SpinForge Solves Max-Cut at 200 Billion Variables on a Single 8-GPU Node and Publishes Its Benchmark Results
Every reported cut is recounted on an independent code path; results, charts and conditions are public at spin-forge.com/proof.
LOS ANGELES, October 18, 2026: SpinForge, a Garmo Labs company, today published its research records and solver benchmark results, including a max-cut run at 200 billion variables on one NVIDIA 8×H100 node, which to our knowledge is the largest max-cut instance solved.
Max-cut, equivalent to finding the ground state of an Ising spin glass, is a standard hard problem behind scheduling, chip design and network partitioning. SpinForge's results:
- 200 billion variables, one node. A max-cut instance with 2×10¹¹ variables and 6×10¹¹ edges reached a 77.42% cut in 11.4 hours on one 8×H100 node.
- 16.5 billion variables, independently recounted. At 1.65×10¹⁰ variables the cut is 89.4392% of 2.475×10¹⁰ edges. The result was recomputed from the saved solution in exact integer arithmetic on the CPU, with no code shared with the solver: relative difference 3.2×10⁻⁷.
- Scaling on D-Wave's Pegasus instances. On 500 Pegasus-native instances at ε = 1%, the GPU solver's time-to-target scaling exponent is 0.44 counted in sweeps (95% CI 0.36–0.55), against published values of 1.69 for quantum annealing with error correction and 1.93 for CPU PT-ICM. The published values cover N = 142 to 1,322 with Gurobi-certified ground states; SpinForge's runs cover N = 210 to 3,432 against the lowest energy found. 250,057 runs were recorded across three benchmark families.
- Head-to-head solver results. SpinForge's Garmo Solvers proved 39 of 40 schedules optimal against CP-SAT's 20 (20 s budget), proved 67 Biq Mac max-cut instances optimal against HiGHS's 49 (148 instances), and recorded 37 wins, 26 ties and 21 losses against PyVRP on vehicle routing at equal budget. Results where SpinForge did not win, including large-instance TSP against LKH-3, are published alongside.
“The results are real. We did the work, and we show it. Every number on our proof page was recounted on a separate code path, and we publish where we lost, not just where we won.”
Patrick Joseph, founder, Garmo Labs
Each result page lists its instances, hardware, budget and charts: spin-forge.com/proof. Full papers and posters are available on request by email.
About SpinForge
SpinForge, a Garmo Labs company, builds optimization and AI-infrastructure software: solvers, checked decisions and engineering for AI performance, complex planning, computational operations and scientific discovery. spin-forge.com
About Garmo Labs
Garmo Labs, LLC (Los Angeles, California) builds chip architectures, software platforms, semiconductor IP and reference systems. garmolabs.com
Media contact
Patrick Joseph, patrick@garmolabs.com