GPU-Accelerated Engine

AcaysiaRT

High-throughput GPU-accelerated chemical simulation runtime. Powers the Acaysia Autopilot's MPPI controller with real-time reactor dynamics. Designed for massively parallel trajectory sampling with deterministic, production-grade reliability.

2B+

Reactor-steps/sec

130x

Faster than PyTorch

0.52ms

P99 Latency

Architecture

Technical Specifications

GPU

CUDA Architecture

Compute capability sm_75, sm_80, sm_87, sm_89. Warp-level primitives for coalesced memory access. Mixed-precision FP32/FP16 with accuracy guarantees. Async memory transfers with CUDA streams.

Physics

Simulation Capabilities

ODE integration: RK4, Dormand-Prince, stiff solvers. Reaction kinetics: Arrhenius, Michaelis-Menten, custom. Thermodynamic models: ideal, Peng-Robinson, NRTL. pH/buffer equilibrium with automatic species balancing.

API

Integration

C++ core with Python bindings (pybind11). Zero-copy tensor interop with PyTorch/JAX. Batched API for MPPI trajectory rollouts. Thread-safe multi-stream execution.

Hardware

Supported Platforms

Data center: A100, H100, L40S. Workstation: RTX 4090, RTX 6000 Ada. Edge: Jetson Orin NX/AGX. Cloud: AWS p4d/p5, Azure NC/ND series.

Benchmarks

Scaling Performance

Measured on NVIDIA A100 80GB with 1M parallel trajectories.

Throughput Scaling

Near-linear scaling with batch size up to hardware limits.

Latency Distribution

Consistent sub-millisecond P99 latency under load.

Memory Efficiency

Optimized memory layout for maximum GPU utilization.

Combined Analysis

Full benchmark suite across multiple GPU architectures.

Applications

MPPI Control Use Cases

Control

Parallel Trajectory Sampling

Evaluate millions of control trajectories simultaneously for optimal MPPI policy computation. Real-time replanning at 100+ Hz control rates.

Safety

Real-time Constraint Enforcement

Hardware-enforced safety bounds on temperature, pressure, and concentration. Constraint violations detected within single inference step.

Reliability

Deterministic Fallback

Bit-exact reproducibility for safety certification. Guaranteed execution time bounds for real-time control system integration.

Additional use cases in development.

Get Started

Simple, Powerful API

Install via pip. Simulate in three lines of code.

Installation pip

# Install AcaysiaRT with CUDA support
pip install acaysia-rt

# Or install from source
pip install -e ".[dev]"
python setup.py build_ext --inplace

Quick Start Python

import torch
import acaysia_rt as art

# Initial reactor state [N, 8]
states = torch.tensor([[
    298.0,   # Temperature (K)
    100.0,   # Volume (L)
    0.002,   # DIC (mol/L)
    0.002,   # Alkalinity (eq/L)
    0.0, 0.0, 0.0, 1.0
]], device="cuda")

# Simulate one timestep
states_next = art.step(states, controls, dt=0.1)

Ready to Accelerate Your Control Systems

Request Early Access