Indigenous GPU-Accelerated Optimization Solver (Sovereign Alternative to Express / CEPLEX)
Mangalore Refinery and Petrochemicals Limited (MRPL) · Miscellaneous · Software
Do not attempt a CPLEX alternative — scope hard to a GPU-accelerated first-order LP solver, benchmark it honestly against HiGHS including the instances where you lose, and that narrow, transparent result is worth far more than a broad claim nobody believes.
Data: Public optimization benchmark sets — MIPLIB, Netlib LP, Mittelmann benchmarks
What it actually is
Refinery scheduling, blending and supply chain planning across Indian industry all run on a handful of foreign optimisation solvers — CPLEX, Gurobi, Xpress — with heavy licence costs and no visibility into the algorithms. The ask is to build an indigenous solver core, GPU-accelerated, covering linear, mixed-integer and quadratic programming. The description is explicit that the hard part is the numerical engine, not the modelling interface.
What to build
A solver core rather than a modelling environment: a linear programming engine using a first-order GPU-parallel method suited to large sparse problems, extended toward mixed-integer support through branch-and-bound with presolve, cutting planes and heuristics, exposed through a standard interface that reads MPS and LP files, and — critically — benchmarked honestly against HiGHS and SCIP on the public MIPLIB and Netlib instances with solve times and solution quality reported per instance rather than in aggregate.
Smallest thing that wins the room
Run your GPU LP solver and an established open-source solver side by side on the same large sparse Netlib or MIPLIB instances, and show your solve times and objective values in a per-instance table where you win on some and lose on others, honestly reported.
How crowded this one gets
A guess, projected from the 2025 statements — the last year where both the submission counts and the winners were published.
Quieter than 63% of the 226 · #84 of 226 by expected field
A normal-sized field. Your idea has to be good, not miraculous.
Why: company-sponsored statements drew the smallest fields of all.
This is a guess, not a fact
Nobody has published 2026’s numbers yet. This is an analysed estimate from last year’s pattern, so please do not take it as the truth — check the live counter on the SIH portal before you decide anything. The range covers the middle half of likely outcomes, so one statement in two lands outside it. Entry closes at 500 ideas per statement, so no range goes past that — a statement that reaches the cap fills and shuts rather than drawing an unlimited crowd. The model reads only three things a team can see before choosing — software or hardware, the theme, and what kind of body posted it — and those explain about a quarter of the variation in last year’s field sizes (R² 0.25 on held-out statements). Trust the band more than the number, and the ordering more than either. It cannot see how good your idea is, which is the part that actually decides it.
The scores
The number is the shorthand. The line under it is the reason.
Acceptance potential
2/5Presented as stated this is unwinnable and a team promising a CPLEX alternative will be measured against CPLEX and lose badly — but there is a narrow honest version, a GPU-accelerated first-order LP solver benchmarked transparently, and a team that scopes to that and reports losses alongside wins can produce genuinely respectable work.
Feasibility
1/5A numerically robust simplex or interior-point implementation with branch-and-cut, presolve and cut generation is a decade-scale engineering effort — HiGHS and SCIP represent many person-years and still trail the commercial solvers, so building a competitive core is not achievable regardless of how much time you are given.
Innovation scope
4/5The algorithmic path is genuinely open and GPU-parallel first-order methods for linear programming are an active, recent research direction, so there is real room to contribute something narrow and legitimate.
Clarity
5/5The description states precisely what is wanted — a solver core not a modelling layer, LP, MILP and QP as initial focus, named algorithm families, sparse matrix exploitation — and links the standard public benchmark suites to evaluate against.
Effort
MassiveEven the scoped-down version, a GPU first-order LP solver with proper sparse handling and benchmark harness, is serious numerical computing work, and the full ask is a multi-year programme.
Demo-ability
MediumA benchmark table against established solvers is rigorous and credible to a technical judge, but it is numbers on a screen with no visual payoff and needs framing for anyone outside operations research.
In its favour
- Green flag: The benchmark suites are named and public, so you have a rigorous, standard, unarguable evaluation protocol from day one
- Green flag: GPU first-order methods for linear programming are a genuine recent research direction, giving you a legitimate narrow scope that is defensible rather than a toy
- Green flag: The sovereignty and licence-cost framing resonates strongly with a PSU judge who pays those licence fees
- Green flag: Almost no student team will attempt solver internals, so the field is essentially empty
Against it
- Red flag: The title invites comparison with CPLEX and Xpress, and any submission judged on that comparison loses — reframe explicitly to a scoped LP contribution or do not take this
- Red flag: Numerical robustness on degenerate and ill-conditioned instances is where solvers actually live or die, and a naive implementation fails on real industrial problems while looking fine on easy ones
- Red flag: The full ask spans LP, MILP and QP with extensibility to nonlinear, which no team can deliver and attempting breadth guarantees nothing works well
- Red flag: Reporting only the instances where you win is the obvious temptation and an operations research judge will ask for the full benchmark table immediately
What you will be writing
- PDLP / primal-dual hybrid gradient first-order LP
- CUDA or CuPy sparse linear algebra
- MPS and LP file format parsing
- Branch-and-bound with presolve
- MIPLIB / Netlib benchmark harness
- C++ or Julia numerical core
- Mathematical optimization
- High-performance computing
- Sovereign software infrastructure
Prior art to read before you start
optimization solver engine development · GPU-accelerated numerical methods · benchmark-driven solver evaluation
Analysed by Claude Opus. Every score above is a judgment call with its reasoning attached — kindly cross-check this against the official statement on the SIH portal before your team commits to it.