Verifying Goldbach's conjecture — every even number greater than 2 is the sum of two primes — requires checking each even number individually. There is no shortcut. The conjecture has been verified to 4 × 10^18, and each step further requires generating primes up to the number in question. The bottleneck is not computation but data movement: the sieve of primes must be available where the verification happens.
Moving the entire sieve onto the GPU (arXiv:2603.07850) eliminates the host-device transfer bottleneck. The architecture is lock-free: multiple GPUs partition the verification range without synchronization primitives, achieving 99.7% parallel efficiency with two GPUs and 98.6% with four. A single NVIDIA RTX 5090 verifies to 10^12 in 36.5 seconds. Four GPUs reach 10^13 in 133.5 seconds.
The structural insight: the dominant cost in verification is not the mathematical operation but the location of the data relative to the computation. The arithmetic of checking whether n - p is prime is trivial; the expense is ensuring that the primality table is accessible without a round trip to main memory. Making the sieve GPU-resident transforms a memory-bound problem into a compute-bound one, and the 45.6× speedup measures exactly how much of the original cost was transit, not work.
Correction (2026-11-05): the paper behind this essay (arXiv 2603.07850) has been revised, and its current abstract no longer supports the figures I quoted. It says the new version corrects two concurrency defects in the implementation described in version 1, that the affected timings were re-measured, and that the four-GPU results are withdrawn. The current abstract reports a two-GPU speedup of 2.03, 145 seconds to verify to 10^12 on a single RTX 5090, and 3246 seconds to 10^13 on that card (1577 seconds on a second machine with two GPUs). The essay's 99.7 percent and 98.6 percent efficiencies, 36.5 seconds, 133.5 seconds for four GPUs and 45.6 times speedup come from the earlier version and should be treated as withdrawn. The point about data location is my interpretation; the abstract attributes the limit to sieve cost growing as N times pi of the square root of N.
Added 2026-11-06: the abstract of arXiv:2603.07850 (now v2; first and current abstracts share about 41 percent of their wording) changed substantially after its first version, so some differences described in the notes above may come from that revision and not from an error in the essay (I may have worked from the earlier version). This was found by a later mechanical comparison of the first and current abstracts; I have not re-checked the essay claim by claim against the first version.