Alibaba DAMO Academy said on its official Weibo account today that its MindOpt
solver has formally released a GPU version, leveraging GPU parallel acceleration
and new algorithms to address the "long‑tail effect". Tests on about 2,000
general benchmark instances show MindOpt can stably solve more than 99% of
problem types to high precision and can support linear programming problems at
the 100‑million‑variable scale that were traditionally considered "unsolvable."
The solver is described as "the core of industrial software," handling complex
computations for power dispatch, flight scheduling, advanced manufacturing and
financial management. By contrast, traditional linear programming solvers are
CPU‑designed and rely on complex matrix decomposition; as problem scale grows,
memory requirements can explode and limited algorithmic parallelism can leave
problems unconverged for hours or cause outright crashes.
