Sandbox::https://plato.asu.edu/ftp/mipfeas.html

30 October 2026     =========================================================
                    Open Source MIPFEAS Benchmark on the MIPLIB2017 Instances
                    =========================================================
                    Michael Bussieck, mbussieck@gams.com, assisted by NVIDIA

The 233 feasible benchmark instances of MIPLIB2017 have been used in this new benchmark. The goal is to determine how fast solvers find good feasible solutions in a given time as measured by the PRIMARY INTEGRAL which is defined as $P(t)= \int_{0}^{T} 2 if z(t)= \infty, 1 if z(t) z^{*} <0 else \frac{|z(t)- z^{*} |}{max (|z(t)|,| z^{*} |,1)} dt / T$ where $z^{*}$ is the optimal solution value, $z(t)$ is the incumbent value at time $t$ . For a detailed explanation there is a blog post on the GAMS web site.

The following codes were run through GAMS for T=600 seconds with 10 threads/workers and otherwise default parameters on an NVIDIA DGX Spark, equipped with an ARM CPU (10 Cortex-X925 performance cores and 10 Cortex-A725 efficiency cores) and a GB10 120GB VRAM GPU. The ReXi solver ran directly on the MPS instances with arguments --workers 10 and --timelimit 600. A script has been used to convert the ReXi logs to tracefiles. Cuopt-b200 runs were performed on a NVIDIA DGX B200 with an Intel� Xeon� Platinum 8570 using 24 threads.

CBC-2.10.13 CBC
cuOpt-26.10 cuOpt
HiGHS-1.15.1 HiGHS
ReXi-Opt-26.10 ReXi
SCIP(spx)-10.1.0 SCIP
SCIPConc(spx)-11.0.0 SCIPConc
This benchmark also includes the commercial solvers COPT (8.0.7), CPLEX (22.2.0.0), and Xpress (9.9.1). While we cannot publish individual results, the vendors agreed to publish results in an aggregated form. So we present them as the virtual mean commercial solver (VMCS).

+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
solver                    cbc   cuopt [cuopt-b200] highs    rexi    scip scipconc    VMCS
Geom. Mean (sh=0.001)  0.0980  0.0088     0.0076  0.0447  0.0103  0.0609   0.0168  0.0122
Geom. Mean [Scaled]   11.1463  1.0000     0.8591  5.0887  1.1720  6.9309   1.9092  1.3908
Feasible found (#)        191     232        231     217     223     207      223     226
Proven Optimal (#)         77      86         85     128     113      82      115     180
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
cuopt-b200 is out of competition due to this observation by GAMS

While the NVIDIA DGX B200 was considered, its extreme cost, constrained 
availability, and comparatively slower CPU disproportionately disfavors 
CPU-based methods, making it unsuitable as a benchmark standard. We still 
include out-of-competition cuOpt runs on the B200 (Intel� Xeon� Platinum 8570, 
24 threads) purely as a high-end exhibition to show what raw enterprise hardware
can do.