Large-scale supply chain optimization is the work of solving planning models with millions of decisions, such as how much of every SKU to send to every store or which truck carries which order, fast enough to use the answer the same day. Planning engines get there with three tools: decomposition, which splits a problem into parts a solver can handle; compute scaling, which adds processors, memory and GPUs; and a deliberate trade between solve time and plan quality. The plan has to be good enough to execute and ready before the warehouse opens.
What Makes Large-Scale Supply Chain Optimization Hard
Size grows by multiplication. A retailer with 20,000 active SKUs, 1,000 stores and a 14-day replenishment horizon already has 280 million SKU, store and day combinations. Each combination can carry several decisions: order quantity, transfer, allocation, markdown. Add DCs, trucks and suppliers, and the model passes hundreds of millions of variables.
Integer decisions make it harder still. A truck either leaves or does not; an order is a whole number of cases; a production line either runs a product this shift or it does not. These yes-or-no and whole-number choices turn a linear program into a mixed integer program (MIP), and MIPs get harder much faster than they get bigger. For background on the model types, see our posts on operations research in supply chain management and linear vs nonlinear optimization.
Decomposition: Splitting the Problem Without Losing the Plan
The classic way to handle a model too big to solve at once is to split it.
- By time: solve the next few days in full detail and later weeks in coarser buckets, then roll the horizon forward each day.
- By geography or product: solve each region or category on its own, then reconcile shared capacity such as DC space or trucks.
- By decision layer: fix one set of decisions first, for example which DC serves which store, then solve the rest with those fixed.
- By coarsening: solve a smaller, aggregated version of the network to get a good starting point or a bound, then refine it.
Academic work shows how far this can go. Jiaze Ma and Victor Zavala used graph sampling and coarsening to solve a supply chain design model with over 38 million variables to within 0.5% of optimal in under 22 minutes (arXiv 2111.01249).
Decomposition has a cost. Every split cuts a link between decisions. Solve regions separately and two regions can both plan to use the same truck. Fix the sourcing first and transportation inherits whatever it was given. Planners then fix the conflicts by hand, which is the work automation was supposed to remove. Good engines split only where the coupling is weak and keep linked decisions, such as fulfillment and transportation, in the same solve.
Compute Scaling: Parallel Solvers, Cloud and GPUs
The second tool is more hardware, used well. Modern MIP solvers run branch-and-bound searches in parallel across many cores, so more processors shorten solve time on hard problems. Large models also need memory: the full constraint matrix, the forecasts and the solver’s search tree have to fit.
Cloud infrastructure changed what is practical here. A planning engine no longer has to fit on one server sized years ago; it can use more compute for the heavy nightly run and less during the day. GPU-accelerated optimization is the newest step. NVIDIA cuOpt, for example, is a GPU-accelerated optimization library for vehicle routing, linear programming and mixed integer programming.
Hardware helps little when the formulation is weak. A model with loose constraints gives the solver a poor bound, so it spends its time proving rather than improving, and doubling the cores barely moves the finish time. Tightening the formulation, removing variables that can never be used and fixing decisions that are already settled often cut solve time more than new servers do.
How Do Engines Trade Solve Time Against Plan Quality?
No daily plan waits for a proof of optimality. A MIP solver reports two numbers as it works: the best plan found so far, and a bound on how good any plan could be. The gap between them is the most it could still improve. Planning engines run with a time limit and an acceptable gap, for example stop at 30 minutes or at a 1% gap, whichever comes first.
Choosing those limits is a business decision. A 1% gap on a transportation plan is real money, but a perfect plan delivered after trucks have left is worth nothing. Engines narrow the trade-off with warm starts (beginning from yesterday’s plan), heuristics that find a good first answer quickly, and tighter model formulations that let the solver prune the search earlier.
Running Optimization Every Day
Daily cadence adds constraints of its own. The run has to fit between the time the ERP and POS data arrive and the time orders must be released. That leaves no room for a failed run, so production engines check data before they solve: missing lead times, negative stock, a supplier calendar that closes every DC on a Tuesday.
A day’s changes often touch only part of the problem. A late shipment, a promotion or a demand spike in one region should trigger a re-solve of only the part it touches. A large-scale engine has to solve the full problem on time every day, with recommendations good enough that planners execute them without edits. That share is measured as user acceptance rate.
How Solvoyo Solves at Scale
Solvoyo is a supply chain decision automation platform. Its engine reconciles cost, service, capacity, lead time, inventory, production and transportation concurrently in a single mixed integer linear programming solve, so a plan that looks optimal in each module is not left infeasible in aggregate. MILP optimization runs across the full network, and NVIDIA cuOpt speeds up solve times, alongside more than 30 statistical and deep learning forecasting models selected per series.
At that scale, a Unilever deployment handles 80 million decisions a week, and a multinational fashion retailer runs 10 million automated decisions a day. Because those decisions come out of one solve, with no conflicts between modules left for planners to fix, over 95% of Solvoyo’s plans execute untouched. See the technology behind the engine, how it works in transportation and fulfillment planning, or request a demo.
Frequently Asked Questions
It is solving supply chain planning models with millions of decisions, such as order quantities for every SKU and store, truck assignments and production schedules, fast enough to use the plan the same day. It relies on mathematical optimization, usually mixed integer programming, combined with decomposition, parallel or GPU computing and a set time limit for each solve.
They split problems where decisions are only loosely linked, keep tightly linked decisions in one model, run solvers in parallel on cloud compute, start from the previous day’s plan and stop at an agreed time limit or optimality gap. Data checks run before each solve, and changes such as a late shipment trigger a re-solve of the affected part of the plan.
The optimality gap is the difference between the best plan a solver has found and a proven bound on the best plan possible. A 1% gap means the plan can improve by at most 1%. Daily planning runs usually stop at a target gap or a time limit, because a near-optimal plan delivered on time is worth more than a perfect plan delivered late.





