two devices helps gain the maximum throughput from each of these devices.
The Traditional Method
The GPU-accelerated method
Fast, flexible results for any workload
Applying GPUs for data warehousing is one of the more interesting applications of a GPU. No longer just for shaving queries from seconds to milliseconds - the GPU's high-throughput orientation can actually power data warehousing applications with surprising effectiveness.
With SQream DB, growing data-sizes are not a problem. In our blog titled "Linearly scaling SQL - It's not a myth!" we describe how combining the throughput-oriented GPU with some clever and best-of-breed data techniques lets SQream DB scale extremely efficiently.
Efficient scaling is a key enabler of complex, analytical queries. Together with the formidable JOIN, that SQream DB runs entirely and effectively in-GPU from its inception, GPU-accelerated data warehouses present the next step in hardware accelerated data processing platforms.
The present and future of GPU accelerated database systems
GPU databases are going to be big in the future. To learn why, watch our recorded talk on GPU databases:
What are they good for? Where do they fit in? What makes them a good fit for your needs?