A GPU-accelerated stats library for batch computing.
Available as a Python module with a core written in C++.
Features
Exastat is a statistical library that evaluates large numbers of statistical tests concurrently on the GPU in a short period of time.
It is delivered as a Python package with a C++ core, aimed at quantitative research and trading desks that need to scan an entire universe of instruments rather than test a single pair.
Statistics
- Cointegration tests between 2 or 3 instruments
- ADF test
- KPSS test
- Hurst exponent
New tests are added and the algorithms refined continuously.
Infrastructure
- Multi-GPU execution
- Human-friendly Python interface
- Streaming and dynamic price updates
- Configurable working memory
Performance
In the following example, each case is a VECM lag search: seven VECM fits, one per lag order from 2 to 8 (the best model based on AIC is returned). The series are daily equity closing prices with 500 observations.
Johansen lag search, 2 to 8 lags: cases per second
Pairs two series per case
- 91statsmodels
one processbaseline - 490Exastat
1× RTX 30605.4× - 18,900Exastat
1× H100207× - 35,700Exastat
2× H100391×
Triples three series per case
- 77statsmodels
one processbaseline - 452Exastat
1× RTX 30605.9× - 13,700Exastat
1× H100178× - 26,300Exastat
2× H100343×
On two H100 GPUs, Exastat processes 35,700 cases per second. This is equivalent to 28 microseconds per case.
statsmodels takes 11 milliseconds per case, run one case at a time in a single process on an AMD EPYC 7532.
How it was measured.
Daily closing prices of real equities, 500 observations per series,
with Exastat 0.1.0 and statsmodels 0.15.0.
Only the solve call is timed.
Exastat values are the median of three runs over 10,000 cases
with 4,096 MiB of compute memory per GPU.
The H100s ran in a cloud container,
the RTX 3060 on an AMD EPYC 7532 host.
statsmodels ran coint_johansen and a VECM fit
at each lag, in one run over 500 cases in a single Python process
on the same EPYC CPU.
Results depend on the setup. Measured on 12 October 2026 on one data set and three machines. Throughput varies with the GPU model and count, the host CPU, driver and CUDA versions, window length, series per case, lag range and the data. Benchmark it on your own hardware and data under a trial or purchased licence.
Requirements
System prerequisites for Python wheels and C++23 native builds.
- Operating system
- Linux x86_64
- glibc 2.38+
- GPU
- Compute capability 7.0+
- V100, T4, A100, L40S, H100
- CUDA
- 12.x, driver 525+
- Python / C++23
- 3.10 – 3.14 / GCC 14+
Pricing
Perpetual licences with 1 to 3 years of updates and maintenance included, in USD, excluding VAT. Sold to businesses only, not to individuals.
Single seat
one named user, one GPU
- Python module and C++ library
- 1 year of updates and maintenance
Team
2 – 5 named seats
- Everything in Single seat
- Use in containers and CI
- Multi-GPU execution
- 2 years of updates and maintenance
Entity licence
one legal entity
- Unlimited internal seats
- 3 years of updates and maintenance
Updates and maintenance
of the licence price per year
- One further year of new releases
- For any licence tier
The licence is perpetual: the key does not expire, and every release published during the updates and maintenance period stays usable after it ends, with no recurring fees. Later releases require the period to be extended, at 20% of the licence price per year. The licence agreement is supplied on request.
Every licence includes training materials, examples, the user guide and agent skills for programming.
Academic licences are free on approval, for non-commercial research and teaching, with a citation requirement.
Evaluation access
Evaluation builds are sent on request to qualifying firms. There is no self-serve download.
- Tell us your setup. GPU model, Linux distribution, Python version, and roughly how many instruments you scan.
- You get the library and a trial licence key. A Python wheel or the C++ library. The licence key is checked offline, with no licence server and no hardware lock. Every feature is enabled during the evaluation. The library stops working when the evaluation expires.
- Benchmark it on your own machines. The figures on this page hold only for the hardware and data they were measured on.
Get in touch
Evaluation access, pricing or academic licences.
Who we are
We are a small team of electrical engineers, financial analysts, and computational scientists amazed by what modern computers and GPUs can achieve. We created the core of Exastat in C++ years before the emergence of LLMs and agentic development practices.