Faster to the optimum than quantum

VERTEX is the world's first wafer-scale solver for the Maximum Cut problem, delivering a 45–165× speedup over quantum solvers and digital annealers on large-scale sparse graphs.

1st
WORLD'S FASTEST MAX CUT

VERTEX is the world's fastest Maximum Cut solver for large sparse graphs.

164.9×
FASTER THAN FUJITSU DA3

On G11, VERTEX reaches the same cut 164.9× faster than the Fujitsu Digital Annealer.

45×
FASTER THAN D-WAVE HYBRID

Across G11–G13, VERTEX beats D-Wave Hybrid's cut and hits the target roughly 45× faster.

19.7×
FASTER THAN TOSHIBA SBM

On G14, VERTEX matches the Toshiba Simulated Bifurcation Machine 19.7× faster.

VercelY CombinatorArgonne National Laboratory

Built Wafer-Scale

VERTEX runs on the Cerebras WSE-3, the largest chip ever built, with 900,000 cores and 44 GB of on-chip SRAM on a single silicon wafer. Every core is wired to its neighbors over an on-chip fabric. The entire sparse graph stays resident on-chip, so every spin updates in lockstep with no off-chip memory traffic.

Single tileNSEWFabricMemorySRAM48kBCache512BRegisters16 General Purpose48 Data Structure8-way 16b SIMD16-way 8b SIMDNSEWSingle dietilestilesWafer Scale Engine12 dies7 dies
VERTEX is distributed across
900,000 independent cores per wafer.
Throughput scales linearly with additional WSEs.

Gset 81: Unsolved for 25 years, solved by VERTEX today. Gset is the canonical Max-Cut benchmark. It contains sparse, frustrated, spin glass systems. Some, like G81, have been unsolved for decades. VERTEX solves them in record time.

FIG 0.1 · G11

Toroidal lattice

An 800-spin toroidal grid. VERTEX reaches the certified optimum, a cut of 564.

FIG 0.2 · G14

Planar graph

An 800-node planar Max-Cut instance. VERTEX reaches the certified optimum, a cut of 3,064.

FIG 0.3 · G81

Largest 2D spin glass

20,000 spins, 40,000 couplings — a 25-year-open benchmark, certified optimal at 14,060.

Certified-optimal G81 assignment20,000 spins · cut 14,060

Best-known solutions, reached in record time

Solver / point
Cut
Time (s)
Δcut
Fₜ
G11800 vertices · 1,600 edges · ±1 · best known 564
D-Wave Hybrid
562
3.000
+2
45.0×
Fujitsu DA3
564
10.983
0
164.9×
VERTEX
564
0.0666
G12800 vertices · 1,600 edges · ±1 · best known 556
D-Wave Hybrid
554
3.000
+2
44.9×
VERTEX
556
0.0668
G13800 vertices · 1,600 edges · ±1 · best known 582
D-Wave Hybrid
580
2.990
+2
44.8×
VERTEX
582
0.0668
G14800 vertices · 4,694 edges · +1 · best known 3,064
D-Wave Hybrid
3,058
3.000
+5
5.9×
Toshiba SBM
3,063
10.000
0
19.7×
VERTEX · fast
3,063
0.5088
Fujitsu DA3
3,064
10.978
0
8.8×
VERTEX · target
3,064
1.2465

Best-known cuts from the Gset suite. D-Wave Hybrid and Toshiba SBM from Vodeb et al. (2024); Fujitsu DA3 from Shaglel et al. (2025). Δcut = cut(VERTEX) − cut(reference); Fₜ = reference time ÷ VERTEX time to an equal-or-higher cut. VERTEX reaches the best-known cut on all four instances.

Fixed-budget quality, 10-second cut

Results from Yang et al. (2025): best cut each solver reaches within a 10-second budget. Δcut is measured against the VERTEX operating point named in each panel.

G11VERTEX cut 564 at 0.0666 s
Solver
Cut @ 10 s
Δcut
GA
486
+78
CIM
550
+14
SB
558
+6
PT
364
+200
SA
430
+134
QIS2
556
+8
QIS3
564
0
D-Wave Neal
564
0
G14VERTEX target cut 3,064 at 1.2465 s
Solver
Cut @ 10 s
Δcut
GA
2,952
+112
CIM
3,027
+37
SB
3,047
+17
PT
2,847
+217
SA
2,903
+161
QIS2
3,032
+32
QIS3
3,060
+4
D-Wave Neal
3,054
+10

Quality-at-budget data from Yang et al. (2025), Table 1; cuts reported as absolute values. A positive Δcut means VERTEX reaches a higher cut. VERTEX matches the best reported G11 cut and exceeds every reported G14 cut.

Speed Limits?


VERTEX turns hours into seconds.

Built by Integrated Reasoning.
Backed by Y Combinator.
Powered by Cerebras.