Heavy.AI Platform
HeavyDB
Benchmarks
Benchmarks of the world’s fastest analytics database.
HeavyDB
Overview
Performance is at the heart of HEAVY.AI.
The HEAVY.AI platform is powered by the GPU-accelerated database HeavyDB, which has been built from the ground up to leverage the massive parallelism and performance of Nvidia GPUs. In various benchmarks against leading cloud-based data warehouses, HeavyDB has been shown to be 1-2 orders of magnitude faster for a given cloud infrastructure budget. This performance advantage means that instead of relying on static reports and analytics, organizations can now query and visualize their largest datasets interactively and in real-time.
Key findings from our benchmarks:
- Dramatically Higher Performance: HeavyDB provides up to 130X higher performance over leading CPU cloud data warehouses.
- Lower cost of ownership: HeavyDB, particularly running on the Nvidia Grace Hopper GH200 Superchip, requires significantly less cloud infrastructure spend than the leading CPU data warehouses tested.
- Faster Ingest Rates: HeavyDB is able to ingest data at significantly higher rates than CPU-based data warehouses, and can process ingest data on CPU while executing queries on GPU for minimal interference between database read and write operations.
#
1
in Analytics Database Performance
130
x
Faster than CPU-based data warehouses
119
MS
per query at a 1TB scale
Star Schema Benchmark
Star Schema Benchmark(SSB) is a popular SQL benchmark designed to evaluate the performance of database systems when handling complex analytical queries. It models a typical data warehouse environment, specifically focusing on a “star schema” data model where a large fact table is surrounded by smaller dimension tables, commonly used in business intelligence and reporting systems. It should be noted that SSB is derived from the TPC-H benchmark but is simplified to emphasize key aspects of star schema-based workloads.For our tests we used a Scale Factor (SF) of 1000, meaning 1000GB (1TB) of data, comprising approximately 6 billion records of data. This is a common scale at which to run this benchmark, and illustrative of data scales our customers commonly target with our system.
For our tests we used a Scale Factor (SF) of 1000, meaning 1000GB (1TB) of data, comprising approximately 6 billion records of data. This is a common scale at which to run this benchmark, and illustrative of data scales our customers commonly target with our system.
Takeaways
- HeavyDB running on an Nvidia 8XA100 server is between 24.9X and 131.1X faster than the CPU-based cloud data warehouses tested.
- HeavyDB is the only database to run queries at the terabyte scale with the performance necessary for interactive exploration and BI, averaging 0.12 seconds per query on the Star Schema Benchmark at Scale Factor 1000, versus 2.96 to 15.63 seconds per query for the competition.
Test Environment
HEAVY.AI Specifications
HeavyDB was deployed on an 8 Nvidia A100 GPU bare metal instance on Vultr Cloud.
Our Cost
$14.00 / hr
GPU
8 X Nvidia A100 GPUs
Memory
640GB
Competitor Costs
A breakdown of the general cost per hour for each competitor environment.
Snowflake Large
$16.00 / hr
Redshift Serverless
$12.00 / hr
Databricks Large
$ 28.00 / hr
BigQuery 400
$24.00 / hr
Results
(In seconds)
| Queries | HeavyDB8XA100 | Snowfalke Large | Databricks Large | Redshift Serverless 32 RPUs | BigQuery 400 Slots |
|---|---|---|---|---|---|
| Q001.1 | 0.03 | 3.03 | 1.90 | 9.94 | 3.31 |
| Q001.2 | 0.03 | 2.13 | 1.77 | 0.78 | 2.97 |
| Q001.3 | 0.03 | 1.75 | 1.88 | 0.72 | 2.77 |
| Q002.1 | 0.08 | 5.25 | 7.65 | 1.99 | 10.00 |
| Q002.2 | 0.10 | 0.45 | 7.43 | 1.64 | 8.51 |
| Q002.3 | 0.06 | 2.74 | 6.81 | 1.33 | 8.17 |
| Q003.1 | 0.11 | 3.67 | 7.63 | 1.44 | 19.01 |
| Q003.2 | 0.23 | 4.79 | 7.64 | 2.74 | 15.26 |
| Q003.3 | 0.22 | 3.67 | 7.63 | 1.44 | 19.01 |
| Q003.4 | 0.22 | 2.62 | 2.82 | 1.41 | 19.00 |
| Q004.1 | 0.14 | 9.35 | 1.27 | 8.64 | 22.65 |
| Q004.2 | 0.14 | 5.74 | 7.64 | 5.05 | 23.76 |
| Q004.3 | 0.16 | 5.66 | 5.59 | 2.08 | 10.93 |
| Total runtime | 1.55 | 59.87 | 87.87 | 38.54 | 203.16 |
Discover how HEAVY.AI can transform your business.
Connect with our expert team today and explore the possibilities to get started!
TPC-H Benchmark
TPC-H Benchmark is an industry-standard SQL benchmark designed to evaluate the performance of decision support systems by simulating complex business-oriented queries. It represents a broad range of ad hoc queries and concurrent data modifications, reflecting real-world applications where databases are used for analytical processing and reporting. Generally the queries are more complex than those in the Star Schema Benchmark, revolving around a snowflake rather than a star schema.
Takeaways
- HeavyDB running on an Nvidia 1XGH200 server is between 3.4X and 21X faster than the CPU-based cloud data warehouses tested, and offered between 10.2X and 126.7X better performance for the price.
- The GH200 Grace Hopper Superchip scales well to working datasets that do not fit entirely in GPU memory (VRAM), due to the ultra-fast NVLink interconnect between CPU and GPU, featuring 900GB/sec of bidirectional bandwidth.
Test Environment
HEAVY.AI Specifications
HeavyDB was deployed on an Nvidia GH200 GPU Grace Hopper Superchip bare metal instance on Vultr Cloud.”
GPU
AMD EPYC™
CPU Memory
2048GB
GPU Memory
96GB
Cost
$2.99/ hr
Competitor Costs
A breakdown of the general cost per hour for each competitor environment.
Snowflake Large
$16.00 / hr
Redshift Serverless
$12.00 / hr
Databricks Large
$28.00 / hr
BIgQuery 400
$24.00 / hr
Results
(In seconds)
| Queries | HeavyDB GH200 | HeavyD B8X100 | Snowfalke Large | Databricks Large | Redshift Serverless 32 RPUs | BigQuery 400 Slots |
|---|---|---|---|---|---|---|
| TPCH01 | 0.04 | 0.05 | 1.66 | 1.51 | 1.08 | 1.94 |
| TPCH02 | 0.11 | 0.43 | 0.93 | 0.88 | 0.51 | 0.89 |
| TPCH03 | 0.03 | 0.22 | 1.83 | 2.84 | 0.41 | 7.77 |
| TPCH04 | 0.43 | 2.38 | 1.51 | 2.05 | 0.36 | 3.67 |
| TPCH05 | 0.05 | 0.07 | 1.75 | 0.79 | 0.62 | 11.19 |
| TPCH06 | 0.01 | 0.02 | 0.28 | 0.56 | 0.13 | 3.67 |
| TPCH07 | 0.06 | 0.08 | 1.22 | 0.66 | 0.42 | 11.19 |
| TPCH08 | 0.10 | 0.09 | 1.82 | 0.87 | 0.40 | 0.86 |
| TPCH09 | 0.38 | 0.23 | 2.71 | 0.60 | 1.06 | 1.45 |
| TPCH10 | 0.06 | 0.26 | 2.01 | 3.63 | 0.73 | 11.22 |
| TPCH11 | 0.12 | 0.27 | 2.92 | 0.55 | 0.46 | 1.20 |
| TPCH12 | 0.02 | 0.04 | 0.86 | 1.48 | 0.29 | 3.12 |
| TPCH13 | 0.07 | 0.22 | 3.38 | 1.69 | 1.62 | 0.92 |
| TPCH14 | 0.03 | 0.03 | 0.70 | 1.07 | 0.21 | 4.61 |
| TPCH15 | 0.09 | 0.45 | 0.77 | 0.67 | 0.76 | 2.61 |
| TPCH16 | 0.45 | 2.70 | 2.26 | 0.69 | 0.55 | 0.91 |
| TPCH17 | 0.31 | 0.82 | 0.94 | 1.73 | 0.44 | 9.92 |
| TPCH18 | 0.29 | 1.04 | 3.20 | 2.21 | 1.31 | 12.92 |
| TPCH19 | 0.09 | 0.05 | 1.02 | 1.23 | 0.84 | 0.65 |
| TPCH20 | 0.78 | 1.68 | 2.13 | 0.65 | 0.65 | 1.52 |
| TPCH21 | 3.54 | 0.63 | 1.98 | 3.76 | ||
| TPCH22 | 0.38 | 0.76 | 0.91 | 2.41 | 0.45 | 1.33 |
| Total runtime | 3.91 | 11.87 | 34.85 | 28.77 | 13.23 | 82.24 |
Latest Articles
Discover the latest innovations at HEAVY.AI. Hear from our experts as they deep-dive into a wide range of topics, including AI, GIS, Analytics, and more.
Become an Insider by subscribing to our newsletter
Be the first to know about our latest releases and industry insights
We care about your data and privacy. View our privacy policy
