Fleet routing and scheduling has long operated under an uncomfortable constraint: the bigger and more complex the problem, the longer it takes to get a solid solution, and the more compromises have to be made getting there.
But a fundamental shift in how optimization problems are solved is changing that calculus entirely.
In a recent webinar, Chris Doersen, JBF’s Principal of Client Engagement, and David Lubinsky, Managing Director of Opsi Systems, discussed why traditional routing technology has hit a ceiling and what GPU-accelerated optimization means for fleet operators who are ready to move beyond it.
Watch the conversation here or keep reading for a summary.
Why Traditional Optimization Has Hit a Wall
For the better part of two decades, routing and scheduling software has relied on CPU-based solvers. These systems work, but they work slowly, and that slowness has consequences that ripple through the entire operation.
David explained, "As the size of the problem gets bigger, the complexity grows exponentially. So, you either have to compromise on the quality of the solution or the amount of time that you wait to get a good solution." In practice, that often means solve times of 30 minutes to two hours for larger fleets. Often with only one shot to get it right, route planners are left making manual adjustments after the fact rather than exploring better answers.
The current state is that you run an optimization, take those results, and "add a little bit of human flavor to it," where route planners tweak things they know the system didn't fully capture. That's not a technology failure by any means; it's a structural limitation. And it means fleet operations are routinely leaving 5%, 10%, even 20% in cost savings untouched just to maintain workable timelines and driver quality of life.
Fleet costs in the US have risen more than 40% over the past four to five years. The gap between what optimization software produces today and what's actually possible isn’t a minor inefficiency. It's a strategic liability.
Solving in Seconds, Not Hours
The breakthrough David and his team at Opsi have developed is Chora, a routing and scheduling engine built around GPU-based computation. The same chips that power AI models like the ones behind large language model chatbots are now being applied to route optimization, with a dramatic difference in results over traditional CPU-based solvers.
Rather than running a single solve across a handful of CPU threads, Chora distributes the problem across thousands of threads simultaneously, leveraging genetic algorithms that thrive in massively parallel environments. The result: a 445-job solve that would take a commercial solver 10 to 15 minutes completed in 30 seconds. And that's running on just one-sixth of a single NVIDIA GPU. With the potential to speed up optimization by simply applying more GPU resources, Chora can scale up to handle almost any routing problem.
"We're talking not just 30% faster type of numbers here — we're talking 15 to 25 times faster. And the quality is not reduced. In fact, you're seeing a couple of percentage points improvement in miles and duration," Chris noted. In benchmarking against both academic and real-world data sets, Chora has consistently delivered 5 to 10% better outcomes than traditional commercial solvers, in a fraction of the time.
The name itself captures the philosophy behind it. Opsi's previous solver was called Solo. Chora (like a chorus) is many voices working together. One voice has limits, but a chorus can do far more.
Real-World Constraints, Finally Modeled Correctly
Speed matters, but so does the quality and realism of the routes that come out the other end. One of the persistent frustrations with older routing systems is that they treat the world as simpler than it is.
A classic example: traffic. Many traditional solvers calculate travel times between any two points without accounting for time of day. That produces routes that send drivers into city centers during morning rush hour, resulting in routes that look fine on paper but fail in the real-world. Chora's processing power allows it to build traffic-sensitive routes from the start, so drivers aren't handed a schedule that simply cannot be executed to plan.
The same applies to driver fatigue rules. Chora models require breaks, lunch stops, and shift limits in a realistic, driver-friendly way, not as afterthoughts bolted onto an otherwise finished plan. 3D load building is another area where the technology closes a significant gap: the system accounts for what can and cannot be stacked, as well as what fits where, rather than forcing operators to artificially cap their stated vehicle capacity at 80%, or even lower, just to ensure trailers can be loaded without running out of capacity.
For heavy urban delivery, chemical compartmentalization, hazmat requirements, or any environment where complexity compounds, the ability to model constraints properly from the start is what separates a route that truly works from one that just gets dispatched and hopes no issues arise during execution.
From One-Time Optimization to Continuous Scenario Planning
Perhaps the most significant operational shift that GPUs enable is the ability to run multiple scenarios and compare them in near real time.
In the webinar demonstration, David ran a 445-job solve, then immediately adjusted driver working hours by one hour and ran it again. The second solve showed that adding that hour reduced the required fleet from 45 vehicles to 41 — a concrete, quantified answer to a question that would otherwise be handled through guesswork or lengthy post-optimization analysis.
The entire demonstration took just over 3 minutes.
Chris noted that this is unlocking questions that are currently treated as strategic, simply because there's no time to ask them without disrupting daily operations.
What happens if we add an extra hour during peak periods?
If there's a weather event and drivers couldn't complete their stops, what does recovery look like?
Can we right-size the fleet to account for the holiday surge without buying trucks that we'll sit on for 10 months?
These are the kinds of decisions fleet managers make constantly, but typically without real data to back them up. When a solve takes 90 minutes, you run it once and commit. When a solve takes 30 seconds, you run it five times and pick the answer that's actually right for your operation.
A Platform for Strategic Fleet Planning
The implications extend well beyond daily routing. The same speed that makes scenario comparison practical on a daily basis opens the door to something fleet operators have long needed: rigorous, route-level strategic modeling.
Where should a new distribution center be located?
What is the true cost of expanding service territory?
What does the right fleet mix look like given projected volume growth over the next three years?
These questions have historically been answered with estimates, simplified models, or consultants running spreadsheet analyses. GPUs make it possible to answer them with full routing solves across hundreds of permutations.
As David noted, some Opsi customers are already running this kind of analysis: "It's just so worth it to be able to really understand where [my depot should be and what the cost of all these different decisions is]." The path from daily operational tool to strategic planning engine is shorter than it might seem, and it's the same underlying capability performing the heavy lifting.
For fleets under pressure to grow delivery capacity without proportionally growing their asset base, this matters. The ability to model exactly how many vehicles a given network configuration requires, not estimate, adds a level of certainty when making capital decisions.
The Bottom Line for Fleet Operators
AI is everywhere in supply chain conversations right now, and it's worth being precise about what's actually happening here. As David put it, large language models aren't solving hard combinatorial routing problems. "You can't go to any of the AIs and give them the data and say, give me a good routing solve on this," he explained. What the AI revolution has done is accelerate the availability and accessibility of the GPU hardware that makes this kind of optimization possible.
The result is a genuine step change in what routing and scheduling technology can deliver. Faster solves, better solutions, realistic constraint modeling, and the ability to answer questions that used to require weeks of analysis, all without asking route planners to fundamentally change how they work — only to change the technology they work with.
Fleets that are still operating on optimization platforms built around CPU-based solvers are carrying a cost they don't have to. The technology to close that gap is available now.
Want to learn more about GPUs and the next generation of logistics technology? Read our whitepaper, Beyond the Limits of Traditional Optimization.
About the Author
Chris Doersen is a Principal of Client Engagement at JBF Consulting, bringing more than 20 years of experience designing, modeling, and implementing logistics solutions for complex transportation networks. He partners with global shippers to drive efficiency, optimize fleet and carrier operations, and enable technology adoption that delivers measurable impact.
Chris has deep expertise with platforms including Blue Yonder, Descartes, Llamasoft, and Appian, and has led large-scale network design and TMS implementations for leading manufacturers and distributors. Known for his analytical rigor and operational insight, Chris helps organizations turn transportation strategy into sustained results.
If your fleet is still running on a routing platform that forces you to choose between solve speed and solution quality, it may be time to take a closer look at what modern optimization can deliver.
Contact us today to learn how our proven approach can deliver measurable benefits for your organization.
FAQs
Traditional CPU-based solvers face an exponential scaling problem: as the number of jobs, stops, constraints, and variables increases, solve times grow dramatically. This forces operators to choose between solution quality and turnaround time — and typically results in runs of 30 minutes to two hours, making real-time scenario comparison impractical.
No. In benchmarking against both academic and real-world data sets, Chora consistently achieves 5 to 10% better outcomes than traditional commercial solvers — even while solving in 5% of the time or less. Speed and quality are no longer a tradeoff.
No. In benchmarking against both academic and real-world data sets, Chora consistently achieves 5 to 10% better outcomes than traditional commercial solvers — even while solving in 5% of the time or less. Speed and quality are no longer a tradeoff.
GPU-powered routing systems can now accurately account for time-of-day traffic patterns, driver fatigue and break rules, 3D load building constraints (including stackability and weight limits), multiple time windows, equipment type requirements, hazmat rules, and compartmentalization — all simultaneously, during the solve rather than as post-processing adjustments.
When a full routing solve takes 30 seconds instead of 90 minutes, it becomes feasible to run hundreds of permutations to support strategic decisions: optimal depot locations, fleet mix analysis, service territory expansions, and annual capacity planning. Questions that previously required simplified estimates can now be answered with actual route-level data.
