A commonly overlooked aspect of genotyping in the design of association studies is that larger effect alleles are likely lower in frequency in a randomly sampled population, but these are not discovered in shallow discovery panels. Additionally, they are often not selected for assay design because they are deemed more likely to be erroneous SNP calls and provide little information to distinguish the discovery panel samples. We have designed a 1 million SNP Affymetrix genotyping array for Oryza sativa and Oryza rufipogon (Asian cultivated and wild rice) from a deep resequencing panel of 150+ lines representing all five major subpopulations of O. sativa as well as O. rufipogon. The array was designed to capture SNP variation specific to each subpopulation as well as SNPs segregating in multiple subpopulations. SNPs were selected considering local mutation and recombination rates with the goal of discovery panel haplotype recovery from the interrogated markers, but subject to stringent constraints to maximize conversion rate. Many low frequency and singleton allele SNPs in the discovery panel can be recovered by LD with "synthetic SNPs" formed by combining 2 or more interrogated markers. Additionally, a fraction of SNPs are replicated in random locations on the array. The concordance between genotype calls at these replicate SNPs directly estimates the fidelity of each sample's data and alleviates any need to run expensive technical replicates. At one SNP per ~400 bp, this is the highest density fixed genotyping array ever created in a plant or animal system.