Alibaba Damo Academy’s MindOpt (敏迭) launches GPU solver, claims to tackle 100‑million‑variable linear programs

Breakthrough claim and lead

Alibaba (阿里巴巴) Damo Academy (达摩院) has released a GPU version of its solver MindOpt (敏迭), saying it overcomes a long‑standing “long‑tail” convergence problem and can handle linear programs with on the order of 100 million variables — problems that traditional solvers often cannot finish. Short sentence: this is a big claim. It has been reported that the new solver blends novel algorithms with deep GPU kernel optimizations to move the industry from “can compute” to “compute accurately” at extreme scale.

How it works (brief)

Solvers are often called the “chip of industrial software”: they underpin power dispatch, flight scheduling, manufacturing and financial risk models. Traditional linear‑programming solvers were designed for CPUs and rely on dense matrix factorizations that explode memory use as problem size grows. GPUs avoid that by converting core work into sparse matrix–vector operations and exploiting massive parallelism, but GPU solvers have typically suffered a late‑stage slowdown in precision gains — the so‑called long‑tail effect. MindOpt’s GPU release reportedly introduces algorithmic acceleration strategies that target this tail and reengineers GPU kernels to sustain convergence on ultra‑large instances.

Reported benchmarks and a real‑world test

It has been reported that Damo’s team tested the GPU MindOpt on roughly 2,000 standard linear‑programming instances. According to the lab, the solver reached high precision on more than 99% of problem types — ahead of peer GPU solvers that reportedly hit 96.7%–98.3% on the same suite — and delivered an average 2.67× speedup with a success‑rate edge of over 14% on large problems. Perhaps most striking: it has been reported that on a production‑scale ad‑allocation case involving about 330 million variables and 16 million constraints, commonly used solvers either crashed or failed to find a feasible solution even after 48 hours, while MindOpt’s GPU build reached reliable precision in roughly 1,700 seconds.

Strategic context and outlook

Why does this matter beyond benchmarks? The solver market has long been dominated by Western vendors, and advanced computing software that can exploit new hardware is increasingly viewed through a geopolitical lens as countries face export controls and tech competition. Damo Academy’s Decision Intelligence Lab head Yin Wotao (印卧涛) said the team will continue to unlock new hardware potential for operations research. It has been reported that MindOpt has already won domestic solver competitions, been highlighted in Ministry of Industry and Information Technology case studies, and is in commercial use for high‑volume decisioning — a sign that China’s industrial‑software stack is seeking to move from catching up to setting its own pace. Who benefits? Industries with exploding computational demands — internet platforms, logistics, power grids and chip design — stand to gain the most.