HeteroDB
HeteroDB is a Tokyo-based database acceleration company founded in 2017 by KaiGai Kohei and Takehiko Kashiwagi to productize GPU-accelerated database technology.
Profile
GPU-accelerated extension for PostgreSQL that offloads analytics and batch SQL queries to NVIDIA GPUs for faster execution.
HeteroDB is a Tokyo-based database acceleration company founded in 2017 by KaiGai Kohei and Takehiko Kashiwagi to productize GPU-accelerated database technology. The company's flagship product, PG-Strom, is a PostgreSQL extension that offloads SQL query processing to GPUs, enabling significant performance gains on analytics and batch workloads. PG-Strom automatically compiles SQL into GPU code and executes queries in parallel across thousands of GPU cores with high-bandwidth memory, reducing CPU bottlenecks and wall-clock execution time for data-heavy operations.
The technology has evolved substantially since KaiGai released the first prototype in 2012; the company now distributes PG-Strom versions supporting PostgreSQL 15 through 18, with binary packages for RHEL, Debian, and Ubuntu environments. In April 2025, HeteroDB shipped PG-Strom v6.0, introducing GPU-accelerated sorting with bitonic algorithms, multi-GPU pinned inner buffers for larger joins, Arrow virtual columns for file-skipping optimization, and enhanced numeric aggregation using 128-bit fixed-point arithmetic to reduce floating-point errors. The company operates as a small team (approximately 6 employees) focused on consulting, support, and development.
HeteroDB participates in NVIDIA's Inception Program, receiving technical support and co-marketing resources. Customer use cases span drug discovery acceleration, geolocation analytics, network security auditing, and high-volume data processing where GPU economics justify the engineering overhead. The company positions itself in the competitive GPU database market alongside players like OmniSci, SQream, Kinetica, and BlazingSQL, though maintains a narrower focus on PostgreSQL integration rather than standalone systems.
Who buys this
- Enterprises processing terabyte-scale data warehouses requiring sub-second query latency for analytics
- Life sciences and drug discovery teams running large-scale molecular simulations and data processing
- Financial services firms performing real-time geolocation and position analytics on high-velocity streams
- Cybersecurity operations analyzing packet captures and audit logs at scale for threat detection
Strengths and what to watch
Strengths
- Deep PostgreSQL integration and upstream expertise from founder KaiGai Kohei, a core PostgreSQL contributor since the 2000s
- Mature product (PG-Strom v6.0 as of April 2025) with support for multiple PostgreSQL versions and enterprise Linux distributions
- NVIDIA Inception Program partnership providing technical validation and joint go-to-market activities
Watch for
- Small team (6 employees) and no disclosed VC funding despite $350M total funding claim in CSV—verify actual capital structure and growth trajectory
- Competitive pressure from specialized GPU database startups and cloud-native alternatives that don't require PostgreSQL expertise
- Dependency on NVIDIA GPU availability and pricing; margins may compress if NVIDIA's pricing power increases or alternative accelerators (AMD, Intel) gain adoption
Key Information
- Industry
- GPU Databases
- Founded
- 2008
- Headquarters
- Tokyo, Japan
Frequently Asked Questions
What is PG-Strom?
PG-Strom is a PostgreSQL extension developed by HeteroDB that offloads SQL query processing to NVIDIA GPUs. It automatically compiles SQL into GPU code, executing queries in parallel across thousands of GPU cores with high-bandwidth memory to accelerate analytics and batch workloads.
How much faster are queries with GPU acceleration?
PG-Strom reduces CPU bottlenecks by executing queries in parallel across thousands of GPU cores with high-bandwidth memory. This significantly decreases wall-clock execution time for data-heavy analytics and batch operations, though specific speedups vary based on workload characteristics, data volume, and query complexity.
What's new in PG-Strom v6.0?
PG-Strom v6.0, released in April 2025, introduced GPU-accelerated bitonic sorting, multi-GPU pinned inner buffers for larger joins, Arrow virtual columns for file-skipping optimization, and enhanced numeric aggregation using 128-bit fixed-point arithmetic. These improvements reduce floating-point errors and boost query performance.
Who should use GPU-accelerated PostgreSQL?
Enterprises processing terabyte-scale data warehouses, life sciences teams running molecular simulations, financial services firms performing real-time geolocation analytics, and cybersecurity operations analyzing packet captures at scale benefit most from GPU acceleration. It's ideal when sub-second query latency and handling high-velocity data streams justify GPU engineering overhead.
Who founded HeteroDB?
HeteroDB was founded in Tokyo in 2017 by KaiGai Kohei and Takehiko Kashiwagi. KaiGai, the Chief Architect and CEO, is a core PostgreSQL contributor with 25 years of database expertise. The company operates as a small team of approximately six employees focused on consulting, support, and development.
How does PG-Strom compare to other GPU databases?
HeteroDB competes with GPU database startups like OmniSci, SQream, Kinetica, and BlazingSQL. PG-Strom's key competitive advantage is deep PostgreSQL integration and upstream expertise rather than building standalone systems. It targets organizations already committed to PostgreSQL seeking significant performance gains without expensive platform migration.
How HeteroDB compares
Direct head-to-head against 3 competitors. Picked by 7wData.
HeteroDB
- Positioning
- GPU-accelerated extension for PostgreSQL that offloads analytics and batch SQL queries to NVIDIA GPUs for faster execution.
- Customer segments
- Enterprises processing terabyte-scale data warehouses requiring sub-second query latency for analytics
- Strengths
- Deep PostgreSQL integration and upstream expertise from founder KaiGai Kohei, a core PostgreSQL contributor since the 2000s
- Watch for
- Small team (6 employees) and no disclosed VC funding despite $350M total funding claim in CSV—verify actual capital structure and growth trajectory
SQream
- Positioning
- Standalone GPU SQL database for petabyte-scale batch analytics, available on-premises and via cloud-managed SQream Blue lakehouse.
- Customer segments
- Telecom, semiconductor, and financial services enterprises processing batch workloads that exceed CPU warehouse capacity.
- Strengths
- SQream Blue benchmarked 3x faster than Databricks Photon at one-third cost on 30TB production workloads.
- Watch for
- Proprietary SQL dialect is not PostgreSQL-compatible; teams porting from standard tooling face significant re-engineering overhead.
- Recent moves
- Appointed Liam Galin as CEO in April 2025 to lead North America expansion and commercial scaling.
Kinetica
- Positioning
- GPU-accelerated real-time analytics database converging spatial, time-series, vector, and streaming workloads in a single SQL interface.
- Customer segments
- Defense agencies, financial institutions, and telecom operators needing sub-second queries on continuously updated high-velocity data.
- Strengths
- Ranked number 1 in the 2025 GigaOm Sonar Report for Real-Time Analytical Databases, with defense-validated deployments at NORAD.
- Watch for
- No disclosed funding since 2017 Series A ($77.5M total); vendor runway and long-term engineering roadmap are unconfirmed.
- Recent moves
- Launched Kinetica Cloud on Azure and Oracle Cloud in December 2024; AWS and Google Cloud availability listed as pending.
HEAVY.AI
- Positioning
- GPU-accelerated visual analytics platform absorbed into NVIDIA in 2025; no longer operating as an independent database vendor.
- Customer segments
- Government intelligence, defense, and enterprise geospatial teams requiring real-time visualization of multi-billion-row location datasets.
- Strengths
- Geospatial query rendering pipeline processes billions of location-tagged rows directly in GPU memory without off-device data movement.
- Watch for
- NVIDIA acquisition (2025) creates consolidation risk: independent support terms, pricing, and product continuity are no longer guaranteed.
- Recent moves
- Acquired by NVIDIA in 2025, ending standalone vendor status and folding platform into NVIDIA accelerated computing portfolio.
Sources
- en.heterodb.com — Company founding date (July 4, 2017), founder names (KaiGai Kohei and Takehiko Kashiwagi), headquarters location (Tokyo, Shinagawa-ku), and legal structure
- heterodb.github.io — Current product versions available (PG-Strom 6.1.0), supported PostgreSQL versions (15-18), and recent distribution center updates (April 2, 2026)
- heterodb.github.io — PG-Strom v6.0 release date (April 1, 2025), feature list including GPU sorting, multi-GPU buffers, Arrow FDW, and enhanced aggregation
- github.com — Open-source repository, development activity, release history, and community engagement
- en.heterodb.com — NVIDIA Inception Program membership, company awards (GPU Technology Conference 2017 best startup), and historical product announcements
- www.linkedin.com — KaiGai Kohei's 25-year career background, PostgreSQL expertise, and role as Chief Architect and CEO