C++20 & CUDA Execution Stack

Indigenous GPU-Accelerated Optimization Solver.

First-order Primal-Dual Hybrid Gradient (PDHG) algorithms and Branch-and-Bound search for massive-scale LP, QP, and MILP industrial workloads.

Formulations
LP • QP • MILP
Scaling Layer
Ruiz Equilibration
Validation
Exact KKT Checker
solver_pipeline.cpp[blend_lp.mps]
PDHG_CUDA_KERNEL
Karush-Kuhn-Tucker Feasibility ChecksPASSED (Tol: 1e-6)
Primal Residual ||Ax - b||_∞4.180e-7
Dual Residual ||Aᵀy + s - c||_∞8.920e-8
Relative Duality Gap |cᵀx - bᵀy|1.120e-7
Complementary Slackness xᵀs0.0000e+00
KKT Conditions Verified
Memory: 18.4 MB
Interconnected Constraint Systems

The Industrial Decision Problem

Industrial operations cannot optimize isolated variables. Altering a single blending quota or unit schedule triggers cascading constraint violations across power, inventory, and transportation.

100K+
Interlinked continuous variables
Coupled
Multi-stage linear constraints
Node: ProductionDecision Variable
x_prod ∈ [0, M]

Directly determines crude throughput and refinery capacity utilization.

Constraint Network (Click nodes)6 Nodes • 8 Bounds
ProductionBlendingSchedulingResourcesInventoryTransportation
Click any decision variable to trace ripple propagation
Mathematical Foundation

How the Solver Operates

Click any step to inspect the underlying mathematical formulation, CUDA execution model, and asymptotic complexity.

STEP 01
Formulation
x₁x₂∇f(x)

Decision Vector Space

min c^T x + 1/2 x^T Q x

High-dimensional parameter space where decision vectors x ∈ ℝⁿ are mapped across objective cost gradients ∇f(x).

CLICK TO EXPAND
STEP 02
Geometry
Feasible K

Constrained Polytope

Ax ≤ b, l ≤ x ≤ u

Constraint hyperplanes intersect to carve out an enclosed convex polytope containing all physically admissible decisions.

CLICK TO EXPAND
STEP 03
Algorithm
x0x1x2x3x4

Primal-Dual Operator Step

x^(k+1) = prox(x^k - τ A^T y^k)

First-order operator splitting navigates high dimensions via GPU-accelerated sparse matrix-vector products.

CLICK TO EXPAND
STEP 04
Verification
OPTIMAL x*Res: 4.12e-9

KKT Certified Optimum

||Ax* - b||_∞ ≤ ε, Gap ≤ ε

Rigorous duality gap computation verifies zero constraint violation and certifies global mathematical optimality.

CLICK TO EXPAND

Architectural Control

A technical comparison of standard proprietary dependencies versus our indigenous stack.

Dependency Stack

Application Layer
Proprietary Solver Interface (Black Box)
Closed Binary Kernel

Indigenous Stack

Application Layer
Our Interface Layer
Our Solver Core (Transparent)
Numerical Layer (CPU / GPU)

Control

Full ownership of the mathematical execution path and optimization parameters.

Transparency

Inspect exactly how constraints are processed and variables bounded.

Extensibility

Modify numerical kernels directly to support specialized problem structures.

Pipeline Architecture

Execution Pipeline

The solver maps high-level mathematical structures down to hardware-optimized numerical operations. Each stage is designed for maximum throughput.

View full architecture details

Reads MPS format problem files and constructs sparse matrix structures.

IN:.mps / .qps files
OUT:In-memory Sparse Matrix
Engine Specifications

Solver Capabilities

A versatile mathematical engine designed for the complexities of modern industrial problems.

Continuous Optimization

Our core engine utilizes the Primal-Dual Hybrid Gradient (PDHG) method for Linear Programming (LP). This first-order approach avoids the costly matrix factorizations required by traditional Interior Point Methods, enabling it to scale efficiently to massive, highly sparse industrial instances. Quadratic Programming (QP) is fully supported for problems requiring non-linear objectives.

Discrete Optimization

For discrete decisions—such as scheduling, equipment activation, or step-function costs—the solver includes a custom Branch-and-Bound (B&B) framework. It recursively partitions the Mixed-Integer Linear Programming (MILP) search space, utilizing the continuous LP solver to evaluate node relaxations and efficiently prune suboptimal branches.

Capability
Status
Description
LP (PDHG)
Implemented
First-order primal-dual hybrid gradient method for large-scale linear programs.
QP
Implemented
Quadratic programming solver for non-linear objective functions.
MILP (B&B)
Implemented
Branch-and-bound framework for mixed-integer linear programming.
Presolve
Implemented
Model reduction techniques including fixed variable elimination and redundant constraint removal.
Ruiz Scaling
Implemented
Iterative matrix equilibration for improved numerical stability.
GPU CUDA
Implemented
Hardware-accelerated numerical kernels for sparse matrix operations.
KKT Validation
Implemented
Rigorous checking of Karush-Kuhn-Tucker optimality conditions.
C API
Implemented
Low-level C bindings for integration with existing industrial systems.
Interior Point
Planned
Second-order method for highly ill-conditioned problems.
Coordinated Architecture

Heterogeneous Computing

GPU acceleration is used selectively for operations that benefit from massive parallelism, coordinated closely with the CPU.

CPU

Orchestration

  • Model management and parsing
  • Presolve and problem reduction
  • Branch-and-bound tree control
  • Heuristics and solver state orchestration
GPU

Acceleration

  • Sparse matrix-vector multiplications (SpMV)
  • Highly parallel numerical kernels
  • Vector additions and dot products
  • Massive throughput for PDHG iterations
Verified Metrics

Performance Benchmarks

No benchmark data yet — Run a benchmark to populate results.

Run in Studio
InstanceClassVariablesConstraintsRuntimeStatus
production_planning_v2.mpsLP125,00084,500--Pending
logistics_network_eu.mpsMILP42,00051,200--Pending
portfolio_risk_model.mpsQP15,00012,000--Pending
STANDALONE & HIGH-THROUGHPUT

Ready to deploy the solver?

Run industrial LP/QP/MILP benchmarks, test custom MPS files, and inspect KKT convergence directly in your browser.

$ solver_solve(&model, &config);KKT CERTIFIED